GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
ManniX-ITA
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ManniX-ITA
Model Tree
Qwen/Qwen3.6-35B-A3B
Input Modalities
Output Modalities
Supported Functionality
wmax aggregation with the LCB + MPE classes up-weighted (1.5) → 72/256 experts dropped per layer, protecting the code-competent experts.num_experts_per_tok = 10) baked into the config → the shipped default. Pass --override-kv qwen35moe.expert_used_count=int:8 to any llama.cpp tool to A/B back to native top-8.| Benchmark | This model | Qwen3.6-35B-A3B (256e) | coder (LCB-only) |
|---|---|---|---|
| GPQA-Diamond | 0.773 | 0.833 | 0.793 |
| MATH-500 | 0.620 | 0.730 | 0.620 |
| AIME | 0.733 | 0.633 | 0.767 |
| LiveCodeBench (v6, 77q) | 0.688 | 0.714 | 0.688 |
| IFEval | 0.730 | 0.960 | 0.840 |
| HumanEval | 0.970 | 0.970 | 0.963 |
| GSM8K | 0.970 | 0.960 | 0.980 |
| ARC-Challenge | 0.944 | 0.935 | 0.933 |
| MultiPL-E | 0.840 | 0.827 | 0.670 |
| Average | 0.808 | 0.840 | 0.806 |
Highlights: best code profile of any prune — MultiPL-E 0.840 (above the teacher; +17pp over the LCB-only coder that this model supersedes), LiveCodeBench 0.688 (tied best), HumanEval 0.970. Average 0.808 sits at the LCB-coder level and within 0.03 of the full teacher.
Aggressive expert pruning makes the model verbose on open-ended reasoning — it over-thinks before answering. This is largely inherited from the base (the 256e teacher does the same on GPQA/AIME) and is bounded by the generation cap; it does not affect the code benches, which have a natural termination anchor.
Response length in characters (content + reasoning), this model vs the 256e teacher; runaway = responses > 20k chars (of 100, or 30/198 for GPQA, 30 for AIME):
| Benchmark | p50 | p90 | max | runaway | 256e runaway |
|---|---|---|---|---|---|
| GPQA | 13.8k | 58.8k | 129k | 58 | 58 (same) |
| AIME | 52.9k | 85.5k | 96k | 25 | 29 |
| IFEval | 12.1k | 56.2k | 81k | 30 | 11 |
| MATH-500 | 2.3k | 14.5k | 76k | 9 | 12 |
| GSM8K | 2.5k | 12.5k | 108k | 6 | 3 |
| ARC | 1.4k | 2.3k | 59k | 7 | 0 |
| HumanEval | 0.8k | 1.4k | 24k | 1 | 2 |
| LiveCodeBench / MultiPL-E | — code path — | tight | tight |
Reading it: GPQA/AIME verbosity is essentially the base model (58 vs 58, 25 vs 29). Only IFEval shows prune-added rumination (30 vs 11) — the trade for the code-targeted drop map. Code and math-with-boxing tasks terminate cleanly. If you want tighter output, a repetition/length penalty at serve time (or top-8 via the override above) reduces the tail.
Qwen3.6-27B-A3B-Coder-MTP-GGUF. Includes the native MTP head (speculative decoding) and a -vision mmproj for multimodal use. imatrix.dat archived in-repo.mannix/qwen3.6-27b-a3b-coder (text) and …-vision tags (with mmproj).