Falcon3-3B-Base-1.58bit-ATLAS (v2.10.0)

This repository contains a highly optimized TQ1 quantized version of the official tiiuae/Falcon3-3B-Base 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	2.11 GB
Inference Speed	7.1 tok/s (hybrid)
Description	22 layers, 3072 hidden, 9216 intermediate — TII Base variant

Architecture

Component	Detail
Base Model	tiiuae/Falcon3-3B-Base
Architecture	falcon3
Layers	22
Hidden Size	3072
Intermediate Size	9216
Attention Heads	12 (GQA, 4 KV heads)
Head Dim	256
RoPE Theta	1000042.0
Vocabulary	131072
Context Window	4096 (NTK-scalable up to 8192)

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 Format

This is a Base model — it generates raw text continuation without instruction-following. Simply provide your prompt:

{prompt}

---

Usage

Python

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

from atlas_infer import AtlasModel

model = AtlasModel("Falcon3-3B-Base-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 Falcon3-3B-Base-tq1.atlas --prompt "What is the capital of France?" --max-tokens 100

SSE Web Server

python atlas_server.py --model Falcon3-3B-Base-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: tiiuae/Falcon3-3B-Base

---

License & Usage Restrictions

This is a quantized derivative work based on the Falcon3 series (original model by the Technology Innovation Institute (TII)), originally released under the Falcon-LLM License.

By downloading or utilizing this file, you agree to be bound by the Falcon-LLM License:

1. Attribution: Any usage or secondary deployment must credit the Technology Innovation Institute (TII).
2. Non-Commercial & Small Commercial Use: Free for academic research, personal projects, and commercial entities with annual revenue under $1,000,000 USD.
3. Commercial Hosting: Entities intending to provide shared, managed hosting of the model or its derivatives as a service must enter into a separate license arrangement with TII.

Disclaimer: This quantized file is provided "as-is". The ATLAS engine itself is Apache 2.0 licensed.