Kimi K2 Thinking

2025-11-06 · Publisher : Moonshot AI · Domain : Artificial Intelligence

Moonshot AI releases Kimi K2 Thinking, an open agentic reasoning model with 1 trillion parameters, capable of 200 to 300 chained tool calls and natively quantized in INT4.

Moonshot AI releases Kimi K2 Thinking, an open agentic reasoning model with 1 trillion parameters, capable of 200 to 300 chained tool calls and natively quantized in INT4. ### Context In November 2025, GPT-5 and Claude Sonnet 4.5 dominate reasoning with tools, and open reasoning models are still lagging behind. Kimi K2 Thinking extends Kimi K2 (July 2025) by scaling both thinking tokens and tool calls. ### What the model brings - Step-by-step reasoning interleaved with tools, stable over 200 to 300 sequential calls without human intervention. - Native INT4 quantization (QAT, quantization-aware training) applied to the MoE components in post-training: about 2 times faster in generation, results published in INT4. - 44.9% on Humanity's Last Exam (HLE) with tools, 60.2% on BrowseComp, 71.3% on SWE-bench Verified according to the vendor. - “Heavy” mode (aggregation of trajectories): 51.0 on HLE, 100 on AIME25. - Input price cut announced by the platform (up to 75%) at the time of release. ### Architecture and specifications 1T-parameter MoE, 32B active, 61 layers, 384 experts including 8 selected and 1 shared, MLA attention, 256K context, text only. Checkpoints in compressed-tensors format (INT4). License: modified MIT (“Kimi K2” mention to be displayed beyond 100 million monthly active users or 20 million dollars in monthly revenue). Weights: huggingface.co/moonshotai/Kimi-K2-Thinking. ### Cost and access | Item | Pricing (USD per million tokens) | |---|---| | Input | 0.60 | | Input (cache hit) | 0.15 | | Output | 2.50 | Pricing from a secondary source (OpenRouter, provider price): the official platform only indicates a cut of up to 75% in input prices, and the current pricing page no longer lists this model (discontinued). Hardware: not disclosed. ### Benchmarks Scores published by the vendor (Hugging Face card), with settings specified in the card's notes. The values marked with an asterisk in the source are re-runs by the vendor. | Benchmark | K2 Thinking | GPT-5 (High) | Claude Sonnet 4.5 (Thinking) | |---|---|---|---| | HLE, with tools | 44.9 | 41.7 | 32.0 | | BrowseComp | 60.2 | 54.9 | 24.1 | | GPQA | 84.5 | 85.7 | 83.4 | | AIME25 without tools | 94.5 | 94.6 | 87.0 | | SWE-bench Verified | 71.3 | 74.9 | 77.2 | | LiveCodeBench v6 | 83.1 | 87.0 | 64.0 | | Terminal-Bench (simulated tools) | 47.1 | 43.8 | 51.0 | Source: Hugging Face card of Kimi-K2-Thinking (vendor measurements). ### Versus closed models On agentic search and HLE with tools, K2 Thinking outperforms GPT-5 (High) and Claude Sonnet 4.5 in this table (for example 60.2 versus 54.9 and 24.1 on BrowseComp). It remains behind on coding: 71.3 on SWE-bench Verified versus 74.9 and 77.2, and on LiveCodeBench against GPT-5. On GPQA and AIME25, it is close to GPT-5. The vendor claims state-of-the-art on HLE and BrowseComp; no third party is cited to confirm these figures in the sources consulted. Open weights and native INT4 make self-hosting more realistic than a 1T model at full precision, though the required hardware is not published. ### Key takeaways Kimi K2 Thinking is an open model presented as ahead of GPT-5 and Claude Sonnet 4.5 on agentic search, according to the vendor. It highlights native INT4 as a way to make a trillion parameters usable.

Tags : LLM, Open Source, Moonshot AI, Kimi, MoE, MIT modifiée, Raisonnement, INT4