Occamy-1.0 · MLX 6-bit
Native MLX · group size 64 · 28.17 GB · text only
Model collection · MLX collection · Checkpoint explorer · Project · Paper
Candidate release — Mac Metal acceptance is pending. Native Linux MLX artifact and bounded inference checks passed. Mac inference and performance remain unverified.
Format
| Property | Value |
|---|---|
| Source | Accio-Lab/occamy-1.0 |
| Source revision | 8f8e0e58a3c9df042be1a3fa2c191fd8047acfb8 |
| Quantization | Native affine 6-bit, group size 64; directly from BF16 |
| Router / shared-expert gates | 8-bit |
| Weight files | 28,168,623,119 bytes · 28.17 GB · 26.23 GiB |
| Export and validation | mlx 0.32.2, mlx-lm 0.31.3, transformers 5.8.1 |
| Inputs | Text only; vision and MTP are not included |
The MLX family includes 8bit · 6bit · 5bit · 4bit · 3bit · mxfp8 · mxfp4 · nvfp4. Compare their sizes in the explorer. Weight file size is not the unified-memory requirement: leave room for the operating system, cache and runtime buffers. This 6-bit export has no paired quality benchmark or established speed ranking; the new 5-bit/MXFP cards report their own paired subset checks.
Run a prompt
On Apple Silicon, use the export versions below. This Mac recipe awaits Metal acceptance.
python -m pip install "mlx==0.32.2" "mlx-lm==0.31.3" "transformers==5.8.1"
python - <<'PY'
from mlx_lm import load, generate
model, tokenizer = load("Accio-Lab/occamy-1.0-MLX-6bit")
prompt = tokenizer.apply_chat_template(
[{"role": "user", "content": "Compute 2+2. Answer briefly."}],
tokenize=False, add_generation_prompt=True, enable_thinking=False,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=128))
PY
Start a local API
Start the server on your own machine:
python -m pip install "mlx==0.32.2" "mlx-lm==0.31.3" "transformers==5.8.1"
mlx_lm.server --model Accio-Lab/occamy-1.0-MLX-6bit \
--host 127.0.0.1 --port 8000 \
--chat-template-args '{"enable_thinking":false}'
In a second terminal, send a request:
curl http://127.0.0.1:8000/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model":"Accio-Lab/occamy-1.0-MLX-6bit","messages":[{"role":"user","content":"Compute 2+2. Answer briefly."}],"temperature":0,"max_tokens":128}'
Use http://127.0.0.1:8000/v1 as the base URL in an OpenAI-compatible client and Accio-Lab/occamy-1.0-MLX-6bit as the model. The stock MLX server passed two authored HTTP smoke checks on Linux. This does not establish Mac, tool-call or agent integration compatibility. The free CPU explorer provides the commands; inference runs on your hardware.
Conversion and validation
All source payload SHA256 hashes were rechecked against the previously verified pinned source. A lossless layout adapter stacks 30,720 separate expert tensors in numeric expert order into 120 groups, then invokes the official Qwen3.5 sanitizer exactly once. Quantization and serialization use native MLX APIs. Reload and inference use stock mlx-lm without an adapter.
Complete checks passed: strict stock reload, every stored floating value finite, native affine dequantization of every quantized row, and byte-identical tokenizer and chat-template files. Eight authored cached greedy fixtures covered English and Chinese text, arithmetic, JSON and multi-turn recall: 8/8 passed, with finite full-vocabulary logits at every decoding step. Two stock server HTTP fixtures passed. No code, tools, vision or full quality benchmark was evaluated in this batch.
Linux validation used MLX CUDA 12 on one NVIDIA B200. Conversion used native Linux CPU kernels. An initial header-version conflict was resolved by installing matching CUDA 12 runtime headers in the isolated validation environment; the checks were rerun. This is not a Mac Metal result.
Validation summary · Complete checks and outputs · HTTP checks · Conversion receipt · Weight hashes
License
Apache 2.0, inherited from Occamy-1.0.
Occamy checkpoints
BF16 · GGUF · FP8 · NVFP4 · MLX 8-bit · MLX 6-bit · MLX 4-bit · MLX 3-bit · MTP head
Compare file sizes, validation scope and deployment commands in the checkpoint explorer.