Release Timeline

Major release competing with GPT-5.2 on coding. Ranked #1 on Code Arena among open-source and domestic models. AIME 2025: 95.7% accuracy.

• LiveCodeBench: 84.9% (beats Claude 4.5)
• SWE-bench Verified: 73.8% (SOTA open-source)

Reasoning and tool-using thinking agent. Can execute 200-300 sequential tool calls without human interference. State-of-the-art agentic capabilities.

• 200-300 sequential tool calls
• LiveCodeBench-v6: 83.1%

Uses Kimi Delta Attention (KDA) for efficient long-context processing. Reduces memory usage and improves generation speed at longer context windows.

• 1M token context window
• Novel Delta Attention mechanism

Latest V3 iteration with improved general capabilities. Enhanced reasoning and coding performance over V3.1.

• SWE-bench Verified: 73.1%
• Improved knowledge and academic tasks

1 trillion parameter MoE with 32B active. State-of-the-art open-source performance on coding benchmarks. Trained for $4.6M, rivaling ChatGPT and Claude.

• SOTA open-source coding performance
• Trained for only $4.6M

Coding-focused model based on Qwen2.5-72B. State-of-the-art among open source on SWE-bench Verified.

• SOTA open-source on SWE-bench Verified
• Built on Qwen2.5-72B foundation

Major reasoning upgrade. AIME 2025 accuracy improved from 70% to 87.5%. Performance approaching O3 and Gemini 2.5 Pro levels.

• Intelligence Index: 68 (Artificial Analysis)
• AIME 2025: 87.5%

Flagship MoE reasoning model. Hybrid thinking/non-thinking modes. Trained on 36T tokens across 119 languages. Outperforms DeepSeek-R1 on 17/23 benchmarks.

• State-of-the-art open-source reasoning model
• Hybrid thinking mode for complex problems

Open-source vision-language MoE model. Efficient multimodal understanding with only 3B active parameters.

• 16B MoE with 3B active
• Efficient multimodal processing

Significant benchmark improvements. MMLU-Pro: 75.9 to 81.2 (+5.3). GPQA: 59.1 to 68.4 (+9.3). AIME: 39.6 to 59.4 (+19.8).

• MMLU-Pro: 81.2
• GPQA: 68.4

Medium-sized reasoning model achieving performance comparable to DeepSeek-R1 (671B). Strong step-by-step reasoning capabilities.

• Competitive with 671B DeepSeek-R1 using only 32B parameters
• ArenaHard: 89.5

Medical domain specialized model. Significantly lower hallucination rate than general models - about 3x lower than DeepSeek.

• 3x lower hallucination rate
• Medical domain specialized

First major Kimi reasoning model. Matches OpenAI o1 in math, coding, and multimodal reasoning. Uses reinforcement learning for dynamic learning.

• Matches OpenAI o1 performance
• Free with no usage limits

Breakthrough MoE model. 671B parameters with only 37B activated per token. Innovative load balancing and multi-token prediction. Trained on 14.8T tokens.

• Outperforms GPT-4o and Claude 3.5 Sonnet on MMLU
• Cost-efficient MoE architecture

Specialized financial model. 95%+ accuracy in banking, insurance, funds, and securities. Overall 93.62% on FLAME-Cer benchmark.

• FLAME-Cer: 93.62%
• Surpasses GPT-4o on financial tasks

State-of-the-art open-source code LLM. Trained on 5.5T tokens. Performance matches GPT-4o on coding tasks. Supports 40+ programming languages.

• HumanEval pass@1: 92.7%
• Aider benchmark: 73.7 (comparable to GPT-4o)

Major upgrade with improved reasoning and 128K context. Surpasses Llama-3.1-405B on several benchmarks. MMLU improved from 84.2 to 86.1.

• Beats Llama-3.1-405B despite smaller size
• 128K context window

Multimodal vision-language model. Strong image understanding, document analysis, and visual reasoning capabilities.

• State-of-the-art open-source VLM
• Dynamic resolution support

Code generation model built on GLM-4-9B. IDE extensions available for VS Code, JetBrains, and more.

• IDE integration available
• Multi-language support

First open-source model to match GPT-4 Turbo on coding. Supports 338 programming languages. 128K context length.

• HumanEval: 90.2%
• MBPP: 76.2%

Efficient open-source version of GLM-4. Strong performance for its size with 128K context support.

• Apache 2.0 license
• 128K context

Cost-optimized version of Baichuan 4. Lowest deployment cost among similar models - runs on 2x RTX 4090.

• Runs on 2x RTX 4090
• Lowest cost in class

Vision-language model based on Llama 3. Strong image understanding and visual reasoning capabilities.

• Built on Llama 3
• Strong OCR capabilities

Fourth generation with 10%+ improvement in general capabilities. Math improved 14%, code improved 9%. Ranked #1 on SuperCLUE Chinese benchmark.

• SuperCLUE score: 80.64 (#1)
• Beats GPT-4-Turbo on Chinese tasks

Fourth generation GLM with significantly improved reasoning and Chinese language capabilities. 128K context window.

• 128K context window
• Strong Chinese understanding

Open-source bilingual model with strong Chinese capabilities. Good balance of performance and efficiency.

• Open source
• Strong Chinese performance