leonsarmiento/Occamy-1.0-6bit-XL-mlx
Occamy-1.0 by Accio-Lab — an agentic co-worker built on Qwen3.6-35B-A3B, quantized for Apple Silicon using the BaseQuant_XL 6/8-bit recipe.
Occamy-1.0 is purpose-built for long-horizon, stateful agent work: coordinated use of search, code, tools, files, structured APIs, and productivity software — with reliable execution, persistent state tracking, recovery, and follow-through. Best among 35B-A3B peers on most agentic benchmarks: 82.20 Claw-Eval avg (+12.7 over its Qwen3.6 base), 49.16 WildClawBench, 59.0 Terminal-Bench 2.1, 65.40 BFCL v4, 91.53 IFEval.
This is a full multimodal build — the vision tower is preserved. (Accio froze the vision encoder/projector during post-training, so vision behavior matches the Qwen3.6 base.)
About XL Quantization
BaseQuant_XL is a fully data-agnostic, static quantization. No calibration dataset, no sensitivity analysis, no importance matrix. Precision is allocated purely by architectural role — routing-critical layers get higher precision, bulk expert parameters get lower precision. The result is a transparent, faithful capture of the source model.
Data-dependent calibration quantizations (iMatrix, AWQ, GPTQ, oQ, oQ4e, etc.) use a calibration set to guide bit allocation. This can produce a skewed representation of the model: domains well-represented in the calibration data (English, popular topics, public or leaked benchmarks) are preserved better, while underrepresented domains (non-English languages, niche use cases, your own data) are preserved worse. XL avoids this trade-off entirely — it generalizes honestly because it is never fit to any particular data distribution.
Quickstart
pip install -U mlx-vlm
python -m mlx_vlm.generate --model leonsarmiento/Occamy-1.0-6bit-XL-mlx --max-tokens 4096 --temperature 1.0 --top-p 0.95 --prompt "Search the repo for where errors from the payment API are swallowed, then propose a fix."
Works with LM Studio — vision mmproj included. Thinking mode is on by default (emits <think>...</think>).
Quantization Strategy
BaseQuant_XL recipe — precision is allocated by layer importance, not applied uniformly:
| Layers | Bits | Rationale |
|---|---|---|
mlp.gate (router), shared_expert_gate, lm_head, shared_expert |
bf16 | Routing decisions and output projection — any quantization noise here causes expert misrouting or output degradation |
embed_tokens, self_attn, linear_attn |
8-bit | Every-token layers — near-lossless, attention quality preserved |
vision_tower, switch_mlp (routed experts) |
6-bit | Bulk parameters — 256 experts with only 8 active per token; redundancy absorbs quantization noise. 6-bit is the sweet spot for routed experts (higher bits can cause overthinking) |
- Bits per weight: 6.808 · Total size: 29.9 GB · Group size: 64
Notes specific to this build:
- Source stores routed experts as individual per-expert tensors (
experts.N.{gate,up,down}_proj, N = 0–255) — stockmlx_vlm0.7.0 sanitize stacks them intoswitch_mlp3D tensors at load; no custom sanitize needed. - Source ships the vision tower as a separate
model-visual.safetensors(transformers 5.x split layout) — picked up natively by the loader. - No MTP layer in the source (
mtp_num_hidden_layers: 0).
Recommended Inference Parameters
Per the source model card:
| Parameter | Value |
|---|---|
temperature |
1.0 |
top_p |
0.95 |
top_k |
20 |
presence_penalty |
1.5 |
max_tokens |
32,768 |
reasoning_parser |
qwen3 |
tool_call_parser |
qwen3_coder |
Thinking is on by default (<think>...</think> before the answer); with a reasoning parser enabled the chain-of-thought is returned in a separate reasoning_content field. Context: 262,144 native (SFT-trained at 131,072).
Model Overview
| Property | Value |
|---|---|
| Architecture | Qwen3.6-family MoE (35B-A3B) + native vision encoder |
| Parameters | 35B total / ~3B active per token |
| Experts | 256 (8 routed + 1 shared) |
| Attention | Hybrid — 30 linear_attn + 10 full attention (40 layers) |
| Modalities | text, image, video → text |
| Context window | 262,144 tokens native |
| Thinking | <think>...</think> — on by default, preserve_thinking supported |
| Tool calling | XML-style (<tool_call><function=...><parameter=...>) |
| License | Apache 2.0 |
Source Model Benchmarks (from Accio-Lab)
| Benchmark | Occamy-1.0 |
|---|---|
| Claw-Eval (avg) | 82.20 |
| WildClawBench | 49.16 |
| AutomationBench Pass@1 | 27.60 |
| Terminal-Bench 2.1 | 59.00 |
| BFCL v4 | 65.40 |
| IFEval | 91.53 |
Trained openly via Accio's Dressage framework: full-parameter SFT (~15,000 trajectories / 403.3M tokens), HDPO, a two-expert merge (Marathon for sustained execution + Sprint for broad agentic capability), and SAO. Weights, dataset (occamy-data-1.0), and training code are all released.