Downloads · 30 days
20
21% of all-time downloads
Robot-Haus/Qwen3.6-35B-A3B-MAYA-oQ3.5e-mtp
Qwen3.6-35B-A3B-MAYA-oQ3.5e-mtp is a machine learning model from Robot-Haus. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for mlx. The card lists the license as apache-2.0.
Downloads · 30 days
20
21% of all-time downloads
All-time downloads
95
Public
Parameters
36B
18.6 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors18.6 GB · 100%
How the weights are stored.
U3235.5B · 99%
From the Hugging Face model README
Monitoring, Analysis and Yield Agent.
The MAYA name marks models calibrated to run as a local AI assistant and is not just a benchmark-chasing quantization. The custom imatrix corpus was built to preserve the weight channels that matter : multi-step agentic reasoning, tool use, network monitoring, expressive professional prose, the ability to express a personality, and the kind of organizational intelligence that lets an AI feel like a real collaborator rather than merely a command executor. The weights provide the capability; the system prompt is what instantiates the specific character on top. If you're building a personal assistant with a defined persona and want the model to have the headroom to inhabit it, and still do real work, then this series was built with that in mind.
A self-quantized MLX checkpoint of Qwen/Qwen3.6-35B-A3B with native MTP (Lightning MTP) support. This is the Q3.5e bf16 variant — the smallest checkpoint in the MAYA series with full MTP tensors intact.
Companion models:
- Recommended for M1/M2:
Robot-Haus/Qwen3.6-35B-A3B-MAYA-oQ3.5e-fp16-mtp— fp16 companion; ~44% faster prefill at long context on M1/M2, and my personal production pick for code and agent tasks.- For better code benchmarks:
Robot-Haus/Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp— Q4 fp16; slightly better LiveCodeBench (+5pp vs. Q3.5e), at the cost of a larger footprint.- Q4 bf16:
Robot-Haus/Qwen3.6-35B-A3B-MAYA-oQ4e-mtp
Qwen/Qwen3.6-35B-A3B — full-precision BF16 directly from Qwen's official release (vision tower + MTP tensors intact)oQ3.5e, bf16 dtype, "Preserve MTP weights" enabled — quantized on oMLX v0.5.5mtp.* tensors preserved at quantization time. Note: mtp_enabled defaults to false for newly loaded models in oMLX — enable via the admin API or model settings UI.HuggingFace has been flooded with Qwen3.x uploads claiming to have MTP retained and the vision tower intact. In my experience, that was rarely the case as conversion pipelines silently strip those components, and there's no standard way to verify from the outside without opening the checkpoint yourself. I started quantizing from the official original weights myself to be certain.
This model was quantized directly from Qwen/Qwen3.6-35B-A3B : the official full-precision source. MTP and vision tensors were verified after quantization by inspecting the safetensors index directly.
The fp16 companion (oQ3.5e-fp16-mtp) is the recommended choice for M1/M2 users. Apple Silicon in that generation has native fp16 compute units that produce a real, large TTFT advantage — measured at ~44% faster prefill at pp32768 in benchmarks. The bf16 dtype (this checkpoint) is the natural choice on M3/M4, where bf16 throughput was significantly improved, or on CUDA hardware.
Intelligence benchmark accuracy is essentially identical between dtypes — fp16 and bf16 variants agree within noise on 7 of 8 benchmarks in cross-dtype safety testing.
All single-request runs: Apple M1 Ultra 128GB, Code (Python) context, Engine: Auto, oMLX v0.5.5.
| Context | MTP on TTFT | MTP on tok/s | MTP off TTFT | MTP off tok/s |
|---|---|---|---|---|
| pp1024 | 1.28s | 66.6 | 1.30s | 67.1 |
| pp4096 | 4.74s | 64.8 | 4.74s | 66.8 |
| pp8192 | 10.1s | 61.1 | 10.0s | 61.0 |
| pp16384 | 23.1s | 57.1 | 23.0s | 56.3 |
| pp32768 | 57.2s | 48.3 | 57.3s | 54.3 |
| pp65536 | 159.5s | 34.8 | 159.6s | 43.3 |
pp32768 highlighted — closest proxy to real working context.
MTP guidance: At bf16, Q3.5e shows a net negative MTP effect at pp32768 (−11% vs. fp16's +3–32% gain). MTP off is the recommendation for bf16 at most context lengths. Turn MTP on only for short-context (<pp4096) single-request workloads where the drafting overhead hasn't yet accumulated.
| Batch | MTP off | MTP on |
|---|---|---|
| 1× | 67.1 tok/s | 66.6 tok/s |
| 2× | 103.6 (1.54×) | 107.1 (1.61×) |
| 4× | 141.9 (2.11×) | 143.8 (2.16×) |
| 8× | 178.0 (2.65×) | — |
Intelligence benchmarks were run on the fp16 companion (oQ3.5e-fp16-mtp); cross-dtype safety testing confirmed fp16 and bf16 agree within noise on 7 of 8 benchmarks, so these numbers apply to this bf16 checkpoint as well. oMLX v0.5.5, Thinking Mode off, batch 8×, Apple M1 Ultra 128GB.
| Model | MMLU | MMLU-Pro | HellaSwag | TruthfulQA | GSM8K | HumanEval | MBPP | LiveCodeBench |
|---|---|---|---|---|---|---|---|---|
| oQ3.5e (this, ~fp16 equivalent) | 81.2% | 65.7% | 92.0% | 84.1% | 90.0% | 92.1% | 83.5% | 44.0% |
| oQ4e-fp16 | 80.8% | 59.3% | 93.0% | 85.7% | 91.0% | 90.2% | 85.5% | 49.0% |
| oQ4e-BF16 | 80.9% | 60.0% | 93.0% | 85.6% | 91.0% | 91.5% | 85.5% | 53.0% |
| oQ5e-fp16 ⚠️ | 80.2% | 62.0% | 93.5% | 86.1% | 90.0% | 76.8%* | 83.0% | 54.0% |
| 4bit-DWQ (prior prod) | 80.9% | 60.0% | 94.0% | 84.8% | 92.0% | 90.9% | 83.5% | 43.0% |
* oQ5e-fp16 HumanEval regression confirmed on independent rerun. Disqualified as production candidate.
The headline finding from the full battery: oQ3.5e is simultaneously the smallest and the highest-accuracy variant on MMLU-Pro and HumanEval, outperforming every larger bit depth on those benchmarks while maintaining a speed advantage at extreme context lengths. The Q4e variants trade a small accuracy drop on MMLU-Pro and HumanEval for a meaningful +5–9pp gain on LiveCodeBench.
The goal was to supplement oMLX's stock imatrix calibration data — not replace it — by adding samples that activate the specific weight channels relevant to a local homelab agent: agentic tool orchestration, coding, shell/sysadmin, and professional writing. The stock corpus already provides strong Hermes-style chat/code/reasoning diversity for Qwen3 models; the additions target the tail of what a homelab assistant actually does day-to-day.
| Category | Samples | Source / License |
|---|---|---|
| Retained originals (chat, code, reasoning, tool-calling) | 2,166 | oMLX stock |
| Professional writing (essays, résumés, cover letters) | 311 | InferencePrince555 + own docs |
| Agentic tool-use traces | 300 | lambda/hermes-agent-reasoning-traces · Apache 2.0 |
| Shell / sysadmin | 200 | ajibawa-2023/Shell-Code-Large · MIT |
| HTML + React | 150 | Own repos + kenhktsui/github-code-permissive-sample |
| World knowledge (Wikipedia) | 150 | wikimedia/wikipedia · CC-BY-SA |
| Swift + AppleScript | 178 | Own repos + HelloImSteven dataset · MIT |
| Organizational / multi-step | 100 | Own docs + hand-authored |
| Structured instruction-writing skills | 23 | Own skill files (curated) |
| Creative | 21 | lemon07r/VellumK2T-Fiction-SFT-01 · MIT |
| Total | 3,629 | All third-party: Apache 2.0 / MIT / CC-BY-SA or equivalent |
100% MoE expert coverage confirmed in the imatrix report (all 256 experts across 123 layers, weakest expert: 37 samples vs. 16 required minimum).
mtp.* tensors preserved at quantization time (481/483 imatrix-weighted). Runtime gate mtp_enabled defaults to false — enable via PUT /admin/api/models/{model_id}/settings with {"mtp_enabled": true} or through the oMLX admin UI.Quantized with oMLX v0.5.5 on Apple M1 Ultra (128GB unified memory) · Robot-Haus