Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF — Model publication¶

Development line: project:qwen3-5-27b-claude-4-6-opus-reasoning-distilled-gguf · thread model-publication
Last event: 2026-03-24 · 1 dated since 2026-03-24 · Researched: 2026-09-06 · confidence: medium

What it is¶

Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF is a user-published GGUF build of Qwen3.5-27B for local analysis, code, and reasoning tasks. The verified files page lists Q3_K_M at 13.3 GB, Q4_K_M at 16.5 GB, and Q8_0 at 28.6 GB. The model card claims SFT+LoRA tuning on Claude Opus 4.6 reasoning trajectories; Anthropic's verified announcement contains no such origin. The 16.5 GB figure is the Q4_K_M file size, not a verified runtime memory measurement. This is a third-party artifact, not a local version of Claude Opus 4.6.

Development line¶

• 2026-03-24 — Hugging Face repository appeared for Qwen3.5-27B Claude 4.6 Opus Reasoning Distilled GGUF. On 2026-03-24, a public repository named Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF appeared on Hugging Face. The dated link shows an artifact was available under that name. Public details do not confirm its contents, origin, or release details.

What changed¶

2026-02-24 — Qwen released the official Qwen3.5-27B base model listed on the GGUF card. 2026-03-24 — Community GGUF files became available; the public card does not confirm this date as a release or weight update.

How to use this¶

1. Check the target GGUF: the verified files page lists Q3_K_M at 13.3 GB, Q4_K_M at 16.5 GB, and Q8_0 at 28.6 GB; pick one file matching available disk and memory. — https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF/tree/main
2. Run hf download <repo> <exact-file> --dry-run before fetching files; then download the file with the full commit SHA pinned in --revision. — https://huggingface.co/docs/huggingface_hub/en/guides/download
3. Start the chosen quantization with llama serve -hf Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M; the card also lists llama cli for one-off runs. — https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF
4. Check GET /health before connecting an agent; HTTP 200 with status ok means the model is loaded. Then check the chat template through /props on the pinned revision. — https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md

Best practices¶

• Pin the revision with the full commit SHA: Hugging Face requires the full hash to download a specific version; run a dry run first to avoid pulling the whole repository by mistake. — https://huggingface.co/docs/huggingface_hub/en/guides/download
• Specify the quantization explicitly (:Q4_K_M) or pass the exact file with -hff: Q4_K_M is the default for -hf, and without it the runtime picks the first file in the repository. — https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md
• Leave Jinja and metadata chat templates enabled on initial launch; override the template only after checking actual model properties through /props. — https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md
• Treat claims about quality, autonomy, VRAM, and training data as author claims: the card calls the build a preview while warning about hallucinations and incompatibilities. — https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF/raw/main/README.md

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The public model card does not show 2026-03-24 as the creation, release, or update date; it only contains a screenshot with that date.
• The card mentions "Update (April 5)" without a year or commit timestamp, so we do not list it as a dated event.
• Independent primary sources have not confirmed the origin of the claimed Claude Opus 4.6 reasoning data, its licensing, or the stated benchmark, VRAM, and speed figures.
• The card calls the final model text-only, but the repository has a vision/multimodal tag and includes mmproj-F32.gguf; test image support locally on the pinned revision with a smoke run.
• Tests for compatibility with specific agents, tool calling, and output quality across quantizations have not been run.

Sources¶

Agent brief¶

• Subject: project:qwen3-5-27b-claude-4-6-opus-reasoning-distilled-gguf, thread model-publication, 1 dated events 2026-03-24 → 2026-03-24.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.