Deepwen 3.6 is a fine-tuned derivative of Qwen/Qwen3.6-35B-A3B (Mixture-of-Experts, ~35B total / ~3B active), forged in DeepSeek traces — its reasoning DNA, effort control and "verify before you answer" discipline come from DeepSeek-V4-Flash-0731. On top of that foundation, it improves Design, Web Graphics and adds specialized skills for AAA GameDev 2D, 3D production workflows: procedural geometry, hard-surface shape language, and Blender asset pipelines.

What the model has

Advanced thinking (DeepSeek style) — the biggest source of performance gain, and the reasoning DNA comes from DeepSeek-V4-Flash-0731: its reasoning-effort system, its thinking discipline, its "verify before you answer" culture. The model reasons before it answers, and its thinking comes from two sources:
- Supervised reasoning training: the vast majority of training examples carry a full reasoning chain as part of the target — the model learns to think before it speaks, not just to parrot.
- Reasoning-effort control: a chat template ported from deepseek-ai/DeepSeek-V4-Flash-0731, with three effort levels — low (default), xhigh, and max ("Beyond maximum — exhaustive, relentless... do not stop reasoning until you have independently verified the solution from multiple angles").
Measurable improvements over the base model across procedural generation, hard-surface design, Blender workflows, lighting, web-graphics (Canvas/Three.js/WebGPU) and UI/design — while the original Qwen capabilities are fully preserved. Only improvements, no losses.
Tool calling and agentic behavior — reinforced through the reasoning training (planning, tool selection, structured output), retaining full native tool-calling support from the Qwen base.

Specialized 3D skills

Procedural 3D generation — explicit blockout gating before high-poly, conditional lightmap workflows, combinatorial validation, non-destructive pipelines.
Hard-surface shape language — stance/relational design, primary volume architecture, motif propagation, panel breakup.
Multi-skill asset workflows — Blender modifier-driven gear recipes, tooth profile generation, PBR game-prep, layered lighting legibility.
Web-graphics — WebGPU pipeline architecture, Three.js/WebGL workflows, Canvas 2D.
UI/design systems — layout, component architecture, visual hierarchy, accessibility.

Training overview

Fine-tuned with a curated, multi-skill supervised dataset. The vast majority of training examples carry full reasoning chains — the model learns to think before it answers, not just to parrot. Training focuses on AAA 3D asset production workflows plus web-graphics expertise (WebGPU, Three.js/WebGL, Canvas 2D) and UI/design systems.

Improvements over the base model

Paired evaluations on held-out tasks (same server, same seeds):

Capability	Improvement
Procedural generation	blockout gating, conditional lightmap, and validation workflows: FAIL → PASS across held-out tasks
Replay safety	base competence suite intact
Shape / hard-surface	consistent across held-out objects
Blender workflow	modifier-driven recipes, tooth profile generation, game-prep UV/PBR, non-destructive ordering
Lighting	layered lighting legibility (bounce and ambient)
Web-graphics	Canvas 2D / Three.js / WebGPU code generation measurably improved
UI/design	layout systems, component architecture, visual hierarchy, accessibility

Quantizations (MoQ)

All files quantized with the Mixture of Quantizations (MoQ) method proposed by Waleed Ahmad: per-tensor type selection (attention/embeddings at higher precision, MLP/experts at more aggressive types) instead of a single type for every tensor.

File	Approx. size	Notes
`Deepwen-3.6-Q2.5-MoQ.gguf`	13.4 GB (12.5 GiB)	aggressive MoQ mix, 2.7 bpw target
`Deepwen-3.6-Q3-MoQ.gguf`	13.4 GB (12.4 GiB)	3.0 bpw target
`Deepwen-3.6-Q4.5-MoQ.gguf`	21.2 GB (19.7 GiB)	4.5 bpw target, sweet spot for local use
`Deepwen-3.6-Q5-MoQ.gguf`	24.7 GB (23.0 GiB)	5.0 bpw target
`Deepwen-3.6-Q6-MoQ.gguf`	28.8 GB (26.8 GiB)	6.5 bpw target
`Deepwen-3.6-Q8-MoQ.gguf`	36.9 GB (34.4 GiB)	near-lossless (Q8_0 ≈ BF16 in practice)
`Deepwen-3.6-mmproj-BF16.gguf`	0.9 GB (0.84 GiB)	multimodal projector (BF16)
`Deepwen-3.6-mmproj-F16.gguf`	0.9 GB (0.84 GiB)	multimodal projector (F16)

Note: the Q6-MoQ was previously hidden due to corrupt offsets and has been re-uploaded and verified (byte-exact).

Original weights (BF16 / safetensors)

The original merged weights are NOT published as a GGUF in this repo. They live as safetensors in the companion repository:

quimmedes/Deepwen-3.6-bf16 — 26 shards (model-00001-of-00026.safetensors ... model-00026-of-00026.safetensors) plus config.json, tokenizer, chat_template.jinja, etc.

If you need a BF16/F32 GGUF, convert from the safetensors with convert_hf_to_gguf.py (llama.cpp), then re-quantize as needed. The GGUF quants above are derived from that exact checkpoint.

Chat templates

Two chat templates are shipped in this repo:

Template	File	Behavior
DeepSeek-style thinking (default, embedded in the GGUF)	`chat_template.jinja` (in-repo, or embedded in the GGUF)	Advanced reasoning with `reasoning_effort` control: `low` (default), `xhigh`, `max`. The `max` level is a relentless "think until verified" mode.
Original Qwen	`chat_template_original.jinja`	Standard Qwen template, no reasoning-effort injection. The model still thinks (`<think>`), but with the base-style behavior — no extra prompting layers.

If you see a lot of hallucination: switch to the original Qwen template (chat_template_original.jinja). It removes the aggressive reasoning-effort prompt injection, which often anchors the model too hard on its own chain-of-thought. The trade-off: you lose the max thinking mode and the reasoning-effort control (low/xhigh/max in chat_template_kwargs).

How to load each template

Embedded (default) — the GGUF already carries the DeepSeek-style template, no extra flag needed:

bash

llama-server -m Deepwen-3.6-Q4.5-MoQ.gguf --host 0.0.0.0 --port 8080
# thinking mode max:
curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
  "messages": [{"role": "user", "content": "design a hard-surface panel breakdown"},
               {"role": "assistant", "content": "<think>..."}],
  "chat_template_kwargs": {"reasoning_effort": "max"}
}'

Original Qwen template (anti-hallucination) — download the file from this repo, then override:

bash

wget https://huggingface.co/quimmedes/Deepwen-3.6/resolve/main/chat_template_original.jinja
llama-server -m Deepwen-3.6-Q4.5-MoQ.gguf --host 0.0.0.0 --port 8080 \
--temp 0.6 \
--jinja --chat-template-file chat_template_original.jinja

With LM Studio: copy chat_template_original.jinja next to the GGUF and name it <model-filename>.jinja (e.g. Deepwen-3.6-Q4.5-MoQ.gguf.jinja) — LM Studio picks it up automatically and uses it instead of the embedded template.

Anti-Repetitions

bash

wget https://huggingface.co/quimmedes/Deepwen-3.6/resolve/main/chat_template_original.jinja
llama-server -m Deepwen-3.6-Q4.5-MoQ.gguf --host 0.0.0.0 --port 8080 \
--temp 0.6 \
--top-p 0.9 \
--top-k 40 \
--repeat-penalty 1.10 \
--repeat-last-n 512 \
--dry-multiplier 0.8 \
--dry-base 1.75 \
--dry-allowed-length 2 \
--jinja --chat-template-file chat_template_original.jinja