Applications

Claude Code ollama launch claude --model itsailor/qwen3.6-fable

OpenCode ollama launch opencode --model itsailor/qwen3.6-fable

Hermes Agent ollama launch hermes --model itsailor/qwen3.6-fable

OpenClaw ollama launch openclaw --model itsailor/qwen3.6-fable

qwen3.6-fable:27b-mlx

NVFP4 quantization of DavidAU’s Qwen3.6-27B Fable-Fusion-711 (Uncensored), built for Ollama’s Apple‑Silicon MLX backend.

A 27B dense, uncensored creative‑writing / roleplay model that runs natively in Ollama on Apple Silicon via the NVFP4 (MLX) path. On a base M4 / 32GB it decodes about 2× faster than the GGUF Q4_K_M build while preserving the model’s behavior.

⚠️ Not for all audiences. This is an uncensored / abliterated model that will follow instructions without moralizing, including adult and NSFW fiction. Use responsibly and in accordance with local law. You are responsible for what you generate.

Quick start

ollama run itsailor/qwen3.6-fable:27b-mlx

This is a thinking model. Ollama puts the chain‑of‑thought in a separate thinking field, so if you only read response with a small token budget it can look empty. For direct answers, disable thinking:

curl http://localhost:11434/api/chat -d '{
  "model": "itsailor/qwen3.6-fable:27b-mlx",
  "messages": [{"role": "user", "content": "寫一句魔王的台詞"}],
  "think": false,
  "options": { "temperature": 0.8, "top_p": 0.92, "repeat_penalty": 1.2, "repeat_last_n": 256 }
}'

Notes - Ships as a bare base (no baked‑in system prompt). Apply your own persona via a derived Modelfile, the system field per request, or /set system. - To reduce repetition on onomatopoeia‑heavy passages, keep repeat_penalty ≈ 1.2 + repeat_last_n 256. Do not set repeat_penalty 1.0 (disables it). - Outputs Traditional/Simplified Chinese and English; instruct it explicitly if you need a specific variant.

Hardware

Apple Silicon, ~22 GB unified memory in use when loaded (100% GPU).
Runs on a base M4 / 32 GB. Very long contexts (≈50k+ tokens) can exhaust 32 GB — use the GGUF build if you need huge contexts on limited RAM.

Benchmark & Performance (Apple M4 base / 32GB, measured) Evaluation

Evaluated locally on Apple Silicon (macOS) using the local-coding-agent-evals benchmark suite. Tests measure prompt evaluation (prefill) throughput, token generation (decode) speed, process memory footprint across varying context sizes, and hard tool-reasoning / coding capabilities.

Speed & Memory Benchmark

Context Size (Approx.)	Prompt Tokens	Generated Tokens	Prefill Speed (`tok/s`)	Decode Speed (`tok/s`)	Peak Process RSS	Total Wall Time
1k words	1,395	1,012	43.85	9.99	5.38 GB	2.5 min
5k words	6,731	1,103	37.63	9.55	1.23 GB	4.9 min
10k words	13,403	888	35.04	8.68	1.22 GB	8.1 min
50k words	66,767	1,272	19.60	3.80	1.22 GB	62.3 min

Hard Tool-Reasoning & Code Capabilities

Evaluated across 5 challenging reasoning tasks encompassing code debugging, security audit, surgical file editing, and tool-call selection:

Task ID	Category	Result	Evaluation Summary
`debug_empty_tokenizer_regression`	Debugging	PASS	Accurately diagnosed empty-string split edge case and provided Pythonic early-return fix.
`review_shell_command_injection`	Code Review / Security	PASS	Identified `shell=True` command injection risk; proposed argument lists and sandboxing.
`choose_minimal_edit_for_cross_platform_path`	Surgical Editing	PASS	Correctly invoked `edit_file` tool with cross-platform `Path(__file__).parent` resolution.
`triage_import_error_after_refactor`	Triage / Tool Selection	Valid Triage	Reasonably requested config and import inspections via clarification fallback.
`debug_mutable_default_cache_leak`	Debugging	PASS	Correctly identified mutable default parameter state leakage and provided `cache=None` fix.

How it was made

Quantized from the original BF16 weights (not from GGUF) with Ollama’s experimental safetensors → NVFP4 importer:

ollama create qwen3.6-fable:27b-mlx --experimental -q nvfp4 -f Modelfile
# Modelfile: a single line ->  FROM <path to BF16 safetensors dir>

No fine‑tuning or weight changes were made — this is a quantization only. Thinking and tool‑calling are inherited from the source and work.

⚠️ Vision does NOT work in this build. Ollama auto‑detects a vision capability badge from the model metadata, but image input was tested and confirmed non‑functional - the NVFP4 (MLX) conversion path does not wire up the vision tower, so images are silently ignored and the model hallucinates a description. Treat this as a text‑only model. If you need working vision, use a different build.

Attribution, lineage & license

This model is a derivative work (quantization only). All credit for the weights and behavior belongs to the upstream authors.

Lineage - Base architecture: Qwen3.6‑27B — © Alibaba Cloud / the Qwen team. - Fine‑tune / merge: DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP — © DavidAU. All the model’s capabilities and character come from this work. - This build: NVFP4 quantization of the above for the Ollama MLX backend. The only modification is quantization/packaging.

License: Apache‑2.0, inherited from the upstream model (as declared on DavidAU’s model card). Your use of this model is governed by the Apache‑2.0 license and by the Qwen license terms for the underlying base model. This redistribution preserves that license; see the Apache License 2.0.

Not affiliated with Alibaba/Qwen, DavidAU, or Ollama. This is an independent community quantization. No warranty; provided “as is.”

If you use or build on this, please also credit DavidAU and the Qwen team - they did the hard part.