Kimi K2 Thinking: Specs, Benchmarks (and the Cost Number Nobody Publishes)

Kimi K2 Thinking is Moonshot AI open-weights reasoning model: a 1-trillion-parameter Mixture-of-Experts with 32B active parameters, a 256K context window, and a Modified MIT license. But there are now several Kimi releases with confusingly similar names, and the benchmark that actually matters — cost per completed task — is one no vendor publishes. Here is what is real, what is vendor-reported, and one first-party number that puts Kimi on a same-tasks cost ladder against Claude, GPT, and DeepSeek.

What Kimi K2 Thinking actually is

Kimi K2 Thinking is the reasoning and agentic variant of Moonshot AI open-weights K2 model family. According to its Hugging Face model card, it is a Mixture-of-Experts network with 1 trillion total parameters and 32 billion active per token — 384 experts, 8 selected plus 1 shared, 61 layers, MLA attention, and a 160k vocabulary. The context window is 256K tokens.

Two things make it interesting for builders. First, it ships with native INT4 quantization via Quantization-Aware Training, which Moonshot says roughly doubles generation speed with lossless quality. Second, it is released under a Modified MIT License with open weights on Hugging Face and an API at platform.moonshot.ai that exposes OpenAI- and Anthropic-compatible endpoints. In other words, you can self-host it or hit it through a gateway without rewriting your client.

The design intent is long agentic tool-calling chains — the kind of workload where a model plans, calls tools, reads results, and keeps going for dozens of steps. That framing matters when you read its benchmarks, because most of its best numbers are reported with tools enabled.

The naming trap: Thinking vs K2.7-Code

Before any benchmark makes sense, you have to untangle the names. Moonshot has shipped roughly five K2 releases in under a year, and the search-popular ones do not share numbers.

• Kimi K2 Thinking — the reasoning/agentic variant with published public-suite scores (HLE, SWE-bench Verified, LiveCodeBench, GPQA). These are the numbers most articles quote.
• Kimi K2.7-Code — a separate, newer release that shipped June 12, 2026. Also a 1T MoE with 32B active and 256K context under Modified MIT. Its headline gains at launch were on Moonshot-proprietary benchmarks, not standard suites.

This is the single biggest source of misinformation about Kimi right now. The SWE-bench 71.3 figure belongs to K2 Thinking. The executed-cost datapoint later in this article belongs to K2.7-Code. Mixing them is how you end up quoting a coding score for the wrong model. We keep them labeled throughout.

Specs and vendor benchmarks

Here are the core specs alongside Moonshot own reported benchmark figures for K2 Thinking. Everything in the benchmark rows is vendor-reported — Moonshot self-published, mostly with tools enabled — so read it as a capability signal, not an independent ranking.

Independent tracker Artificial Analysis places K2 Thinking at 41 on its current Intelligence Index — above the roughly 30 open-weights average, and it registered 67 at its November launch on the prior scale — but flags a real cost hazard: it is unusually verbose. It burned about 100M tokens to complete the index, among the most Artificial Analysis has recorded and roughly 2.5x the average model. That matters because a low per-token price can still produce a high per-task bill if the model writes far more tokens. Headline pricing hides this; executed-cost benchmarks do not.

K2.7-Code, by contrast, leaned at launch on Moonshot-proprietary benchmarks: Kimi Code Bench v2 62.0 (vs 50.9 for K2.6), Program Bench 53.6, and MLS Bench Lite 35.1, plus a claim of roughly 30% fewer reasoning tokens than K2.6. Those launch figures cannot be reproduced or compared to SWE-bench, so we do not rank on them; independent Intelligence Index and public-suite numbers have only appeared since. For a broader view of open models, see our best open-source LLM 2026 roundup.

The cost number nobody publishes

Every outlet reports either Moonshot proprietary benchmarks or generic vendor SWE-bench numbers. Nobody puts Kimi on a same-tasks, actually-executed cost ladder against closed frontier models. That is the gap the DataLLM Lab first-party benchmark fills.

In July 2026 we ran 13 models against 9 tasks where each model must generate code, then have it pass hidden tests. Cost is measured in real dollars per 1,000 tasks; latency is average seconds per task. Kimi K2.7-Code scored a perfect 9/9 at $1.34 per 1,000 tasks and 10.4s per task — squarely mid-pack.

The headline is not Kimi rank — it is that 10 of 13 models scored a perfect 9/9. Quality has converged on standard coding tasks, so the decision variable is no longer the leaderboard; it is cost per completed task, where the spread is 88x for the same result. Full methodology is in our LLM coding cost benchmark.

Where it sits: DeepSeek, GLM, Claude

On the open-weights reasoning/coding landscape, Kimi K2, DeepSeek V4, and GLM sit within a few points of one another on composite trackers. On the current Artificial Analysis Intelligence Index, K2.6 registers 44, with DeepSeek V4 and GLM close behind, and all three land near 80% on vendor SWE-bench Verified. These are directional, secondary-source numbers — treat them as a cluster, not a precise order.

The differences are practical, not on the scoreboard: Kimi is often favored for single-shot, hard tickets; DeepSeek for long-horizon agentic runs and the cheapest API coding. If you are weighing those two directly, our Kimi vs DeepSeek comparison goes deeper, and DeepSeek vs Claude covers the open-versus-frontier tradeoff. For token pricing specifics, see Kimi API pricing.

On pricing as of July 2026: Moonshot-direct K2 Thinking runs about $0.60 per 1M input and $2.50 per 1M output. K2.7-Code is higher — observed reseller rates are roughly $1.10 input and $4.80 output. All of this is provider-dependent and moving fast, so state it qualitatively and re-check before you commit.

A decision rule for picking Kimi

Given quality convergence, here is a simple rule that survives the next Moonshot release:

• Standard coding tasks: pick on cost-per-completed-task, not benchmark rank. Kimi K2.7-Code at 9/9 for $1.34/1,000 is a strong open-weight middle option — cheaper than the Claude/GPT frontier, pricier than Qwen3 Coder Next or DeepSeek V4-Flash.
• Hard, single-shot reasoning tickets: K2 Thinking earns its verbosity; budget for the extra tokens it generates.
• Long-horizon agentic pipelines: weigh DeepSeek V4, and always measure real per-task cost — a low per-token price plus high verbosity can invert the ranking.
• Data residency / self-hosting: the Modified MIT open weights are the differentiator versus closed frontier models.

Because Moonshot ships fast, do not memorize a score. Memorize the framing: benchmarks have saturated on standard tasks, so measure cost per completed task on your workload before choosing.

FAQ