Kimi K2 Thinking

Kimi K2 Thinking is the first open-weights model to beat state-of-the-art closed models including GPT-5 and Claude 4.5 Sonnet across HLE (44.9%), BrowseComp (60.2%), and SWE-Bench Verified (71.3%). Built with native INT4 Quantization-Aware Training, it delivers 2× generation speed and halved memory requirements versus FP8 variants. The 256K context window enables 200–300 sequential tool calls without human intervention across complex multi-step workflows. Deployment requires 512GB RAM and ≥32GB VRAM for 4-bit precision at a 600GB model size, with day-0 support for vLLM and MLX on Mac. Independent verification by Artificial Analysis confirmed a top ranking on Tau2 Bench Telecom at 93%, validating Moonshot's claims beyond self-reported data. The November 2025 release marks a potential inflection point for open-source vs closed model parity in frontier AI capabilities.

VentureBeat reports Moonshot AI released Kimi K2 Thinking on November 6, 2025, outperforming GPT-5, Claude Sonnet 4.5 (Thinking), and Grok-4 on several reasoning, coding, and agentic-tool benchmarks. The MoE model has one trillion total parameters with 32B activated per inference and executes up to 200–300 sequential tool calls autonomously. The model is released under a Modified MIT License granting full commercial and derivative rights, with a light-touch attribution clause only triggering above 100 million MAUs or $20M monthly revenue. Weights and code are hosted on Hugging Face, with APIs accessible via platform.moonshot.ai and a chat interface on kimi.com.

Kimi K2 Thinking is Moonshot AI's open-source reasoning model that scores 44.9% on Humanity's Last Exam and 60.2% on BrowseComp, outperforming GPT-5, Claude, and DeepSeek on those benchmarks. The model chains 200–300 sequential tool calls while maintaining coherent reasoning, addressing the degradation most models show after dozens of prompts. Its 1-trillion parameter MoE architecture activates only 32 billion parameters per inference, and full open weights enable inspection and fine-tuning. Released in November 2025, K2 Thinking represents a shift from conversational speed toward sustained multi-step problem-solving, which the Lambda blog frames as pivotal for production agentic workloads. Benchmark comparisons show it trails Claude only on SWE-bench Verified (71.3% vs 77.2%) but leads on HLE and BrowseComp. Lambda notes that production deployment still requires serious GPU capacity despite open availability.

Interconnects analyst Nathan Lambert highlights Kimi K2 Thinking as a 1T-total, 32B-active parameter reasoning MoE with 256K context length and interleaved thinking in agentic tool-use. The model beats leading closed models on benchmarks like Humanity's Last Exam and BrowseComp, though GPT-5 and Claude Sonnet 4.5 still top several evaluations. Lambert notes the writing style of the original Kimi K2 Instruct model is preserved through extended thinking RL training, making K2 Thinking a joy to use. He frames the release as the closest open models have come to the closed frontier of performance, underscoring the rapid rise of Chinese AI labs.

Kimi K2 Thinking is the latest open-source thinking model from Moonshot AI, built as a thinking agent that reasons step-by-step while dynamically invoking tools. It is a 1-trillion parameter Mixture-of-Experts model activating 32B per inference, with native INT4 quantization and a 256K context window enabling lossless speed-up in low-latency mode. The model sets new state-of-the-art on HLE, BrowseComp, and other benchmarks by scaling multi-step reasoning depth and maintaining stable tool-use across 200–300 sequential calls. It supports deep thinking with tool orchestration, stable long-horizon agency that surpasses prior models degrading after 30–50 steps, and is distributed under a modified MIT license on Hugging Face.

Moonshot AI introduced Kimi K2 Thinking as its best open-source thinking model, built as a thinking agent that reasons step by step while using tools. The model achieves state-of-the-art performance on Humanity's Last Exam (44.9% with tools), BrowseComp (60.2%), and SWE-Bench Verified (71.3%), with major gains in reasoning, agentic search, and coding. K2 Thinking can execute 200–300 sequential tool calls without human interference, reasoning coherently across hundreds of steps to solve complex problems. It marks Moonshot's test-time scaling effort by scaling both thinking tokens and tool calling steps, and is live on kimi.com with API access.

BenchmarkList's aggregator page for MoonshotAI Kimi K2 Thinking lists 70 benchmark rows with a snapshot timestamp of 2026-06-10. Agentic evaluations show a median 29th-percentile placement across 17 evals, including Tau2-Bench Telecom Success Rate at 93.0% (89th percentile), Terminal-Bench Hard Success Rate at 31.1% (7 The page includes comparison columns referencing other models such as Fable 5.1, Claude Opus 5, Kimi K3, Qwen3.8-2.4T-A95B, Qwen3.8-Flash-Next, GLM 5.3, and GLM 5.3 Flash, with field-leader deltas computed against them — for example, Fable 5 leading Terminal-Bench Hard at 62.9%. Because some of these comparator model n

CNBC reported on November 6, 2025 that Beijing-based Moonshot AI, backed by Alibaba, released Kimi K2 Thinking — its second major model update in four months — with claims of superior "agentic" capabilities versus OpenAI's ChatGPT. The model can automatically select 200 to 300 tools to complete tasks without explicit s Kimi K2 Thinking launched amid escalating U.S.–China AI competition, with Nvidia CEO Jensen Huang publicly urging the U.S. to maintain pace. CNBC frames Moonshot's positioning alongside DeepSeek, noting both Chinese labs have emphasized cost efficiency relative to U.S. hyperscalers. The article does not provide API spe