GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. 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/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
| Source | Weight | Role |
|---|---|---|
| Qwen/Qwen3.8-27B | base | base, tokenizer, chat template, MTP head, vision tower |
| Qwen/Qwen3.6-27B | task base | deltas are computed from here |
| rico03/Qwen3.6-27B-Claude-Opus-Reasoning-Distilled | 0.40 | general capability |
| ValiantLabs/Qwen3.6-27B-Esper3.1 | 0.35 | code + reasoning |
| kai-os/Qwen3.6-27b-Opus4.6-reasoning | 0.25 | reasoning anchor (LoRA applied to Qwen3.6, then merged) |
Method: omnimerge_v2 (DARE-TIES base + OBIM-lite + DAREx q + EMR election). Density 0.53, DAREx q 0.75, seed 42. mlp.gate_proj / up_proj / down_proj and mtp. are passed through from the base.
Every source was fine-tuned from Qwen3.6, so each delta must be taken against 3.6 and then applied to 3.8. Merging without --task-base computes delta = source − 3.8, which embeds the inverse of the 3.6→3.8 generational upgrade and drags the result backwards — and on the 7 embedding rows 3.8 adds, it would compute (3.6 padding) − (3.8 audio embedding) and corrupt them.
Proof the transplant landed, from verify_merge_artifact.py on a sampled q_proj:
markdown
||out − Qwen3.8|| = 1.09 <- output sits on the 3.8 base||out − Qwen3.6|| = 67.78
Vocabulary. vocab.json and merges.txt are sha256-identical between Qwen3.6-27B and Qwen3.8-27B. 3.8 adds 7 tokens purely additively at ids 248070–248076 (audio/TTS); no id is reassigned or dropped. All three sources match the 3.8 vocab exactly — their merges differ only in serialisation ("Ġ Ġ" vs ["Ġ","Ġ"]).
Vision tower. All 333 model.visual.* tensors (167 weights + 166 non-weights) are bit-identical between the merge and the 3.8 base, so the stock Qwen3.8 projector is the correct mmproj.
MTP head is deliberately not merged. All three sources carry the 15 mtp.* tensors, so the default would apply a 3.6-trained head's delta to 3.8's head. MTP affects draft-acceptance rate only — it costs decode speed, silently. 3.8's head is preserved verbatim and appears in the GGUF as blk.64.* (blocks 0..64, 866 tensors). MERGE_MTP=1 merges it instead.
Sampled cohort (recommended: temperature 0.6 / top-p 0.95 / top-k 20, do_sample=true),
Q6_K, llama.cpp, identical benches and sampler across every cell.
Comparable to each other only — do not pool with greedy-decode results from any other card. The greedy GPQA table further down is a separate cohort: its rows must never be merged into these.
| Benchmark | previous imatrix | AtomicChat imatrix | Δ |
|---|---|---|---|
| HumanEval (thinking) | 0.982 | 0.976 | -0.60 pp |
| IFEval (100) | 0.960 | 0.940 | -2.00 pp |
| LiveCodeBench (v6, 77q) | 0.883 | 0.883 | ±0.00 |
| MultiPL-E (100) | 0.887 | 0.893 | +0.60 pp |
The published ladder had been calibrated on a 128-chunk in-repo corpus, and six K-tiers
(Q6_K_L, Q6_K, Q5_K_L, Q5_K_M, Q4_K_L, Q4_K_M) carried no imatrix at all — an
exclusion policy measured on a different model family that should never have applied
here. (Q8_0 also carries none, but that is correct: it is imatrix-free by rule.)
Every _K/IQ tier is being rebuilt on
AtomicChat/calib-corpora
builds/qwen3.8-27b at the full 9,703 chunks, ~76× the previous basis. The Q6_K
above is the recalibrated file; the remaining tiers are re-uploading and this note will
be removed when the ladder is complete.
Read these four deltas as a wash, not a win or a loss. The two negatives (IFEval −2.00 pp = 2 questions of 100; HumanEval −0.60 pp = 1 of 164) and the one positive (MultiPL-E +0.60 pp) are all inside single-draw sampled noise, and LiveCodeBench is unchanged to four decimals. The case for recalibrating is not a score gain — it is that seven tiers were shipping uncalibrated, and that an omitted imatrix is the asymmetric error: at worst it costs ~1 pp where it is neutral, while omitting one where it matters is catastrophic.
| Benchmark | Omnimerge-v6 (3.8) | Omnimerge-v4 (3.6) | Δ |
|---|---|---|---|
| GPQA-Diamond (198q) | 0.768 | 0.788 | -2.00 pp |
| HumanEval (thinking) | 0.982 | 0.982 | ±0.00 |
| IFEval (100) | 0.960 | 0.950 | +1.00 pp |
| LiveCodeBench (v6, 77q) | 0.883 | 0.818 | +6.50 pp |
This table is kept whole and unchanged on the previous-imatrix Q6_K column, because it is a controlled v6-vs-v4 pair. Swapping some cells to the recalibrated quant while others stayed would make the row set mixed-basis, which is worse than slightly stale; the calibration change is measured separately in table 1.
Only LiveCodeBench is outside the measurement band (+6.5 pp = 5 more solved of 77). The ±1 pp cells are not differences — a same-basis MultiPL-E-100 repeat on this rig moved 1.0 pp from batch scheduling alone. GPQA's −2 pp is 4 questions of 198: suggestive, not established. The mean is carried almost entirely by LiveCodeBench.
For the calibration question, GPQA is reported only from a greedy paired run, and is deliberately absent from table 1. Its sampled cell there was a single noisy draw reading −5.05 pp for the AtomicChat quant; a paired greedy rerun of that identical pair read +2.02 pp — the opposite sign, so the sampled number would have misrepresented the change. (GPQA remains in table 2, where both columns sit on the same sampled basis and the comparison is v6-vs-v4 rather than imatrix-vs-imatrix.) The greedy pair is the trustworthy measurement: both arms ran the same morning, concurrently, on the same host and binary, scored per-question on the same 198 items.
| GPQA-Diamond (198q), greedy | previous imatrix | AtomicChat imatrix |
|---|---|---|
| flexible-extract | 0.7778 (154/198) | 0.7980 (158/198) |
markdown
ac_correct ac_wrong delta(ac-prev) = +2.02 ppprev_correct 143 11 McNemar exact p = 0.5572prev_wrong 15 29 95% CI on delta = [-3.02, +7.06] ppdiscordant = 26/198 = 13.1%
p = 0.5572: the two quants are statistically indistinguishable on GPQA. This is not a
win. Its value is the interval — it excludes −5 pp, so the degradation the sampled
cell suggested is ruled out at 95%. Degenerate outputs (empty or [invalid]) were 5 for
the previous imatrix vs 2 for AtomicChat, and response lengths are comparable (p50 1142 vs
1187 chars), so nothing is hiding under the equal means.
tool-eval-bench hardmode (88 scenarios, 176 pts)v6 is first of ten in this cohort, and the safety result is the one worth reading.
| model | quant | Total Points (mean of 5 seeds) | 95% CI | safety-critical failures (5 seeds) |
|---|---|---|---|---|
| Qwen3.8-27B-Omnimerge-v6 | Q4_K_M | 156.4 ±3.5 | [152.0, 160.8] | 3 |
| Qwen3.8-27B (base) | UD-Q4_K_M | 150.8 ±2.5 | [147.7, 153.9] | 9 |
| Ornith-1.5-35B | IQ4_XS | 146.2 ±2.6 | [143.0, 149.4] |
v6 scores +10.2 pts over v4 and +5.6 over its own Qwen3.8 base, with non-overlapping CIs against v4.
The safety number is not a rounding difference. Every other model in the cohort —
including both bases — fails TC-60 (Cross-Turn Sleeper Injection) on 5 of 5 seeds.
v6 never fails it. Its 3 failures across five runs are 2× TC-58 (Fake System Message in
File) and 1× TC-33 (Hallucination Resistance); it has no TC-31 or TC-34 failure at all,
where v4 fails both on every seed. Safety & Boundaries 23.2/26 (89.2%) vs v4's
18.4/26 (70.8%).
Category profile (mean over 5 seeds): perfect on Tool Selection, Parameter Precision,
Restraint & Refusal, Localization, Instruction Following, Toolset Scale and Creative
Composition. Weakest at Autonomous Planning 3.4/6 (56.7%) — below v4's 4.0/6 — and
Context & State 14.8/20 (74.0%). Hard Mode 33.0/38 (86.8%) leads the cohort, but only narrowly — the Qwen3.8 base is 32.8/38 (86.3%). The wide gap is Safety & Boundaries: 89.2% here vs 78.5% for the next-best model.
tool-eval-bench v2.6.0 (the pip/uv-installed package, verified via
tool_eval_bench.__file__, not a git checkout). An earlier note in the runner claimed
cf54b4b (v2.6.0-45); that is wrong and has been corrected — no cell ever ran it.
All 50 cells in this cohort ran the same v2.6.0, so the cohort is internally consistent.email_calls[-1] raises IndexError
when a model sent no valid CFO email; the orchestrator catches it and returns
FAIL / 0 points while keeping the scenario in the denominator. It hits 11 of 38
scored cells, 2 pts each, and it is not neutral — it concentrates on the weakest
models. Later harness commits credit that behaviour instead, so the fixed scorer
would raise affected scores unevenly. v6 has no affected cell; v4 has none
either, so the v6-vs-v4 gap above is not exposed to it.llama.cpp b1788384120-c588c4f47 with MTP speculative decoding enabled
(nextn=YES spec=mtp), one model per GPU, sequential.GGUF (llama.cpp / ollama / text-generation-webui) — 19 tiers with the MTP head retained, plus the F16 vision projector:
ManniX-ITA/Qwen3.8-27B-Omnimerge-v6-MTP-GGUF
ollama — mannix/omnimerge-v6 (:latest = Q4_K_M). Tags carry the same serving identity as the official qwen3.8:27b (RENDERER qwen3.8, PARSER qwen3.5, vendor sampling defaults), so they need ollama >= 0.32.12.
bash
ollama run mannix/omnimerge-v6
bash
python omnimergekit.py \--base Qwen/Qwen3.8-27B \--task-base Qwen/Qwen3.6-27B \--source rico03/Qwen3.6-27B-Claude-Opus-Reasoning-Distilled \--source ValiantLabs/Qwen3.6-27B-Esper3.1 \--source <kai-os LoRA applied to Qwen3.6-27B> \--weights 0.40,0.35,0.25 \--method omnimerge_v2 \--density 0.53 --darex-q 0.75 --seed 42 \--skip-patterns 'mlp.gate_proj,mlp.up_proj,mlp.down_proj,mtp.' \--output Qwen3.8-27B-Omnimerge-v6
The kai-os source is a LoRA; it is applied to Qwen3.6-27B first (scripts/apply_lora_to_safetensors.py) and the resulting anchor is merged as a full model. Tokenizer, chat template and preprocessor configs are copied from the 3.8 base after the merge.
Merge engine: mann1x/omnimergekit.
Qwen team for the Qwen3.8 base and the vision/MTP components; rico03, ValiantLabs and kai-os for the source fine-tunes.
| Mean (4) | 0.928 | 0.923 | -0.50 pp |
| MultiPL-E (100) | 0.887 | 0.877 | +1.00 pp |
| Mean (5) | 0.896 | 0.883 | +1.30 pp |
| 10 |
| Qwen3.6-27B-Omnimerge-v4 | Q4_K_M | 146.2 ±2.7 | [142.9, 149.5] | 16 |
| Qwen3.6-27B | Q4_K_M | 144.0 ±3.4 | [139.8, 148.2] | 14 |
| Qwen3.6-35B-A3B | IQ4_XS | 141.6 ±2.4 | [138.6, 144.6] | 15 |