Deepwen-3.6-bf16

Deepwen 3.6 (BF16)

Deepwen 3.6 is a fine-tuned derivative of Qwen/Qwen3.6-35B-A3B (MoE: 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 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

Format

• safetensors, bfloat16 (26 shards, ~67 GB)
• Base dtype: bf16; Mamba SSM tensors in f32
• Includes the full model (no MTP layers stripped at this level)

Usage

Standard transformers / vLLM / SGLang loading:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("quimmedes/Deepwen-3.6-bf16", torch_dtype="bfloat16")
tok = AutoTokenizer.from_pretrained("quimmedes/Deepwen-3.6-bf16")

To enable advanced thinking at maximum effort:

{
  "messages": [{"role": "user", "content": "..."}],
  "chat_template_kwargs": {"reasoning_effort": "max"}
}

Quantized versions

GGUF quantizations with MoQ (mixed) tensor types are published in:

quimmedes/Deepwen-3.6 — levels Q2.5 (IQ2_XXS), Q3, Q4.5, Q5, Q6, Q8, all without MTP (733 tensors), plus vision mmproj (BF16/F16).

Disclosures

• Base model: Qwen/Qwen3.6-35B-A3B — Copyright © Alibaba Group / Qwen Team. All rights to the base model and its weights remain with the original authors.
• The base model is released under the Apache License 2.0; this derivative inherits that license.
• Qwen 3.6 is a copyrighted, trademarked model family of Alibaba. "Deepwen 3.6" is an independent fine-tuned derivative and is not affiliated with, endorsed by, or sponsored by Alibaba / Qwen. The "Qwen" name is used solely to identify the base model.
• The embedded reasoning-effort prompts are adapted from the chat template of deepseek-ai/DeepSeek-V4-Flash-0731; DeepSeek remains the copyright holder of those prompt texts.
• MoQ quantization method: "Mixture of Quantizations" proposed by Waleed Ahmad (https://huggingface.co/w-ahmad).
• This model is provided as-is, without warranties of any kind, for research and local experimentation.