AIAny
AI Model2026
Icon for item

DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-MTP-GGUF

Post-trained Qwen3.8-27B variant using the COLD FUSION (GAIN+Unsloth) tuning to reduce internal reasoning-token use and improve instruction following while keeping base capabilities. Deliverables include 256k-context-compatible GGUF quants (regular and MTP, NEO IMATRIX), vision support via an mmproj, and three reasoning-effort modes (xhigh/medium/low).

Introduction

Most practical LLM improvements trade compute for longer, costlier internal “thinking” traces. This model applies the COLD FUSION (GAIN+Unsloth) post-training technique to reformat and shrink reasoning blocks so the same reasoning power is produced with far fewer thinking tokens — often 1/10 to 1/2 of typical Qwen reasoning length — while preserving or improving benchmarks.

Key capabilities
  • Reduced reasoning token footprint: median reductions reported between ~50% and 90% vs baseline Qwen3.8 reasoning blocks, enabling cheaper multi-turn and long-horizon use. This is the core, deliberate design goal of the tuning.
  • Quantized releases: multiple GGUF quants are provided (regular and MTP) built with NEO IMATRIX adjustments; some tensors kept at 16-bit for output fidelity. MTP quants target higher throughput via multi-token prediction.
  • Multi-mode reasoning control: supports xhigh (default), medium and low reasoning_effort via the model's chat-template; preserves thinking blocks by default but can be disabled.
  • Native multimodal/vision support: vision activated via a separate mmproj shipped with the release; model remains compatible with common inference stacks (HuggingFace Transformers, vLLM, TokenSpeed, etc.).
  • Large context: released artifacts expect long-context usage (256k native), with guidance for YaRN overrides to extend to 1M tokens in supported engines.
Who it’s for and tradeoffs

Great fit if you need a deployment-friendly Qwen3.8 27B variant optimized for lower inference token cost and multi-turn agentic workflows (coding, long-horizon agents, multimodal tasks) and if you plan to run quantized GGUFs in local/edge inference stacks. The model includes both MAX-quality and LOW-speed quants (IQ4_XS, Q6_K) to trade VRAM for throughput. Look elsewhere if your workload requires absolute upstream Qwen3.8-vanilla fidelity across every benchmark variant (the author rates this tune about a 1–2 level change vs their heavier Fable-Fusion work) or if your target environment lacks support for the provided GGUF/MTP tooling — testing both regular and MTP quants is advised because MTP acceptance rates affect real-world speed.

Where it fits

This release is a community-maintained, post-trained distribution of a Qwen3.8-27B derivative focused on inference efficiency and reasoning reformatting. Use it for lower-cost local inference, agentic stacks, and multimodal prototypes where smaller reasoning footprints and GGUF quant support matter most. For production critical deployments, validate reasoning-effort settings and quants on your workload (humidity of prompts, multi-turn behavior and MTP token acceptance all affect outcomes).

Information

  • Websitehuggingface.co
  • OrganizationsDavidAU, Nightmedia, TeichAI
  • AuthorsDavidAU
  • Published date2026/08/17

Categories

More Items

Hugging Face
AI Model2026

Runs text generation, summarization, and question answering directly on Windows devices as a 2.3B-parameter instruction-tuned model. It supports eight languages and a 32K context window, trading broad world knowledge and deep reasoning for local execution and privacy.

Hugging Face
AI Model2026

Handles text, image, and video understanding with sparse attention designed for contexts up to 1 million tokens. It has 770B total parameters with 49B activated per token, targeting long-document analysis, multimodal reasoning, and extended workflows.

Hugging Face
AI Model2026

Estimates metric depth for every pixel in a photo, video, or live feed, with sharp object boundaries and stable results across frames. Multiple model sizes are available; weights are restricted to non-commercial research and academic development.