Duplicate from prism-ml/Ternary-Bonsai-27B-mlx-2bit

Co-authored-by: Pasha <pashak@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

LICENSE.txt

@@ -0,0 +1,177 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+

NOTICE.txt

@@ -0,0 +1,4 @@
+This software is copyright 2026-present Prism ML, Inc. It is available under the Apache 2.0 license.
+If you publicly deploy or redistribute this software, we would appreciate attribution such as: "Created using Bonsai by Prism ML."
+
+This software is built from Qwen3.6-27B, Copyright 2026 Alibaba Cloud, which is available under the Apache 2.0 License: https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/LICENSE

README.md

@@ -0,0 +1,285 @@
+---
+license: apache-2.0
+library_name: mlx
+pipeline_tag: text-generation
+tags:
+- conversational
+- ternary
+- 2-bit
+- mlx
+- cuda
+- metal
+- on-device
+- hybrid-attention
+- prismml
+- bonsai
+base_model:
+- Qwen/Qwen3.6-27B
+---
+
+<p align="center">
+  <img src="./assets/bonsai-logo.svg" width="280" alt="Bonsai">
+</p>
+
+<p align="center">
+  <a href="https://prismml.com"><b>Prism ML Website</b></a> &nbsp;|&nbsp;
+  <a href="https://github.com/PrismML-Eng/Bonsai-demo"><b>Whitepaper</b></a> &nbsp;|&nbsp;
+  <a href="https://github.com/PrismML-Eng/Bonsai-demo"><b>Demo &amp; Examples</b></a> &nbsp;|&nbsp;
+  <a href="https://discord.gg/prismml"><b>Discord</b></a>
+</p>
+
+# Ternary Bonsai 27B
+
+Full 27B-class reasoning in ternary transformer weights — on everyday laptops
+
+> **\~9.4x** smaller than FP16 (ideal) | **95%** of FP16 intelligence retained | **\~26 tok/s** on an Apple M5 Pro laptop
+
+## Highlights
+
+- **\~7.2 GB** deployed footprint (down from \~54 GB FP16) — full 27B-class reasoning on a standard laptop or a single GPU
+- **95% of FP16 intelligence retained**: 80.49 average across 15 thinking-mode benchmarks — a *higher* score than the conventional IQ2_XXS build (72.73) at less than two-thirds of its footprint
+- **Retains thinking, reasoning, and agentic behavior** deep in the sub-4-bit regime, where conventional low-bit representations collapse: math within two points of full precision (93.40), coding at 85.96, agentic tool use at 74.01
+- **End-to-end ternary language weights** across embeddings, attention projections, MLP projections, and LM head, at a *true* 1.71 bits per weight — no high-precision escape hatches behind a low-bit label; the vision tower ships in compact 4-bit HQQ
+- **262K-token context** on-device, kept practical by the Qwen3.6-27B hybrid-attention backbone (\~75% linear attention) and 4-bit KV-cache quantization
+- **Custom 2-bit hybrid-attention kernels** on Apple MLX (Python, Swift) and CUDA — packed weights are consumed directly, never expanded back to FP16
+- **Ships with a DSpark speculative-decoding drafter layer** trained against the Bonsai 27B target — a lossless **1.34x** decode speedup on the CUDA serving path
+- **1-bit companion**: also available as [1-bit Bonsai 27B](https://huggingface.co/prism-ml/Bonsai-27B-mlx-1bit) (\~3.9 GB), the phone-class operating point that fits an iPhone 17 Pro Max
+
+## Resources
+
+- **[Whitepaper](https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-27b-whitepaper.pdf)** — full methodology, benchmarks, and measurement notes
+- **[Demo & examples](https://github.com/PrismML-Eng/Bonsai-demo)** — serving, benchmarking, and integrating Bonsai
+- **Low-bit kernels**: [MLX fork](https://github.com/PrismML-Eng/mlx) (Apple Silicon) · [mlx-swift fork](https://github.com/PrismML-Eng/mlx-swift) (iOS/macOS) · [llama.cpp fork](https://github.com/PrismML-Eng/llama.cpp) (CUDA)
+- **[Discord](https://discord.gg/prismml)** — join the community for support, discussion, and updates
+
+## Model Overview
+
+| Item              | Specification                                                                                    |
+| :---------------- | :----------------------------------------------------------------------------------------------- |
+| Base model        | Derived from Qwen3.6-27B, a 27B hybrid-attention causal language model (architecture unchanged)  |
+| Parameters        | \~27.3B ternary language weights (\~24.8B backbone across 64 blocks + \~2.5B embedding/LM head) + \~0.46B vision tower (27 blocks) |
+| Architecture      | Hybrid attention (\~75% linear / \~25% full attention), SwiGLU MLP, RoPE, RMSNorm                   |
+| Context length    | 262K tokens (full-context capable on-device, enabled by the predominantly linear-attention backbone) |
+| KV cache          | Near-lossless 4-bit KV quantization; the hybrid backbone grows a full-attention cache on only 16 of 64 layers (\~4.3 GB at the full 262K window) |
+| Weight format     | Ternary g128: {−1, 0, +1} weights with FP16 group-wise scaling                                    |
+| Low-bit coverage  | Embeddings, attention projections, MLP projections, LM head                                       |
+| Vision tower      | HQQ 4-bit; optional \~0.63 GB mmproj pack, loaded only for image input                             |
+| Deployed size     | **\~7.2 GB** (5.9 GB ideal at 1.71 bits/weight; see below)                                         |
+| Acceleration      | DSpark speculative-decoding drafter layer provided                                                |
+| Backends          | Apple MLX (Python, Swift) and CUDA                                                                |
+| License           | Apache 2.0                                                                                        |
+
+## Weight Representation: Ternary g128
+
+Each weight takes a value from {−1, 0, +1}, with one shared FP16 scale factor for every group of 128 weights. A ternary value carries log₂3 ≈ 1.585 bits of information, so the effective storage cost is **\~1.71 bits/weight** (ternary code + 16-bit scale amortized over 128 weights) — an idealized \~9.4x reduction vs FP16.
+
+Relative to the binary format, the extra zero state gives a more expressive weight alphabet and recovers more of the full-precision model's behavior, which makes ternary the **quality-oriented operating point** of the Bonsai 27B family.
+
+### Memory Requirement
+
+| Format              | True bits/weight | Ideal size | Deployed size | Reduction (ideal) |
+| :------------------ | ---------------: | ---------: | ------------: | ----------------: |
+| FP16 (baseline)     | 16.0             | \~54 GB     | —             | 1.0x              |
+| **Ternary g128**    | **1.71**         | **5.9 GB** | **\~7.2 GB**   | **\~9.4x**         |
+
+Today's kernels store each ternary value in a 2-bit slot (2.125 bits/weight deployed), so the deployed footprint sits above the representation's information-theoretic minimum until native ternary kernels close the gap. The deployed figure describes the language model alone — the only component that must stay resident for text inference; a negligible tail of normalization and scale parameters remains in higher precision.
+
+Unlike conventional low-bit builds — whose advertised labels understate their true average bit-width (a widely-used "2-bit" build of Qwen3.6-27B is really 2.8 bits/weight at 9.4 GB) — the Bonsai representation carries a bit-width that matches its name.
+
+### Shipped Components
+
+Two optional components ship alongside the language model (on-disk sizes from the GGUF packaging):
+
+| Component      | Pack                               | Size     | Residency                          |
+| :------------- | :--------------------------------- | -------: | :--------------------------------- |
+| Language model | 2-bit g128 slots                   | 7.17 GB  | resident                            |
+| DSpark drafter | Q4_1 (default)                     | 1.95 GB  | optional — speculative decoding     |
+| DSpark drafter | bf16 (reference)                   | 7.29 GB  | optional                            |
+| Vision tower   | mmproj HQQ 4-bit (Q8_0 container)  | 0.63 GB  | optional — multimodal input only    |
+| Vision tower   | mmproj BF16 (reference)            | 0.93 GB  | optional                            |
+
+The vision tower is usually offloaded: it sits outside the accelerator's resident budget and is loaded only when an image actually arrives, so text-only serving never pays for it. A group-64 ternary pack (7.59 GB) is also published, matching the 64-value-group Q2_0 packing in llama.cpp — the same native g128 representation with each scale repeated per 64-value block.
+
+### MLX Packaging
+
+The published MLX pack measures **8.49 GB** on disk (safetensors, vision tower bundled in the same file). MLX's grouped low-bit format stores both a scale and a bias per group; Bonsai's scale-only weights are packed by setting `s_mlx = 2·s_g` and `bias = −s_g`, which reproduces ±s_g exactly — the bias carries no new information — but stores two FP16 values per group where the native format stores one. The effective rate is therefore 2.25 bits/weight versus the deployed 2.125. This is a current MLX limitation; once MLX supports scale-only group formats, the MLX pack matches the native rate.
+
+### Peak Memory at Context
+
+What a device must actually accommodate is *peak* memory — weights plus KV cache plus activations and runtime buffers (\~1.3 GB across backends). Measured, language model only, no KV-cache compression (sizes in decimal GB; the Q4_K_XL row is derived from its weight footprint plus the same measured cache-and-overhead build-up, all other rows directly measured):
+
+| Build                                | Weights | 4K ctx | 10K ctx | 100K ctx |
+| :----------------------------------- | ------: | -----: | ------: | -------: |
+| **Ternary Bonsai (llama.cpp Q2_0)**  | 7.15    | 8.4    | 8.7     | 14.7     |
+| **Ternary Bonsai (MLX 2-bit)**       | 7.57    | 9.2    | 9.6     | 15.5     |
+| Qwen3.6-27B "4-bit" (Q4_K_XL)        | 17.6    | 19.2   | 19.6    | 25.6     |
+| 27B 16-bit (GGUF bf16)               | 51.25   | 52.6   | 53.3    | 59.3     |
+
+The ternary build holds a **100K-token context at 14.7–15.5 GB without any KV-cache compression** — a budget that fits mainstream laptops outright; the conventional Q4_K_XL build needs \~25.6 GB before the first long document is loaded. These peaks are the conservative case, with the cache left at FP16. Enabling the 4-bit KV cache shrinks the context-dependent term \~4x: the ternary build's 100K peak drops to \~10.1 GB, and the full 262K window fits in \~12.8 GB peak.
+
+## Best Practices
+
+### Generation Parameters
+
+| Parameter   | Suggested |
+| :---------- | :-------- |
+| Temperature | 0.7       |
+| Top-p       | 0.95      |
+| Top-k       | 20        |
+
+These are the settings used for all reported benchmark results (thinking mode).
+
+### System Prompt
+
+You can use a simple system prompt such as:
+
+```
+You are a helpful assistant
+```
+
+## Quickstart
+
+### MLX (Apple Silicon — Mac)
+
+Use the [Bonsai-demo repo](https://github.com/PrismML-Eng/Bonsai-demo) to run ternary Bonsai 27B on MacBook.
+
+### CUDA (NVIDIA GPUs)
+
+CUDA inference uses the GGUF pack of the same weights: fused 2-bit GEMM kernels in our [llama.cpp fork](https://github.com/PrismML-Eng/llama.cpp) unpack the ternary codes and apply the group-wise scales inside the matrix multiply — the model is never expanded into a dense FP16 tensor in memory. See **[Ternary-Bonsai-27B-gguf](https://huggingface.co/prism-ml/Ternary-Bonsai-27B-gguf)** for the full CUDA, Metal, and server quickstarts.
+
+> **Deploying to a phone?** The ternary build (\~7.2 GB) exceeds the \~6 GB per-app iOS memory budget and is laptop/GPU-only. Use the [1-bit companion](https://huggingface.co/prism-ml/Bonsai-27B-mlx-1bit) (\~3.9 GB), which fits an iPhone 17 Pro Max.
+
+## Cross-Platform Throughput
+
+`tg128` is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); `pp512` is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s. Rows are measured with llama.cpp (Metal/CUDA, custom low-bit kernels) on the GGUF pack of the same weights.
+
+| Platform                     | Footprint | TG128 (tok/s) | PP512 (tok/s) |
+| :--------------------------- | --------: | ------------: | ------------: |
+| Laptop (Apple M5 Max, Metal) | 7.2 GB    | 44.0          | 830           |
+| Laptop (Apple M5 Pro, Metal) | 7.2 GB    | 26.2          | 393           |
+| Laptop (Apple M4 Pro, Metal) | 7.2 GB    | 18.0          | 125           |
+| Single GPU (H100, CUDA)      | 7.2 GB    | 98.0          | 2596          |
+
+On the laptop, the FP16 baseline (\~54 GB) and even conventional "4-bit" builds (17.6 GB) do not fit at all — the meaningful statement is not a speedup ratio but that a 27B model runs interactively on an everyday laptop. The measured decode streams \~186 GB/s of weights on the M5 Pro, confirming the memory-bandwidth-dominated profile that the low-bit representation is built to exploit. The H100 row is the exception that proves the rule: at batch size 1 a datacenter GPU is limited by kernel-launch and synchronization latency rather than weight bandwidth, so the ternary and binary variants converge there (98 vs 104.8 tok/s) despite their \~1.9x difference in bytes per step.
+
+## Speculative Decoding: DSpark
+
+Ternary Bonsai 27B ships with a **DSpark** drafter layer trained against the low-bit target — a semi-autoregressive drafter with confidence-scheduled verification. Speculative decoding is lossless: verification preserves the target distribution exactly, so accepted tokens are indistinguishable from ordinary generation.
+
+The drafter is a compact **six-layer block-parallel transformer** conditioned on hidden states tapped from five evenly spaced layers of the target; its drafter-unique weights add roughly **0.5 GB at serving precision** (embeddings and output head are shared with the resident target). It follows the DSpark recipe with a diffusion-flavored block-denoising objective, survival-probability-weighted distillation, per-source-normalized hidden-state taps, and a draft block size chosen from a measured verify-cost model of the serving stack. The drafter ships 4-bit quantized — the \~1.95 GB Q4_1 pack is the default; it drafts faster than the bf16 reference at essentially unchanged draft quality, and because verification preserves the target distribution exactly, drafter precision affects only speed, never output quality.
+
+On the CUDA serving path the drafter is a measured net win — an accepted length of τ ≈ 3.7 at draft depth k = 4 turns into a **1.34x** end-to-end decode speedup on H100 (98 → 131.8 tok/s). On Apple Silicon the batch-1 verification pass does not yet amortize, so the drafter layer is not enabled by default on-device.
+
+## Benchmarks
+
+Evaluated with EvalScope + vLLM on NVIDIA H100 under identical infrastructure, decoding, and scoring, in **thinking mode** — where the model's full reasoning is exercised and the sub-4-bit collapse of conventional methods is most visible. 15 benchmarks across six skill categories. For cross-family context the table also includes Gemma-4-31B, a model of the same capability tier, with its conventional low-bit builds — the collapse below 4 bits is a property of the methods, not of one base model. Bit-widths are true averages; "vs FP16" is relative to the Qwen3.6-27B FP16 reference.
+
+| Variant                                                                    | True bpw | Footprint  | Thinking avg | vs FP16    |
+| :-------------------------------------------------------------------------- | -------: | ---------: | -----------: | ---------: |
+| Qwen3.6-27B FP16                                                            | 16.0     | 54 GB      | 85.07        | 100%       |
+| Qwen3.6-27B Q4_K_XL ("4-bit")                                               | 5.2      | 17.6 GB    | 84.99        | 99.9%      |
+| Qwen3.6-27B IQ2_XXS ("2-bit")                                                | 2.8      | 9.4 GB     | 72.73        | 85.5%      |
+| Gemma-4-31B FP16                                                            | 16.0     | 61.5 GB    | 84.58        | 99.4%      |
+| Gemma-4-31B QAT ("4-bit")                                                   | 6.0      | 23.3 GB    | 83.41        | 98.0%      |
+| Gemma-4-31B Q2_K_XL ("2-bit")                                               | 3.0      | 11.8 GB    | 73.31        | 86.2%      |
+| **Ternary Bonsai 27B**                                                      | **1.71** | **5.9 GB** | **80.49**    | **94.6%**  |
+| [1-bit Bonsai 27B](https://huggingface.co/prism-ml/Bonsai-27B-mlx-1bit)     | 1.125    | 3.9 GB     | 76.11        | 89.5%      |
+
+At 5.9 GB, Ternary Bonsai 27B outscores both sub-4-bit conventional builds by more than seven points at one-half to two-thirds of their size.
+
+The aggregate gap also understates *how* the conventional builds fail: their degradation is selective, concentrated on the benchmarks that demand sustained chains of reasoning. IQ2_XXS falls to 57.5 on AIME26 and 56.4 on LiveCodeBench while still scoring 88.93 on MMLU-Redux — which is why casual testing misses the collapse. Ternary Bonsai holds exactly these benchmarks, keeping AIME at 87.5–90.8 and LiveCodeBench at 82.8.
+
+### By Skill Category
+
+| Category                | Benchmarks                          | FP16  | Ternary 27B |
+| :---------------------- | :---------------------------------- | ----: | ----------: |
+| Knowledge & reasoning   | MMLU-Redux, MuSR                    | 83.15 | 76.96       |
+| Math                    | GSM8K, MATH-500, AIME25, AIME26     | 95.33 | 93.40       |
+| Coding                  | HumanEval+, MBPP+, LiveCodeBench    | 88.74 | 85.96       |
+| Instruction following   | IFEval, IFBench                     | 78.47 | 71.77       |
+| Agentic / tool calling  | BFCL v3, τ²-Bench                   | 80.00 | 74.01       |
+| Vision                  | MMMU-Pro, OCR Bench v2              | 72.61 | 65.19       |
+| **Overall (15)**        |                                     | **85.07** | **80.49** |
+
+The reasoning backbone comes through intact: math stays within two points of full precision (93.40), coding at 85.96, and the ternary model spends its extra footprint to hold the most demanding categories — agentic tool use at 74.01 and vision at 65.19 — the behaviors that conventional sub-4-bit representations lose first.
+
+### Full Per-Benchmark Results
+
+<details>
+<summary>Expand full per-benchmark results (thinking mode)</summary>
+
+| Benchmark              | FP16  | Ternary 27B |
+| :--------------------- | ----: | ----------: |
+| MMLU-Redux             | 93.42 | 88.05       |
+| MuSR                   | 72.88 | 65.87       |
+| GSM8K                  | 95.30 | 96.06       |
+| MATH-500               | 99.40 | 99.20       |
+| AIME25                 | 93.29 | 90.84       |
+| AIME26                 | 93.33 | 87.50       |
+| HumanEval+             | 95.12 | 93.90       |
+| MBPP+                  | 83.33 | 81.22       |
+| LiveCodeBench          | 87.77 | 82.75       |
+| IFEval                 | 88.91 | 85.03       |
+| IFBench (prompt-loose) | 68.03 | 58.50       |
+| BFCL v3                | 77.10 | 74.41       |
+| τ²-Bench               | 82.90 | 73.61       |
+| MMMU-Pro               | 79.94 | 68.96       |
+| OCR Bench v2           | 65.28 | 61.42       |
+| **Average (15)**       | **85.07** | **80.49** |
+
+</details>
+
+## Intelligence Density
+
+Intelligence density captures the ratio of a model's capability to its deployed size:
+
+```
+D = -log2(1 - score/100) / size_GB
+```
+
+| Variant                                                                    | Size (GB) | Benchmark avg | Intelligence Density (1/GB) |
+| :-------------------------------------------------------------------------- | --------: | -----------: | --------------------------: |
+| [1-bit Bonsai 27B](https://huggingface.co/prism-ml/Bonsai-27B-mlx-1bit)     | 3.9       | 76.11        | 0.530                       |
+| **Ternary Bonsai 27B**                                                      | **5.9**   | 80.49        | **0.400**                   |
+| Qwen3.6-27B IQ2_XXS                                                          | 9.4       | 72.73        | 0.199                       |
+| Gemma-4-31B Q2_K_XL                                                         | 11.8      | 73.31        | 0.162                       |
+| Qwen3.6-27B Q4_K_XL                                                         | 17.6      | 84.99        | 0.155                       |
+| Gemma-4-31B QAT                                                             | 23.3      | 83.41        | 0.111                       |
+| Qwen3.6-27B FP16                                                            | 54        | 85.07        | 0.051                       |
+| Gemma-4-31B FP16                                                            | 61.5      | 84.58        | 0.044                       |
+
+Ternary Bonsai 27B delivers **2x** the density of the densest conventional build (IQ2_XXS at 0.199) and nearly **8x** FP16 — no conventional build of Qwen3.6-27B or Gemma-4-31B exceeds 0.2. Each stored gigabyte is translated into far more usable intelligence.
+
+## Use Cases
+
+- **Laptop-local 27B agents**: full 27B reasoning and tool use on a standard laptop at \~26 tok/s, with the 262K context available for long-document analysis, full-repository code work, and other tasks that depend on holding a large working set in context
+- **Privacy-sensitive and offline settings**: on-device execution keeps prompts and data on the device by construction, and works with intermittent or no connectivity
+- **Single-GPU and commodity-GPU serving**: 27B-class quality from a single consumer or entry-level datacenter GPU, with headroom for larger batches, longer contexts, or co-resident models — combined with the KV-cache quantization, high-throughput serving and long-context document analysis become practical on a single 24 GB GPU
+- **Quality-first low-bit deployment**: when the deployment target has laptop-class memory or better, ternary is the operating point that retains the most of the full-precision model's behavior
+
+## Limitations
+
+- **The quality–footprint trade-off**: the ternary model retains 94.6% of the full-precision average, and the gap is modest and predictable — the reasoning core (math, coding) stays within a few points of baseline, with the difference concentrated in the most demanding categories; deployments that need the last few points of accuracy can still reach for the full-precision model where its footprint is not a constraint
+- **Does not fit a phone**: at \~7.2 GB the ternary build exceeds the \~6 GB per-app iOS memory budget; use the [1-bit companion](https://huggingface.co/prism-ml/Bonsai-27B-mlx-1bit) for phone deployment
+- **Served in 2-bit slots today**: the deployed footprint (\~7.2 GB) sits above the representation's \~5.9 GB native target; native ternary kernels are an active engineering target and would return the remaining bandwidth and footprint advantage directly as latency and energy improvements
+- **Agentic coding** (long-horizon, multi-file, run-test-and-repair workflows) is not yet a strong target of this release; a Bonsai 27B variant tuned for agentic coding is next on the roadmap
+- **KV compression headroom**: this release standardizes on a 4-bit KV cache; Bonsai's tolerance to KV-cache error grows with context length, and early results show the key cache can be pushed toward the sub-2-bit regime — a path to still longer contexts within a fixed device-memory budget
+
+## Citation
+
+If you use Ternary Bonsai 27B, please cite:
+
+```bibtex
+@techreport{bonsai27b,
+    title   = {Bonsai 27B: Full 27B-Class Reasoning in Binary and Ternary
+               Transformer Weights --- on Laptops and Phones},
+    author  = {Prism ML},
+    year    = {2026},
+    month   = {July},
+    url     = {https://prismml.com}
+}
+```
+
+## Contact
+
+For questions, feedback, or collaboration inquiries: **contact@prismml.com**

chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if (preserve_thinking is defined and preserve_thinking is true) or (loop.index0 > ns.last_query_index) %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

config.json

@@ -0,0 +1,149 @@
+{
+  "architectures": [
+    "Qwen3_5ForConditionalGeneration"
+  ],
+  "image_token_id": 248056,
+  "language_model_only": false,
+  "model_type": "qwen3_5",
+  "text_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "attn_output_gate": true,
+    "bos_token_id": 248044,
+    "dtype": "bfloat16",
+    "eos_token_id": 248046,
+    "full_attention_interval": 4,
+    "head_dim": 256,
+    "hidden_act": "silu",
+    "hidden_size": 5120,
+    "initializer_range": 0.02,
+    "intermediate_size": 17408,
+    "layer_types": [
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention"
+    ],
+    "linear_conv_kernel_dim": 4,
+    "linear_key_head_dim": 128,
+    "linear_num_key_heads": 16,
+    "linear_num_value_heads": 48,
+    "linear_value_head_dim": 128,
+    "mamba_ssm_dtype": "float32",
+    "max_position_embeddings": 262144,
+    "model_type": "qwen3_5_text",
+    "mtp_num_hidden_layers": 1,
+    "mtp_use_dedicated_embeddings": false,
+    "num_attention_heads": 24,
+    "num_hidden_layers": 64,
+    "num_key_value_heads": 4,
+    "output_gate_type": "swish",
+    "pad_token_id": null,
+    "partial_rotary_factor": 0.25,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "mrope_interleaved": true,
+      "mrope_section": [
+        11,
+        11,
+        10
+      ],
+      "partial_rotary_factor": 0.25,
+      "rope_theta": 10000000,
+      "rope_type": "default"
+    },
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "vocab_size": 248320,
+    "quantization": {
+      "group_size": 128,
+      "bits": 2
+    }
+  },
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.1",
+  "video_token_id": 248057,
+  "vision_config": {
+    "deepstack_visual_indexes": [],
+    "depth": 27,
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1152,
+    "in_channels": 3,
+    "initializer_range": 0.02,
+    "intermediate_size": 4304,
+    "model_type": "qwen3_5",
+    "num_heads": 16,
+    "num_position_embeddings": 2304,
+    "out_hidden_size": 5120,
+    "patch_size": 16,
+    "spatial_merge_size": 2,
+    "temporal_patch_size": 2
+  },
+  "vision_end_token_id": 248054,
+  "vision_start_token_id": 248053,
+  "quantization": {
+    "group_size": 128,
+    "bits": 2
+  },
+  "eos_token_id": 248046
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8acd4597893ea7004e2d7336c3cf6e3157b8896592bbcf066db004021e45846b
+size 8490785104

preprocessor_config.json

@@ -0,0 +1,21 @@
+{
+    "size": {
+        "longest_edge": 16777216,
+        "shortest_edge": 65536
+    },
+    "patch_size": 16,
+    "temporal_patch_size": 2,
+    "merge_size": 2,
+    "image_mean": [
+        0.5,
+        0.5,
+        0.5
+    ],
+    "image_std": [
+        0.5,
+        0.5,
+        0.5
+    ],
+    "processor_class": "Qwen3VLProcessor",
+    "image_processor_type": "Qwen2VLImageProcessorFast"
+}

processor_config.json

@@ -0,0 +1,3 @@
+{
+  "processor_class": "Qwen3VLProcessor"
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87a7830d63fcf43bf241c3c5242e96e62dd3fdc29224ca26fed8ea333db72de4
+size 19989343

tokenizer_config.json

@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "split_special_tokens": false,
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

video_preprocessor_config.json

@@ -0,0 +1,21 @@
+{
+    "size": {
+        "longest_edge": 25165824,
+        "shortest_edge": 4096
+    },
+    "patch_size": 16,
+    "temporal_patch_size": 2,
+    "merge_size": 2,
+    "image_mean": [
+        0.5,
+        0.5,
+        0.5
+    ],
+    "image_std": [
+        0.5,
+        0.5,
+        0.5
+    ],
+    "processor_class": "Qwen3VLProcessor",
+    "video_processor_type": "Qwen3VLVideoProcessor"
+}