MergeAILab/Merge-27B-MTP-Reasoning-v1

Merge-27B-MTP-Reasoning-v1

Merge AI Lab · dense reference model

A LoRA supervised fine-tune of Unsloth's Qwen3.6-27B-MTP, trained locally on a single AMD Radeon 8060S (ROCm) at Merge. Multi-token-prediction (MTP) heads and the vision tower are preserved from the base; only the language tower is adapted.

This is one of two models we are publishing from the same internal programme. It is the dense reference — the quality bar that our Mixture-of-Experts sibling, Merge-35B-A3B-Reasoning-v5-b1, was built to chase. Both were developed against the same capability targets and scored on the same internal benchmark.

We are releasing these because they are honest, reproducible results from a small lab running on one consumer GPU, and because the recipe — what moved the score and what did not — is more useful shared than kept.

What it was built for

Both Merge models were developed and evaluated against one capability set, drawn from the work our internal agents actually do:

The end purpose is internal: both models drive MCP servers and tool-using agents across our copy, design and implementation pipeline. They are meant to be the local model behind an agent that plans a page, writes the copy in pt-PT, and implements the component — not a general-purpose chat assistant.

The SFT corpus itself is reasoning-trace distillation (below). The capability list above is what the model was selected and measured against, not a description of the training mix.

Out of scope: anything safety-critical. This fine-tune adds no safety training of its own — the base model's alignment is all that is present.

Results

Scored on our private internal benchmark (described below). Both rows are the same quantization, the same harness, the same grader.

+5 points over the base, and rank 1 of +50 local models on our rolling table — ahead of its own MoE sibling and every other local build we have tested.

A second independent run scored 93 — a 1-point run-to-run spread.

About the benchmark

We are not publishing the benchmark, but here is what it is and is not.

It is an 8-task suite built from real briefs in our own production workflow — not synthetic puzzles and not sampled from any public set. Each task is graded against a hidden rubric with withheld answer keys, and the tasks cover exactly the capability list above: design decomposition, planning a content structure from a thin brief, writing and adapting frontend components, pt-PT localization quality, surgical CSS repair from visual intent, client-side React/Next interaction, and a typed Next.js server route backed by PostgreSQL.

Scoring:

• —Technical /100 — seven tasks, raw scores summed and rescaled to 100. This is the ranking metric, and full credit is reachable without any tool use.
• —Vision bonus /12 — one image-grounded task, scored separately so text-only models are not penalized.
• —Tool-grounding bonus /19 — quality of tool use, awarded only against verifiable session-log evidence. Also separate from the ranking score.

Deliberately, the technical score judges output quality, not protocol compliance: a model is not punished on its main score for clumsy tool orchestration, because a bad plan executed cleanly is still a bad plan.

Honest limits, stated plainly:

• —n = 2 runs, 8 tasks, single-operator rubric grading. Treat 94 as indicative, not precise. The gap to the next model is within plausible run-to-run noise.
• —Scores are not comparable to public leaderboards. Different tasks, different rubric, different grader. A 94 here means "94 on our bar", nothing more.
• —Models were scored at their own preferred sampling presets, not one shared preset — this model at temp 0.1, the 35B MoE at temp 0.6. That reflects each model at its best, and is not a controlled A/B.
• —Retrieval answer keys were denylisted at run time and verified absent from the outputs.

Training

Packing is disabled deliberately so reasoning traces are never split across samples.

Training data

1,078 reasoning conversations combining two public TeichAI datasets — Claude Opus 4.5 (250 examples) and Opus 4.6 (887 examples) — deduplicated and filtered to an 8k-token budget, split 90/10 train/eval. Plain messages turns; no tools, and no private, client or company data.

Provenance disclosure: these are distilled traces generated by Claude Opus models. Anyone redistributing or building on this model should check both the TeichAI dataset licenses and Anthropic's terms covering the use of model outputs.

Training dynamics

Smooth convergence, no instability; final grad-norm 0.38.

Files

Q4_K_M is the only quantization we ship, because it is the only one we benchmarked. If you want another, convert from the BF16 GGUF — we would rather publish one measured quant than five unmeasured ones.

Embedded GGUF metadata: general.name = Merged Merge 27b Mtp Reasoning, general.basename = merged-merge, general.finetune = mtp-reasoning.

Sampling settings

These are the settings the 94/100 was produced at. This model is stable at low temperature.

The low temperature is deliberate: this model is used for precise technical and structured output, and it does not degrade into repetition there.

If you prefer more varied prose, the Qwen3 reasoning-mode defaults (temp 0.6, top_p 0.95, top_k 20) also work well, but the reported score was not measured there.

Note: the GGUF embeds temp 1.0, top_p 0.95, top_k 20. Set the values above explicitly — some runtimes pick up the embedded defaults.

Usage

llama.cpp:

llama-server \
  -m Merge-27B-MTP-Reasoning-v1-mtp-Q4_K_M.gguf \
  --mmproj mmproj-F32.gguf \
  --jinja --reasoning-format auto --reasoning-budget 6000 \
  --temp 0.1 --top-k 40 --top-p 0.9 --min-p 0.05 --repeat-penalty 1.05 \
  -c 124096 -ngl 999 --flash-attn on

The model is trained to emit long reasoning traces; a low --reasoning-budget will truncate them mid-thought.

Limitations

• —Trained at 1024-token max length, well below the 124k context it is served at. Long-context behaviour is inherited from the base model, not tuned.
• —English and European Portuguese in practice; other languages are base-model behaviour.
• —Distilled reasoning style can be verbose, and can produce confident-sounding traces that reach a wrong conclusion.
• —Slower per token than the MoE sibling. If throughput matters more than the last two points of quality, use the 35B-A3B.
• —No added safety tuning.

Licensing

Released under Apache 2.0, matching the Qwen3.6-27B base as published by Unsloth (unsloth/Qwen3.6-27B-MTP-GGUF, unsloth/Qwen3.6-27B — both apache-2.0, derived from Qwen/Qwen3.6-27B).

Two further conditions apply to anyone building on this model, and they are not ours to grant: the TeichAI dataset licenses, and Anthropic's terms on the use of Claude outputs — the training traces are Claude-distilled. Check both before redistributing.

Citation

@misc{merge27b_mtp_reasoning_v1,
  title  = {Merge-27B-MTP-Reasoning-v1},
  author = {Mergeinto.digital},
  year   = {2026},
  note   = {LoRA SFT of Unsloth Qwen3.6-27B-MTP on distilled reasoning traces}
}