The MAYA series
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.
So this is simply Qwen3.6-35B-A3B quantized using a custom imatrix calibration corpus that would ensure activation of the specific weights which benefit this use case.
Qwen3.6-35B-A3B-MAYA-oQ4e-fp16-mtp
A self-quantized MLX checkpoint of
Qwen/Qwen3.6-35B-A3B with native MTP (Lightning MTP) support, built for daily use as a personal AI assistant on Apple Silicon. This is the
Q4 fp16 variant — it earns its extra size on code-intensive and multi-step reasoning tasks, with a clear edge on LiveCodeBench (+5pp vs. Q3.5e fp16).
Companion models:
- For better decode speed and a smaller footprint:
Robot-Haus/Qwen3.6-35B-A3B-MAYA-oQ3.5e-fp16-mtp — my personal production pick; handles code and agent tasks well at a smaller size with faster throughput.
- For bf16 weights at the same bit depth:
Robot-Haus/Qwen3.6-35B-A3B-MAYA-oQ4e-mtp
What this is
- Source:
Qwen/Qwen3.6-35B-A3B — full-precision BF16 directly from Qwen's official release (vision tower + MTP tensors intact, no prior stripping or reformatting)
- Quantization: oMLX oQe, level
oQ4e, fp16 dtype, "Preserve MTP weights" enabled, and quantized and benchmarked on oMLX v0.5.5; partial re-benchmark on v0.6.0
- Hardware: Apple M1 Ultra, 128GB unified memory
- Calibration: Custom homelab corpus (3,629 samples) extending oMLX's stock calibration data with agentic tool-use traces, shell/sysadmin, professional writing, Swift + AppleScript, Objective C, HTML/React, and world knowledge. Full composition below.
- MTP:
mtp.* tensors preserved at quantization time. Draft acceptance rate typically in the 74–79% range. Note: mtp_enabled defaults to false for newly loaded models in oMLX. Enable lightning MTP via the admin API or model settings UI.
Why two models: Q4 vs Q3.5e
A full quantization battery was run across Q3.5e through Q6e at both bf16 and fp16. The main tradeoff between these two:
| oQ3.5e-fp16 | oQ4e-fp16 (this model) |
|---|
| MMLU-Pro | 65.7% | 59.3% |
| HumanEval | 92.1% | 90.2% |
| LiveCodeBench | 44.0% | 49.0% |
| Decode tok/s (pp32768 MTP on) | 58.9 | 58.4 |
| TTFT (pp32768 MTP on) | 31.9s | 31.6s |
| Peak memory | ~19–24 GB | ~22–27 GB |
Q3.5e-fp16 surprises on overall accuracy (it's the highest MMLU-Pro and HumanEval of any variant tested) and has the smallest footprint. That's the production pick for most workloads. Q4e-fp16 (this model) is the better choice when the task is specifically code-intensive — the +5pp LiveCodeBench gap is consistent and meaningful.
Performance Benchmarks
All single-request runs: Apple M1 Ultra 128GB, Code (Python) context, Engine: Auto, oMLX v0.5.5.
Note: A pp8192 decode-speed anomaly was observed on oMLX 0.6.0 across multiple models. pp8192 is excluded from any 0.6.0 data; 0.5.5 numbers remain the primary reference for this model.
oMLX v0.5.5
| Context | MTP on TTFT | MTP on tok/s | MTP off TTFT | MTP off tok/s |
|---|
| pp1024 | 0.76s | 71.0 | 0.73s | 65.8 |
| pp4096 | 2.81s | 71.6 | 2.75s | 62.5 |
| pp8192 | 5.78s | 65.7 | 5.79s | 60.1 |
| pp16384 | 12.98s | 62.9 | 12.95s | 60.6 |
| pp32768 | 31.60s | 58.4 | 31.64s | 57.2 |
| pp65536 | 85.89s | 50.4 | 85.86s | 49.3 |
pp32768 highlighted — closest proxy to real working context.
Continuous Batching (pp1024 / tg128) — v0.5.5
| Batch | MTP off | MTP on |
|---|
| 1× | 65.8 tok/s | 71.0 tok/s |
| 2× | 113.9 (1.73×) | 117.8 (1.66×) |
| 4× | 155.8 (2.37×) | 162.0 (2.28×) |
| 8× | 221.2 (3.36×) | 222.1 (3.13×) |
MTP: Context-Length-Dependent Effect
Q4e-fp16 has the cleanest, most consistently positive MTP result of the whole comparison battery:
- pp1024: +7.9% | pp4096: +14.6% | pp8192: +9.3% | pp16384: +3.8% | pp32768: +2.1% | pp65536: +2.2%
MTP is positive at every single context length tested for Q4e-fp16 — including the extreme end where every bf16 variant turns flat or negative. Keep MTP on through pp65536. Turn it off above pp100K (drafting overhead compounds past that point regardless of bit depth or dtype).
Intelligence Benchmarks
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-fp16 | 81.2% | 65.7% | 92.0% | 84.1% | 90.0% | 92.1% | 83.5% | 44.0% |
| oQ4e-fp16 (this) | 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% |
| mlx-community 4bit | 81.9% | 58.7% | 94.5% | 84.8% | 93.0% | 90.2% | 82.5% | 51.0% |
* oQ5e-fp16 HumanEval regression confirmed on independent rerun (76.8%). Disqualified as production candidate.
Calibration Corpus
Built to preserve the weight channels most relevant to agentic tool use, structured professional writing, shell and systems work, and coding tasks. Extends oMLX's stock Hermes/chat/reasoning mix rather than replacing it.
| 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 |
Why a custom corpus? oMLX's imatrix calibration determines which weight channels receive precision-preserving treatment during quantization. The additions above target the specific registers this model is used for — tool orchestration, shell/systems reasoning, professional prose, and Swift/AppleScript — without discarding the general diversity that research on Qwen3 shows is already a strong calibration baseline.
Technical Notes
- MTP:
mtp.* tensors preserved at quantization. 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.
- imatrix coverage: 481/483 tensors received imatrix-weighted quantization; 100% MoE expert coverage (all 256 experts across 123 layers).
- Vision tower: Preserved from the original source weights.
- fp16 vs bf16: fp16 gives a large, real TTFT advantage on M1/M2 (~44% faster prefill at pp32768 vs. equivalent bf16). On M3/M4, bf16 is equally capable and the
oQ4e-mtp checkpoint is the natural choice.
- fp16 accuracy safety check: fp16 and bf16 agree within noise on 7 of 8 benchmarks. LiveCodeBench shows -4pp (49 vs. 53 for bf16) — plausibly sampling variance on this benchmark's 100-question set, not a regression. Full discussion in the GUIDE.
Quantized with oMLX v0.5.5 on Apple M1 Ultra (128GB unified memory) · Robot-Haus
