---

license: cc-by-nc-sa-4.0

base_model: devon7y/WikiQwen-35B-A3B

tags:

• gguf
• llama.cpp
• qwen3.6
• moe
• how-to
• instructional
• illustrated
• step-by-step

---

WikiQwen-35B-A3B-GGUF

A Q4_K_M GGUF of devon7y/WikiQwen-35B-A3B, built for llama.cpp-compatible runtimes.

Why this model is interesting

WikiQwen is not a conventional general-chat fine-tune. Its trained behavior is:

Ask almost anything → get a structured how-to article.

Typical generations include:

• a # How to ... title
• numbered methods / parts / steps
• an [IMAGE: caption] line before each step
• tips, warnings, and things-you'll-need sections when useful

Those image captions are deliberately written so they can be passed to the companion WikiQwen-Illustrator models to create step illustrations.

The upstream model is based on Qwen3.6-35B-A3B, a mixture-of-experts model with roughly 35B total parameters and ~3B active per token.

This GGUF

• Quantization: Q4_K_M
• File: WikiQwen-35B-A3B-Q4_K_M.gguf
• Size: 19.71 GiB
• SHA256: eed0b1fe4daeefd36e752a90e121cfeb3942ea566bdad31d795063de2e7d8ac1
• Source revision: 9fd1a0a89ad213c05716383168844fc2d3e20749
• Conversion: source safetensors → BF16 GGUF → Q4_K_M
• MTP handling: converted with --no-mtp because the source config declares one MTP/NextN layer but the published source tensors contain no corresponding MTP weights

Run with llama.cpp

llama-cli -m WikiQwen-35B-A3B-Q4_K_M.gguf -ngl 999 -c 8192 --reasoning off

OpenAI-compatible server:

llama-server -m WikiQwen-35B-A3B-Q4_K_M.gguf -ngl 999 -c 8192 --reasoning off

Recommended prompting

The upstream author recommends no system prompt. The article-writing behavior is trained into the model.

Example user prompt:

How do I keep basil alive on a windowsill?

The source model was trained with thinking disabled. Keep reasoning off for behavior closest to upstream.

The upstream demo sampling recipe is approximately:

• temperature: 0.7
• top_p: 0.9
• repetition penalty: 1.05
• up to ~1800 new tokens for a full article

Long generations can occasionally repeat a step; upstream notes that its demo stops or trims repeated / unfinished steps.

Upstream training highlights

According to the source model card:

• about 118k how-to articles were drawn from the Kiwix March 2023 archive
• 107,054 training conversations and 300 validation conversations were used
• every training article was converted into a user-message → structured-article conversation
• each step includes an [IMAGE: ...] caption
• roughly 49% of captions were based on the step's real Creative Commons illustration
• the LoRA used rank 32, alpha 64, dropout 0.05
• the adapter was merged into the published model
• training used assistant-only loss, one epoch, sequence length up to 4096

The fine-tune mainly teaches the article format and response habit, not factual verification. The upstream author explicitly warns that health, safety, legal, financial, and other consequential advice should be independently checked.

Validation of this quant

This exact Q4_K_M file passed a real llama-server smoke gate:

• model load: PASS
• exact-output instruction: PASS
• factual short answer: PASS
• arithmetic: PASS
• system/user separation: PASS
• multi-turn context: PASS
• normal stop behavior: PASS
• no chat-control-token leakage observed

The first arithmetic smoke run used an overly small 64-token cap and truncated before the final answer. Re-running with 128 tokens completed normally and returned 323 for 17×19.

License and attribution

This derivative follows the upstream CC BY-NC-SA 4.0 license. The upstream model was trained on text and illustrations by wikiHow contributors from a March 2023 Kiwix archive under CC BY-NC-SA 3.0.

All model and dataset credit belongs to the original authors and contributors. This GGUF repository only provides a quantized runtime format.