Qwen3.8-27B Cold Fusion Demo

Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M
Apple MLX oQ8e · 1,048,576-token YaRN · Metal-only · ~30GB

⚠️ The MLX quant cannot run on Spaces (Linux, no Metal). Live chat below proxies to the identical base weights (DavidAU/…, same merge) via Inference Providers (featherless-ai). For full-fidelity local runs, see the Deploy on Mac tab.

Billed to the token owner. Get a token (fine-grained → Inference Providers). Saved only in this browser's localStorage.

Enable thinking (<think>)

Temp Max tokens

Model

Build exact Qwen3.8 prompts (<think> CoT vs suppressed) for local MLX use.

System

User

Enable thinking

Formatted prompt

Tokenizer snippet

9-for-9 vs Claude Opus 4.6 Max

Suite	Focus	Qwen3.8-27B TURBO	Opus 4.6 Max	Margin
SWE-bench Pro	Agentic SWE	61.7%	53.4%	+8.3%
LiveCodeBench v6	Live problem solving	90.3%	88.8%	+1.5%
QwenSWEBench	Repo debugging	79.0%	63.8%	+15.2%
OSWorld-Verified	OS control	84.3%	72.7%	+11.6%
AndroidWorld	Mobile autonomy	81.9%	62.0%	+19.9%
IFBench	Constraint following	79.5%	62.5%	+17.0%
CoWorkBench	Multi-file workflows	70.7%	68.2%	+2.5%
ARC-C	Scientific abstraction	735	~710–720	frontier tier
ARC-E	Commonsense	882	~870	exceeds frontier

Architecture: 75% Gated Delta (linear) attention + 25% GQA, GAIN Cold-Fusion merge, Project Heretic abliteration, MTP dual-stream drafting (1.7–2.2× on Apple Silicon), importance-matrix calibrated oQ8e, baked-in 1M YaRN RoPE.

Run the real MLX model on your Mac

Option 1 — Anvil (recommended)

curl -fsSL https://anvil-llm.github.io/anvil/install.sh | sh
anvil run hf:Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M --ctx 1048576 --type-k turbo4 --type-v turbo3
anvil serve hf:Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M --port 8080 --ctx 1048576

Option 2 — MLX-LM

pip install --upgrade mlx-lm
python -m mlx_lm.generate --model Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M --prompt "<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\n" --max-tokens 1024 --temp 0.6
python -m mlx_lm.server --model Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e-1M --port 8080