GLM-4.1V-9B-Thinking

// Overview

GLM-4.1V-9B-Thinking is an open-weight vision-language model released on July 1, 2025 by Z.ai (the Zhipu AI / GLM team) together with Tsinghua University's KEG lab. It is the first reasoning-focused entry in the GLM-V multimodal line, built on the GLM-4-9B-0414 base and paired with an AIMv2-Huge vision encoder. Despite being a roughly 9-billion-parameter model, it is positioned to compete with much larger systems on multimodal reasoning.

The defining feature is its 'thinking' paradigm: GLM-4.1V-9B-Thinking produces an explicit chain-of-thought before answering, which the team trained with a method they call Reinforcement Learning with Curriculum Sampling (RLCS). The model accepts text, images, and video, supports a 64K-token context, and handles arbitrary aspect ratios and image resolutions up to 4K via 2D-RoPE positional encoding.

Because the weights ship under an MIT license, GLM-4.1V-9B-Thinking can be downloaded, fine-tuned, and self-hosted without API fees, and its small size makes it practical to run on a single modern GPU. It is also offered as a hosted API through Z.ai's platform and third-party inference providers such as SiliconFlow for teams that prefer not to manage their own infrastructure.

// Benchmarks

1. MMStar72.9%
2. MathVista80.7%
3. MMMU68%
4. MMMU-Pro57.1%
5. AI2D87.9%
6. MMBench-V1.1-EN85.8%
7. OCRBench84.2%
8. VideoMME (with subtitles)73.6%
9. WeMath63.8%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Hosted API pricing via SiliconFlow. The model weights are open under MIT, so self-hosting incurs no per-token fee.

Pricing source ↗

// Strengths

• Open-weight (MIT) — free to download, fine-tune, and self-host with no usage restrictions
• Strong multimodal reasoning for its size: leads 10B-class models on 23 of 28 reported benchmarks and beats Qwen2.5-VL-72B on 18 of them
• Explicit chain-of-thought 'thinking' output improves accuracy and interpretability on STEM and math-vision tasks
• Handles arbitrary aspect ratios and up to 4K image resolution, plus video input
• Small enough (9B) to run on a single GPU, keeping inference cheap

// Best for

• Solving STEM and math problems presented as images, diagrams, or charts
• Document and chart understanding, including OCR-heavy and long-document pages
• Video understanding and question answering
• GUI / screenshot understanding and UI navigation agents
• Self-hosted multimodal reasoning where data privacy or cost rules out a closed API

// How to access

// GLM-V (vision-language) — every version

The full lineage of the GLM-V (vision-language) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.