Qwen3.5-Coder

Coding-optimized variants of the official Qwen3.5 MoE models — full vision capability retained, tuned for precise code generation via lower temperature. Based on Alibaba’s Qwen3.5 distributed through the Ollama registry.

Two tags available:

Same architecture. Same quantization. Coding-optimized temperature across both.

What is this?

Qwen3.5 is a hybrid Mixture-of-Experts model from Alibaba’s Qwen team featuring a novel Gated DeltaNet + sparse MoE architecture. Despite large total parameter counts, only ~3B are activated per token, making inference efficient regardless of model size.

These Coder variants take the official Ollama Qwen3.5 models and tune their parameters for coding workloads — lower temperature for more deterministic output, presence penalty to reduce repetition. Vision capability is fully preserved.

Why Coder?

OpenCode and similar coding tools don’t support per-session parameter overrides — they use whatever is baked into the Ollama model. These variants provide coding-optimized defaults without sacrificing any model capability.

• Use :35b for fast, capable coding on standard hardware
• Use :122b for maximum reasoning depth on high-memory systems

Architecture

The 122b model has a larger embedding dimension (3072 vs 2048), enabling richer per-token representations and stronger reasoning depth for complex problems.

Capabilities

• Completion
• Vision (image input supported)
• Tool calling
• Thinking (extended reasoning mode)

Usage

35B — Fast coding

ollama pull mdq100/qwen3.5-coder:35b
ollama run mdq100/qwen3.5-coder:35b

Parameters:

temperature:      0.6
top_p:            0.95
top_k:            20

122B — Maximum reasoning

ollama pull mdq100/qwen3.5-coder:122b
ollama run mdq100/qwen3.5-coder:122b

Parameters:

temperature:      0.6
top_p:            0.95
top_k:            20
num_ctx:          131072

OpenCode

Project-level (:35b):

{
  "model": "ollama/mdq100/qwen3.5-coder:35b"
}

Project-level (:122b for complex tasks):

{
  "model": "ollama/mdq100/qwen3.5-coder:122b"
}

Benchmarks

Scores from the base Qwen3.5 model (BF16, full precision). Q4_K_M quantization may show minor variance (~1-2%).

Coding

Knowledge & Reasoning

Instruction Following

Long Context

Vision

Comparison: Coder vs Flash-Coder

Use mdq100/qwen3.5-coder when you need vision support or prefer Ollama-validated builds. Use mdq100/qwen3.5-flash:35b-code for better quantization quality without vision.

Credits

• Original model: Qwen3.5 by Alibaba Qwen Team
• Base distribution: Ollama registry
• Coder variant (coding-optimized parameters): packaged for Ollama + OpenCode