Qwen3.5-Flash 35B

A text-only, thinking-capable variant of Qwen3.5-35B-A3B — leaner and faster by removing the CLIP vision projector. Based on Unsloth’s Q4_K_M quantization of Alibaba’s Qwen3.5-35B-A3B.

Two tags are available under this model:

Same weights. Same architecture. Different temperature.

What is this?

Qwen3.5-35B-A3B is a hybrid Mixture-of-Experts model from Alibaba’s Qwen team featuring a novel Gated DeltaNet + sparse MoE architecture. Despite 34.7B total parameters, only ~3B are activated per token, making inference efficient.

This Flash variant strips the CLIP vision projector to produce a clean, text-only model. The LLM weights are unchanged — only vision input is removed.

Why Flash?

The original Qwen3.5-35B-A3B includes a 446M-parameter CLIP vision encoder. Removing it: - Eliminates vision input (no image processing) - Reduces load time and memory overhead - Avoids compatibility issues with vision loading in current Ollama versions - Keeps the full language and reasoning capability intact

Why two tags?

OpenCode and similar coding tools don’t support per-session parameter overrides — they use whatever is baked into the Ollama model. A dedicated coding tag avoids manually tuning parameters per session.

• Use :35b for general reasoning, chat, and instruction following
• Use :35b-code for coding inside OpenCode or similar assistants

Architecture

Capabilities

• Completion
• Tool calling
• Thinking (extended reasoning mode)

Usage

General purpose

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

Parameters:

temperature: 1.0
top_p: 0.95
top_k: 20
presence_penalty: 1.5

Coding optimized

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

Parameters:

temperature: 0.6
top_p: 0.95
top_k: 20

OpenCode

Add to your project’s opencode.json:

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

Or globally in ~/.config/opencode/opencode.json.

Benchmarks

Scores below are for the base Qwen3.5-35B-A3B model (BF16, full precision). Q4_K_M quantization may show minor variance (~1-2%).

Coding

Knowledge & Reasoning

Instruction Following

Long Context

Math & STEM

Multilingual

Vision benchmarks (MMMU, MathVision, etc.) are not applicable to this Flash variant.

Unsloth Improvements

• Unsloth Dynamic 2.0 — improved quantization algorithm that preserves more accuracy at lower bit depth, outperforming standard Q4_K_M
• Improved imatrix calibration — curated calibration data covering chat, coding, long context, and tool-calling
• Reduced KL divergence — output distribution stays closer to the full-precision original
• Early availability — published within hours of the original model release

Credits

• Original model: Qwen3.5-35B-A3B by Alibaba Qwen Team
• GGUF quantization: Unsloth
• Flash variant (vision removed, coding tag temperature-optimized): packaged for Ollama + OpenCode