Qwen3.8-27B-Cold-Fusion-Remix-oQ4e

sluttybutfast/Qwen3.8-27B-Cold-Fusion-Remix-oQ4e

Standing on the shoulder of giants:

• Language model: from DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1 (finetuned, great reasoning reductions)
• Quantization: using oQ (oMLX v0.6.2) mixed-precision
• MTP head dropped in favor of size (acceptance rate not worth it on recommended temp)
• Chat template: based on froggeric/Qwen-Fixed-Chat-Templates (v21.3)
• Custom template additions to further improve behaviour and tool handling
• For superior caching, run with oMLX
• Important!: use recommended settings (DO NOT USE WITH LOW TEMP!!, force sampling in oMLX if necessary)

Recommended settings (forced in oMLX):

• temp 0.8 (0.4-1.0 should be safe, i don't recommend greedy temp)
• top_p=0.95
• min_p=0.02 (chop some noise, 0.0 is also fine)
• top_k=20
• repetition_penalty=1.05 (harmless antiloop, use 1.0 if you prefer)
• presence_penalty=0.0
• thinking_budget=none (better to limit output tokens and use reasoning_effort)
• enable_thinking=true (default)
• preserve_thinking=true (default)
• reasoning_effort=medium (default, uses custom instruction)
• output_token_limit=12000 (might need to increase when using high reasoning)
• KV quant: TQ4

The model uses a tuned froggeric chat_template with default reasoning_preserve=true and custom system prompt addons to lower thinking loop probability and provide some general reasoning guidance. If you prefer the original chat_template replace chat_template.json with chat_templat.json.org or any you like. (minimum should be the unsloth one, as the Qwen default template sucks.)

If you think you need MTP, use f.e. pyros-vault/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-oQ4e-mtp and replace chat_template.jinja with the one in this model.