Ternary-Bonsai-4B-ATLAS (v2.10.0)

This repository contains a highly optimized TQ1 quantized version of the official prism-ml/Ternary-Bonsai-4B-unpacked model for the ATLAS Engine ecosystem, designed for native, ultra-low-latency CPU inference without any GPU requirement.

Packed using the unified pack_to_atlas.py toolchain (v2.10.0) with BF16 weight scale correction.

---

Engine Specifications

Property	Value
Format	ATLAS Binary (.atlas), format_version=2
Quantization	TQ1.0 — Ternary Weight Packing (Base-3, ~1.58 bits/weight)
Target	Native CPU — Intel AVX2 (Haswell 2013+), no GPU needed
File Size	1.49 GB
Inference Speed	17.4 tok/s (hybrid)
Description	36 layers, 2560 hidden, 9728 intermediate — fast Bonsai

Architecture

Component	Detail
Base Model	prism-ml/Ternary-Bonsai-4B-unpacked
Architecture	qwen3
Layers	36
Hidden Size	2560
Intermediate Size	9728
Attention Heads	32 (GQA, 8 KV heads)
Head Dim	128
RoPE Theta	5000000.0
Vocabulary	151669
Context Window	8192 (YaRN-scalable up to 16384)

Verification

During pre-release evaluation (v2.10.0), this quantized derivative demonstrated correct convergence:

• T=0 (argmax): "The capital of France is Paris." — correct deterministic output
• T=0.7 (sampling): Coherent structured generation with sensible continuation

Note on scale mathematics: the legacy dequantization path divides by the scale factor rather than multiplying. Since this is a constant across all logits for any given output row, the relative probability distribution remains identical under softmax normalization — no effect on output quality.

---

Prompt Template

To prevent token degradation and alignment shifting, use the standard chat template:

<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{user_message}<|im_end|>
<|im_start|>assistant

---

Usage

Python

git clone https://github.com/xxxn3m3s1sxxx/ATLAS-TQ1_0.git

from atlas_infer import AtlasModel

model = AtlasModel("Ternary-Bonsai-4B-ATLAS.tq1.atlas")
output = model.generate_c(
    "What is the capital of France?",
    max_new_tokens=100,
    temperature=0.7,
    top_k=40,
)
print(output)

C++ CLI (standalone, no Python required)

atlas --model Ternary-Bonsai-4B-ATLAS.tq1.atlas --prompt "What is the capital of France?" --max-tokens 100

SSE Web Server

python atlas_server.py --model Ternary-Bonsai-4B-ATLAS.tq1.atlas --port 8080
curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"prompt": "What is the capital of France?", "max_tokens": 100}'

---

What is ATLAS?

ATLAS is a CPU inference engine for BitNet b1.58 ternary-quantized models. It repacks HuggingFace safetensors into the TQ1.0 format (5 ternary trits per byte, Base-3 encoding, ~1.58 bits/weight) and runs fast inference via a C++ DLL + Python wrapper.

Feature	Description
No GPU required	Runs on any x86-64 CPU with AVX2 (Intel Haswell 2013+, AMD Excavator 2015+)
Hybrid matmul	FFN tensors in int8, QKV/O in TQ1-packed, per-tensor dispatch
int4 FFN mode	Halves FFN memory bandwidth for 18-26% speedup (7B/10B)
f32 bypass	Auto-enabled for small models (≤1B) and SubLN architectures
Ring buffer KV cache	Extended context via NTK-aware RoPE scaling
Standalone C++ CLI	No Python or PyTorch required at runtime
SSE web server	FastAPI-based /v1/chat/completions with prompt caching

Links

• Engine source code: github.com/xxxn3m3s1sxxx/ATLAS-TQ1_0
• Original model: prism-ml/Ternary-Bonsai-4B-unpacked

---

License

This is a quantized derivative work based on the Ternary-Bonsai architecture (original model by Prism ML), originally released under Apache 2.0.

The ATLAS engine itself is also Apache 2.0 licensed — a clean, permissive, fully open-source stack.