Moonshot's Kimi K2 Thinking Puts an Open-Weight Model at the Top of the Agent Charts

Beijing-based Moonshot AI released Kimi K2 Thinking on November 6, 2025, an open-weight reasoning model built to work like an agent: it plans, calls tools, reads the results and keeps going. On Moonshot's own tests it edged past closed frontier models on several research-style benchmarks while costing far less to run through the API.

What Moonshot shipped

K2 Thinking is a mixture-of-experts model with roughly one trillion total parameters, of which about 32 billion activate for each token. It supports a 256k-token context window and was trained and served with native INT4 quantization, which keeps inference costs down. The weights are on Hugging Face under a modified MIT license that allows commercial use, with one condition: products above 100 million monthly users or $20 million in monthly revenue must display "Kimi K2" in the interface.

The headline numbers come from Moonshot's own evaluations. The model scored 44.9% on Humanity's Last Exam with tools enabled and 60.2% on BrowseComp, a test of persistent web research. It also reported 71.3% on SWE-Bench Verified and 83.1% on LiveCodeBench v6. No independent lab had reproduced these at launch, so treat them as vendor claims.

Built for long tool loops

The pitch is less about raw knowledge than stamina. Moonshot says K2 Thinking can chain 200 to 300 sequential tool calls without a human nudging it, interleaving reasoning with search, a code interpreter and browsing. That matters because agentic failures usually come from drift: after a few dozen steps, models lose the plot. A model that holds its goal across hundreds of steps can attempt jobs such as multi-source research or debugging a repo end to end.

Price is the second story

Moonshot's API charges $0.60 per million input tokens on a cache miss ($0.15 on a hit) and $2.50 per million output tokens. GPT-5 listed at $1.25 in and $10 out. For agent workloads, which burn through output tokens, a four-fold gap compounds quickly. And because the weights are public, companies with the hardware can host the model themselves, which sidesteps per-token pricing and data-residency worries altogether.

Why it matters

For two years the open-versus-closed question has been whether public models would trail the frontier by a comfortable margin. K2 Thinking suggested the margin was shrinking on agentic tasks, and it fit a broader push by Chinese labs toward open weights. It also raised a measurement problem: leaderboards built on vendor-run evaluations are hard to compare, since tool access, token budgets and scaffolding vary.

The benchmark mix also flatters the model's strengths. BrowseComp rewards patience and search; it says little about coding on a large repository or about safety behavior. A fair read is that K2 Thinking is a credible open alternative for research agents, not proof that open models have caught every closed one.

Reading the other numbers

BrowseComp is only one row of the scorecard. Moonshot also reported 85.7% on GPQA Diamond, a graduate-level science test, against 84.5% it listed for GPT-5, and said the model matched GPT-5 on AIME 2025 and HMMT 2025, two math competitions. On Seal-0, a benchmark of search-heavy questions with noisy results, it posted 56.3%. Those are strong figures for a model whose weights anyone can download, but the comparisons are drawn from the vendor's own harness. The VentureBeat write-up notes that GPT-5 regains parity in some heavy-mode configurations, which is the usual way a lead on a single benchmark narrows once rivals tune their settings.

The architecture explains some of the economics. A mixture-of-experts model routes each token through a small slice of its parameters, here about 32 billion out of a trillion. Compute per token tracks the active slice, while memory requirements track the full size. That is why the model is cheap to call through an API run by a provider with the right hardware, yet far from trivial to host on a single workstation. Native INT4 quantization shrinks the memory footprint and speeds decoding, which Moonshot treats as a feature of the training recipe rather than an afterthought.

There is also a competitive-dynamics point. Chinese labs have leaned into open weights as a distribution strategy: free downloads build developer mindshare, and revenue comes from hosted inference and enterprise deals. Western labs mostly keep their strongest models behind APIs, so developers who want to fine-tune, audit or run models on their own servers have increasingly looked east. Whether procurement teams will accept that, given licensing terms and geopolitical scrutiny, is a separate question from whether the benchmark numbers hold up.

What to watch

• Independent reproductions on BrowseComp and HLE from third-party evaluators.
• Whether US labs respond on price, or on their own open-weight releases.
• How enterprises weigh a modified-MIT license from a Chinese lab against self-hosting savings.
• Moonshot's follow-up models and whether the long tool-call horizon holds up outside benchmarks.

Sources

1. Moonshot's Kimi K2 Thinking emerges as leading open source AI, outperforming GPT-5, Claude Sonnet 4.5 on key benchmarks — VentureBeat
2. Kimi K2 Thinking — Simon Willison's Weblog
3. Kimi-K2-Thinking model card — Hugging Face
4. Kimi K2 Thinking: The open-source model giving closed AI labs a run for their money — TechTalks