K2-Think is a 32B open-weights general reasoning model with strong competitive math performance. Benchmarks: AIME 2024 90.83, AIME 2025 81.24, GPQA-Diamond 71.08, LiveCodeBench v5 63.97.
Added Jul 26, 2025
Context Window
128.0K
Max Output
32.8K
Input Price (Auto)
$0.17/1M
Output Price (Auto)
$0.68/1M
Cache Read (Auto)
$0.085/1M
Benchmarks
Benchmarks
Performance metrics and benchmarks
No benchmark data is available yet for this model.
Providers
Auto routing is available for this model. Explicit provider selection is not available.
Loading provider options…
Related text models
Compare K2-Think with similar models from the same provider or model family.
Qwen3.8 Max 0902
alibaba/qwen3.8-max-0902Qwen3.8 Max 0902 is Alibaba's September 2 checkpoint of its flagship Qwen3.8 Max model for coding, knowledge work, data analysis, and long-running agent workflows. It supports text, image, video, PDF input, selectable thinking, tool calling, structured output, and a near-million-token context window.
Gemini 3.8 Flash
google/gemini-3.8-flashGoogle's fast multimodal model for agentic workloads, including coding, tool use, image understanding, PDF and document extraction, audio, and video. Its capabilities, limits, reasoning behavior, and pricing currently mirror Gemini 3.7 Flash.
Muse Spark 1.3
meta/muse-spark-1.3Meta's Muse Spark 1.3 is a frontier multimodal reasoning model for long-horizon coding and agentic workflows, with strong gains in computer use, browsing, professional tool use, codebase understanding, instruction following, and million-token retrieval. It accepts text, images, audio, video, and files, supports tool calling and structured output, and always reasons before answering.
Muse Spark 1.3 Contributor
meta/muse-spark-1.3-contributorMeta's Muse Spark 1.3 Contributor is a frontier multimodal reasoning model for long-horizon coding and agentic workflows, with strong gains in computer use, browsing, professional tool use, codebase understanding, and million-token retrieval. It accepts text, images, audio, video, and files, supports tool calling and structured output, and always reasons before answering. Prompts and outputs may be used by Meta for training and to improve its products.
Claude Fable 5.1
anthropic/claude-fable-5.1Claude Fable 5.1 improves on Fable 5 across agentic coding, long-running workflows, front-end and visual code generation, finance, analysis, and knowledge work, with more concise plans and summaries. Anthropic retains prompts and outputs for 30 days; Zero Data Retention is not available.
Mercury 2.5 Preview
inception/mercury-2.5-previewMercury 2.5 Preview is Inception's latest and most intelligent diffusion language model. Instead of generating tokens strictly one at a time, it produces and refines multiple tokens in parallel, reaching up to 1,107 tokens per second on standard GPUs. It delivers a 10+ point intelligence gain over Mercury 2, with tunable reasoning, parallel tool calls, schema-aligned JSON output, and a 260K context window. It is built for latency-sensitive production work such as search agents, voice pipelines, customer support, rapid coding iteration, and coding subagents.