Qwopus3.6-27B-Coder-MTP¶

Development line: project:qwopus3-6-27b-coder-mtp · thread qwopus3-6-27b-coder-mtp-development
Last event: 2026-06-18 · 1 dated since 2026-06-18 · Researched: 2026-09-05 · confidence: medium

What it is¶

Qwopus3.6-27B-Coder-MTP is a Qwopus3.6-27B-v2 coding fine-tune with MTP draft heads for local repository work.

• Code generation for repositories.
• Debugging and patching code.
• Structured tool calls.

Available GGUF quants range from 13.5 GB Q3_K_M to 29 GB Q8_0. The community-reported SWE-bench score is 335/500 in a Q5_K_M thinking-off run, with other suites pending. We use it as an experimental local coding model, not as an independently validated production authority.

Development line¶

• 2026-06-18 — Qwopus3.6-27B-Coder-MTP GGUF artifact linked. We linked Qwopus3.6-27B-Coder-MTP-GGUF, its Qwopus3.6-27B-v2 base, and the Qwen MTP-to-GGUF conversion workflow as one local speculative-decoding path.

What changed¶

2026-06-18: Qwopus3.6-27B-Coder-MTP-GGUF, its Qwopus3.6-27B-v2 base, and the Qwen MTP-to-GGUF conversion workflow were linked as one local speculative-decoding path.

How to use this¶

From 2026-06-18, treat the linked GGUF artifact as a candidate Coder-MTP deployment format. Verify provenance, runtime compatibility, licence, and performance against the linked v2 model before use.

1. Choose a GGUF quantization to fit storage and memory. The repository lists Q3_K_M through Q8_0 builds. — https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF/tree/main
2. Run the GGUF in a llama.cpp-compatible local application. The model card names llama.cpp, Ollama and LM Studio as compatible targets. — https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF
3. Keep tool definitions and the system prompt aligned with the training format for agents. Handle text inside tags in the client. — https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF/blob/main/README.md

Best practices¶

• Treat the reported 67.0% SWE-bench score as a single setup, not a cross-harness ranking. The card marks the release experimental and leaves other suites pending. — https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF/blob/main/README.md
• Run the preflight before downloading weights when creating or repackaging an MTP GGUF. A source/target configuration mismatch or insufficient disk blocks the build. — https://raw.githubusercontent.com/R6410418/Jackrong-llm-finetuning-guide/main/qwen-mtp-gguf/README.md
• Smoke-test the GGUF before upload or deployment. Use a matching MTP-head source model rather than assuming heads transfer across variants. — https://raw.githubusercontent.com/R6410418/Jackrong-llm-finetuning-guide/main/qwen-mtp-gguf/README.md

Superseded by this¶

• 2026-07-09: a super-squashed GGUF cleanup commit replaced the prior granular repository history. It shows no new model capability.

Still unknown¶

• No dated primary source proves the benchmark, throughput, or MTP speed claims held on 2026-06-18. We do not attach those claims to that event.
• Hugging Face metadata labels the GGUF as image-text-to-text, Transformers and a 0.5B CLIP architecture, but the README describes a 27B coder GGUF. This conflict makes the Transformers, vLLM and image examples unverified for deployment.
• No independent benchmark reproduction, compatibility matrix, or release changelog exists for the 2026-06-18 release point.

Sources¶

Agent brief¶

• Subject: project:qwopus3-6-27b-coder-mtp, thread qwopus3-6-27b-coder-mtp-development, 1 dated events 2026-06-18 → 2026-06-18.
• Practical note: From 2026-06-18, treat the linked GGUF artifact as a candidate Coder-MTP deployment format. Verify provenance, runtime compatibility, licence, and performance against the linked v2 model before use.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.