Qwen3.6-34B-80L-Fable-5-Heretic¶

Development line: project:qwen3-6-34b-80l-fable-5-heretic · thread qwen3-6-34b-80l-fable-5-heretic
Last event: 2026-06-29 · 1 dated since 2026-06-29 · Researched: 2026-09-05 · confidence: medium

What it is¶

Qwen3.6-34B-80L-Fable-5-Heretic is a BF16 derivative of Qwen3.6-27B for self-hosting with Transformers or vLLM.

• 80 layers, ~34B parameters, and context claimed up to 256K.
• QLoRA training on 4 665 Fable-5 agentic CoT traces.
• MTP weights in the repository for speculative decoding.

Development line¶

• 2026-06-29 — Qwen3.6-34B-80L-Fable-5-Heretic model page recorded on Hugging Face. A Qwen3.6-27B derivative expanded from 64 to 80 layers, claiming ~34B parameters, BF16 weights, 4 665 QLoRA traces, and 15 MTP weights. We found no independent dated confirmation of publication on 2026-06-29.

What changed¶

2026-06-29 — We recorded the link to hiebo/Qwen3.6-34B-80L-Fable-5-Heretic. The model card details this step: the Qwen3.6-27B derivative expands from 64 to 80 layers, claiming ~34B parameters, BF16 weights, 4 665 QLoRA traces, and 15 MTP weights. We found no independent dated confirmation of publication on 2026-06-29.

After 2026-06-29 — We found no verified dated changes to the model. Mirrors on Hugging Face redistribute the same artifacts rather than a new version.

How to use this¶

From 2026-06-29, locate the Qwen3.6-34B-80L-Fable-5-Heretic entry on Hugging Face and verify the model card, files, licensing, provenance, and compatibility before adopting it.

1. Load the exact ID hiebo/Qwen3.6-34B-80L-Fable-5-Heretic through Transformers with device_map="auto" and test a short prompt before loading the full context. — https://huggingface.co/hiebo/Qwen3.6-34B-80L-Fable-5-Heretic
2. Serve with the vLLM command from the model card using BF16 and an explicit --max-model-len. Start with a lower limit until you measure KV-cache and GPU memory. — https://huggingface.co/hiebo/Qwen3.6-34B-80L-Fable-5-Heretic

Best practices¶

• Do not carry over base model settings automatically: the custom derivative changes the layer count and adds MTP weights, so verify loader, tokenizer, and generation pipeline compatibility with a local smoke test first. — https://huggingface.co/hiebo/Qwen3.6-34B-80L-Fable-5-Heretic
• Treat the claimed 78% acceptance rate and 2× throughput as a model card hypothesis; measure acceptance, latency, and quality on your own task suite. — https://huggingface.co/hiebo/Qwen3.6-34B-80L-Fable-5-Heretic
• The base Qwen model recommends a large generation limit for complex tasks; treat that as an experimental setting, not proof of derivative quality. — https://huggingface.co/Qwen/Qwen3.6-27B

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The model card is an author claim from a personal repository; we found no independent evaluations, reproducible training runs, or verified speedups.
• Primary sources omit exact publication and update dates, so we cannot confirm the release occurred on 2026-06-29.
• The schema lacks event_findings and new_events fields; we recorded relevant findings in what_changed and unknowns.

Sources¶

Agent brief¶

• Subject: project:qwen3-6-34b-80l-fable-5-heretic, thread qwen3-6-34b-80l-fable-5-heretic, 1 dated events 2026-06-29 → 2026-06-29.
• Practical note: From 2026-06-29, practitioners can locate the Qwen3.6-34B-80L-Fable-5-Heretic Hugging Face entry and should verify the model card, files, licensing, provenance, and compatibility before adopting it.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.