---
license: apache-2.0
base_model: Qwen/Qwen3.6-27B
pipeline_tag: text-generation
library_name: llama.cpp
tags:
- agent
- coding
- reasoning
- tool-use
- function-calling
- gguf
- quantized
- cuda
- metal
---
# BTL-3 Compact
### The complete 27B agent model in one 8.39 GB file
**Under 2.5 effective bits per parameter · 98.4% held-out tool-contract retention**
[Full BTL-3](https://huggingface.co/badtheorylabs/BTL-3) ·
[Runtime source](https://github.com/Badtheorylabs/BTL-3) ·
[Bad Theory Labs](https://www.badtheorylabs.com/) ·
[Discord](https://discord.gg/QJBCcB7bF)
> **Canonical behavior refresh — July 24, 2026:** the file at the existing
> `model/BTL-3-Compact-AVQ2.gguf` path now embeds the rank-8 step-160 Thinking
> Escape behavior adapter. The packed decoder and rank-32 output correction are
> unchanged. The new model SHA-256 is
> `0a4d9ddee49e5aa93586a792bd4d452ea837229d49d22e54212dde87a5c9888a`.
> Thinking remains disabled by default; thinking-enabled coding still fails its
> release gate and should be treated as experimental.
## Introducing BTL-3 Compact
BTL-3 is a 27B open-weight agent model built for agentic coding, structural
tool use, repository work, failure recovery, and long multi-turn execution.
BTL-3 Compact packages the complete text model into one **8.39 GB** native
GGUF—smaller than an 8B model stored in FP16 and under **2.5 effective artifact
bits per parameter**.
On the step-160 candidate's held-out teacher-correct tool gate, BTL-3 Compact
retained **62 of 63 behaviors**, yielding **98.4% conditional retention**.
## Overview
BTL-3 Compact is the portable native edition of BTL-3. It packages the
complete text model—including the compressed decoder, vocabulary matrices,
output correction, and behavior repair—into a single **8,392,369,600-byte
GGUF**.
The runtime consumes the packed representation directly. It does not download,
load, or reconstruct the original BF16 checkpoint.
## Highlights
- Complete Qwen3.6-27B-derived text model in **8.39 GB decimal / 7.82 GiB**.
- Full 64-layer mixed AVQ2/UniSVQ representation with targeted precision
islands and compact behavioral corrections.
- **98.4% conditional retention (62/63)** on the step-160 candidate's held-out
teacher-correct tool gate.
- Exact native CUDA and Metal execution with no persistent dense
reconstruction.
- OpenAI-compatible local server plus LM Studio and Ollama CLI integrations.
- Verified macOS arm64 package and a clearly separated DGX Spark CUDA preview.
## Model specification
| Item | Specification |
|---|---|
| Model lineage | Qwen3.6-27B → BTL-3 RL-0013 |
| Scope | Text-only coding, reasoning, and tool use |
| Layers | 64 |
| Model file | `model/BTL-3-Compact-AVQ2.gguf` |
| Deployed model bytes | 8,392,369,600 |
| Deployed model size | 8.39 GB / 7.82 GiB |
| Architectural context | 262,144 tokens |
| Packed tensors | 2,416 |
| Runtime | BTL packed llama.cpp |
| License | Apache-2.0 model, MIT runtime |
Usable context depends on device memory, KV cache, and runtime workspace.
Start with 4K on a 16 GB Mac, 16K on a 12–16 GB GPU, or 32K on a 24 GB GPU,
then increase the window after measuring headroom.
## Measured retention
The candidate gate contains 63 turns across single, parallel, sequential,
parallel-multiple, and abstention behavior. It is an internal contract-retention
evaluation, not a public frontier benchmark.
| Metric | Result |
|---|---:|
| Teacher-correct cases | **63/63** |
| BTL-3 Compact | **62/63** |
| Conditional retention | **98.4% (62/63)** |
| Single-call retention | **100% (23/23)** |
| Parallel-call retention | **100% (14/14)** |
| Parallel-multiple retention | **100% (2/2)** |
| Sequential-call retention | **87.5% (7/8)** |
| Abstention retention | **100% (16/16)** |
The complete candidate report is included in
[`evidence/thinking-escape-v3-tool-gate.json`](evidence/thinking-escape-v3-tool-gate.json).
## Native performance
| Device | Prompt processing | Generation | Runtime status |
|---|---:|---:|---|
| RTX PRO 6000 Blackwell 96 GB | **84.70 tok/s** | **43.16 tok/s** | Exact GGUF, full CUDA offload |
| Apple M2 16 GB | **2.30 tok/s** | **2.48 tok/s** | Exact GGUF, Metal compatibility smoke |
The RTX result is the mean of three native runs with a 512-token prompt and
128 generated tokens. The M2 measurement is a compatibility smoke on the
entry-level 16 GB system, not a performance projection for newer Apple
hardware.
## Runtime support
| Target | Status | Package |
|---|---|---|
| macOS arm64 / Apple Metal | **Verified** | `runtimes/supported/BTL-3-Compact-macos-arm64` |
| Linux arm64 / NVIDIA CUDA / DGX Spark | Preview | `runtimes/preview/BTL-3-Compact-linux-arm64-cuda` |
| OpenAI-compatible HTTP | **Verified** | Native `btl3-server` |
| LM Studio | **Supported through included generator** | `integrations/btl3-native` |
| Ollama CLI | **Supported through included bridge** | `integrations/ollama` |
Stock Ollama and the stock LM Studio GGUF engine do not decode AVQ2 directly.
The included integrations start or connect to BTL's native runner while
preserving the familiar client interface.
## Quickstart
> **Current reasoning-policy note:** BTL-3 Compact is released with thinking
> disabled by default. Use the supported non-thinking path for chat, coding,
> and tools. The experimental thinking override is currently discouraged
> because it can repeat procedural reasoning or fail to terminate. On the
> step-160 held-out smoke, 5/12 prompts reached `` and a final answer;
> thinking-enabled coding reached 0/3. The default no-thinking path passed
> direct-answer stopping and executable Python validation on the exact GGUF.
### Install the verified macOS package
```bash
python3 tools/install_consumer_bundle.py \
--runtime runtimes/supported/BTL-3-Compact-macos-arm64 \
--model model/BTL-3-Compact-AVQ2.gguf
```
### Start the server directly
```bash
BTL3_MODEL="$PWD/model/BTL-3-Compact-AVQ2.gguf" \
BTL3_CTX_SIZE=4096 \
runtimes/supported/BTL-3-Compact-macos-arm64/bin/btl3-server
```
### Call the OpenAI-compatible API
```bash
curl http://127.0.0.1:8080/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{
"model": "BTL-3",
"messages": [
{
"role": "user",
"content": "Write a retrying fetch helper and include tests."
}
],
"chat_template_kwargs": {"enable_thinking": false},
"stream": true
}'
```
## Representation
The release combines:
- packed AVQ2 decoder tensors;
- affine INT4 tensors and two measured INT4 demotions;
- selected higher-precision islands;
- packed embedding and output matrices;
- a rank-32 output correction;
- a compact behavior adapter.
All 2,416 tensor payloads were byte-verified during export. Native CUDA parity
was checked against the reference representation, and the exact exported GGUF
subsequently completed autoregressive generation on CUDA and Metal.
## Artifact integrity
| Artifact | Value |
|---|---|
| File | `model/BTL-3-Compact-AVQ2.gguf` |
| Bytes | `8,392,369,600` |
| SHA-256 | `0a4d9ddee49e5aa93586a792bd4d452ea837229d49d22e54212dde87a5c9888a` |
Verify the complete release directory:
```bash
shasum -a 256 -c SHA256SUMS
```
`RELEASE_MANIFEST.json` records the exact model identity and separates
verified runtimes from preview packages.
## Intended use
- local coding and debugging;
- private repository and terminal agents;
- structured single, sequential, and parallel tool use;
- offline or self-hosted OpenAI-compatible inference;
- consumer and workstation deployments where the BF16 model is impractical.
## Operational guidance
Run generated code and tool calls in a sandbox. Require explicit confirmation
before destructive, privileged, financial, or otherwise high-impact actions.
## License and citation
The model artifact is Apache-2.0. The included runtime is MIT-licensed; see
the packaged licenses and third-party notices.
```bibtex
@software{btl3_compact_2026,
title = {BTL-3 Compact: A Native 8.39 GB Agentic Coding Model},
author = {Bad Theory Labs},
year = {2026},
url = {https://huggingface.co/badtheorylabs/BTL-3-Compact}
}
```
For questions and release updates, visit
[Bad Theory Labs](https://www.badtheorylabs.com/) or join the
[community Discord](https://discord.gg/QJBCcB7bF).