NovaeonStudio/Occamy-1.0-oQ8-fp16-mtp

[image]

Occamy-1.0 · oQ8-fp16-mtp — a Novaeon.Studio build

A fast, super-agentic local worker for Apple Silicon. This is an oMLX oQ8 (near-uniform 8‑bit, router fp16) quantization of `Accio-Lab/occamy-1.0` — the 35B‑A3B agentic co‑worker — repacked for the Mac and shipped with the official experimental MTP head merged in. Built, tuned, and benchmarked on an Apple M5 Max (128 GB) by [Novaeon.Studio](https://novaeon.studio).

Why this build exists: we run local agents on Apple Silicon and wanted Occamy's tool‑use reliability at 8‑bit fidelity, in oMLX‑native format, with a documented optimal serving profile — not a generic dump. Everything below is measured on our hardware, not copied.

Best for & why it's here

Best for: an alternative 35B-A3B agentic / tool-use worker for Apple Silicon when you want a second option in the same size/speed class as our seat.

Why we published it: we built and benchmarked Occamy-1.0 as a seat candidate — it scored 6/6 on our agentic probe and was briefly our fleet seat, then superseded in-class by the faster, uncensored empero distill. We keep it up as a clean, reference oMLX/oQ8 build for anyone comparing 35B-A3B options.

Highlights

• —Super‑agentic. 6/6 on our controlled tool‑use probe suite (tool selection, parallel calls, sequential tool chains, argument fidelity, correct abstention, no hallucinated tools).
• —8‑bit fidelity. Near‑uniform 8‑bit with fp16 router — the highest‑fidelity published Apple‑Silicon Occamy build.
• —Tuned for speed. Ships a documented optimal oMLX profile (ANE‑prefill + TurboQuant‑KV) that sustains throughput far into long context.
• —Vision intact. Multimodal (text + image → text); the vision tower is preserved from the base.

Measured benchmarks (Apple M5 Max 128 GB · oMLX)

All numbers below were measured on our hardware, on these exact weights, with the seat isolated (no other load) using the optimal profile.

Decode throughput vs context

[image]

_{¹ First large‑prefill compile spikes; steady‑state prefill is fast. Cold load ≈ 18–36 s (39 GB weights).}

Agentic & tool‑use — 35B‑A3B open agentic class

Occamy‑1.0 is top‑tier for tool‑use and function‑calling among same‑size (35B‑A3B) open agentic MoE models. The numbers below are the published Accio‑Lab benchmark table. Our retired Novaeon seat was a Qwen3.6‑35B‑A3B tune, so its base Qwen3.6‑35B‑A3B is included as its reference point:

[image]

Our own probe on this build — 6/6 (temp 0): tool selection, parallel tool calls, sequential tool chains, argument fidelity, abstention when no tool is needed, and no hallucinated tools. This bespoke probe was only run on this build (6/6) and our retired general seat (5/6), so it is not comparable to the public models above — the retired seat is shown purely as a local reference.

<details><summary>Internal 6‑probe detail (this build vs retired seat)</summary>

[image] </details>

• —IFEval (ours, this build): 84.58 avg — measured locally against these quantized weights (thinking off), scored with the official IFEval verifier (541 prompts, 834 instructions).

_{Base card reports IFEval 91.53; the gap reflects 8‑bit quantization plus our thinking‑off / no‑CoT eval setup. Number is real and reproducible on these weights.}

Optimal oMLX settings (figured out empirically)

We swept the tuning levers on this build. Result — because the grafted MTP head does not accelerate decode under oQ8 (below), ANE‑prefill wins here (there is no MTP dispatch overhead to fight), and TurboQuant‑KV keeps long‑context decode high:

<details><summary>Drop‑in <code>~/.omlx/model_settings.json</code> entry</summary>

{
  "max_context_window": 262144,
  "turboquant_kv_enabled": true,
  "turboquant_kv_bits": 8.0,
  "turboquant_skip_last": true,
  "qwen35_ane_prefill_enabled": true,
  "qwen35_ane_prefill_sequence_length": 2048,
  "qwen35_ane_prefill_dual_ane": true,
  "qwen35_ane_prefill_gdn": true,
  "qwen35_ane_prefill_cpu_enabled": true,
  "qwen35_oq_a8_enabled": false,
  "moe_expert_offload_enabled": false,
  "mtp_enabled": false
}

</details>

On the MTP head: we merged Accio's official experimental occamy-1.0-MTP head (converted to the MLX language_model.mtp.* layout). oMLX recognizes it (mtp_compatible: true), but under oQ8 speculative acceptance is negligible — MTP on/off decode is identical in our A/B. Accio notes the head "requires the same decode kernels for verification" and validated it only on BF16/NVFP4, so the oQ8 verifier rejects nearly every draft. The head is shipped for correctness and future‑runtime support; we serve with mtp_enabled: false.

Quickstart

# oMLX (recommended, Apple Silicon)
omlx serve NovaeonStudio/Occamy-1.0-oQ8-fp16-mtp --port 8000
# then hit the OpenAI-compatible endpoint at http://127.0.0.1:8000/v1

# mlx-vlm
pip install -U mlx-vlm
python -m mlx_vlm.generate --model NovaeonStudio/Occamy-1.0-oQ8-fp16-mtp \
  --max-tokens 4096 --temperature 1.0 --top-p 0.95 \
  --prompt "Search the repo for where payment-API errors get swallowed, then propose a fix."

Recommended inference parameters (from the base card)

temperature 1.0 · top_p 0.95 · top_k 20 · presence_penalty 1.5 · reasoning_parser qwen3 · tool_call_parser qwen3_coder. Thinking is on by default (<think>…</think>, returned in reasoning_content); pass enable_thinking: false for fast, plain instruction‑following.

How it compares in its class (35B‑A3B)

Against the other 35B‑A3B open models, Occamy‑1.0 leads its size class on the core agentic / co‑work benchmarks and is competitive on tool‑calling, coding, and instruction‑following.

[image]

Full table (best in each row in bold; this build is a quantization of Occamy-1.0, so the Occamy column reflects the base weights — our locally‑measured IFEval 84.6 and agentic 6/6 above corroborate the quant holds):

_{Numbers reproduced from the Accio‑Lab Occamy‑1.0 benchmark table, same‑size class. * official model‑card / Artificial Analysis result. Larger models (GPT‑5.6 Sol, Qwen3.8‑Max, DeepSeek V4, GLM‑5.2) are out of this size class and score higher on several tasks — see the base card. BFCL v4 and Terminal‑Bench were not re‑run locally on this quant.}

Attribution & license

• —Base model: `Accio-Lab/occamy-1.0` — Apache‑2.0. Project: accio-lab.github.io/occamy · Paper: arXiv:2609.11977 · Training framework: Dressage.
• —MTP head: `Accio-Lab/occamy-1.0-MTP` (experimental).
• —This quantization & serving profile: Novaeon.Studio, 2026. Released under Apache‑2.0, same as the base.

This is an independent community quantization. It is not endorsed by Accio‑Lab. All credit for the model's capabilities belongs to the Occamy‑1.0 authors.

@misc{novaeon2026occamyoq8,
  title  = {Occamy-1.0 oQ8-fp16-mtp: an oMLX build for Apple Silicon},
  author = {Novaeon.Studio},
  year   = {2026},
  note   = {Quantization of Accio-Lab/occamy-1.0},
  url    = {https://huggingface.co/NovaeonStudio/Occamy-1.0-oQ8-fp16-mtp}
}

_{novæon — digital business architecture + AI · <a href="https://novaeon.studio">novaeon.studio</a>}