Kimi

Why this family matters

Moonshot AI's Kimi line solves two distinct problems with two distinct models. Kimi K2 is the chat-and-tools workhorse, and the K2.6-thinking variant lands at Quality Score 96.1 (#19 of 197 models we track), which puts it inside the open-weights frontier cluster against the strongest open competitors. Kimi VL is the vision-language line.

The structural fact pulling teams onto K2 is the long-context profile. The 0905 checkpoint ships with a 262K-token context window at $0.6 input / $2.5 output per million, and the K2.6-thinking variant pairs strong reasoning quality with that same context envelope. For teams running document-heavy or long-conversation workloads on the open-API side, that combination is the headline reason to evaluate.

Which variant to start with

Default to moonshot-kimi-k2 for text-only workloads. The 0905 checkpoint at 262K context and $0.6 / $2.5 per million is the practical entry point. Step up to the K2.6-thinking variant (QS 96.1, $0.57 / $2.3 per million) when the workload visibly benefits from explicit reasoning behaviour.

When to deviate:

• Image-grounded document workloads: consider moonshot-kimi-vl-a3b, the family's vision-language line. Use it when the workload is layout-aware extraction or image-grounded reasoning where running OCR to text and then a chat-tier LLM loses information. Caveat: benchmark coverage for Kimi VL in our current index is incomplete (no pricing, no context window, no scores at last verification). Treat this variant as a directional pick to evaluate against your own data, not a pre-validated recommendation.
• Long chain-of-thought reasoning: K2.6-thinking is the family's reasoning ceiling on our index. Compare against DeepSeek R1 and Claude Opus 4.5+ thinking on the specific benchmark that matters before committing.
• Cheapest Moonshot tier: the K2 0711 and base checkpoints are slightly older but priced at $0.57 / $2.3 per million with 131K context. The 0905 update moves both context (131K to 262K) and price (cheaper input, same output) in the right direction; there is no strong reason to start on 0711 unless a deployment is already pinned.

Where the data is weak

We aggregate benchmark scores from multiple sources but coverage on this family is uneven. Specifically:

• Kimi VL has effectively no public benchmark or pricing data in our current pull. The variant is registered (Kimi-VL-A3B with thinking and non-thinking modes) but context window, per-million pricing, and per-benchmark scores are all unset. We surface it for completeness; treat it as exploratory until the data fills in.
• K2 has many minor checkpoints (0711, 0905, base, instruct, thinking, thinking-turbo, K2.5-thinking, K2.6-thinking). The difference between K2 0711 (QS 71.1) and K2.6-thinking (QS 96.1) is large enough that "Kimi K2 quality" is not a useful single number; the variant table is the load-bearing artifact, not a family-level Quality Score.
• Pricing on this page is the published API list price. Moonshot ships through several inference providers in addition to its own API; list price is a calibration anchor, not the cost ceiling.
• Long-context behaviour on K2 deserves its own evaluation. A 262K context window in the API does not mean uniform recall across that range; verify with a needle-in-haystack-style test on your specific document distribution before committing.

If you are making a procurement decision, the variant table on this page is the load-bearing artifact. Cross-check pricing against Moonshot's own docs before you commit.

When to reach for which alternative

• Cheapest competent long-context API: DeepSeek V4 Flash ships 1M context at $0.094 / $0.188 with QS 75.7, which is a stronger cost-and-context anchor than K2 0905 if quality is acceptable at Flash's tier. K2.6-thinking still beats Flash on raw quality score; the choice depends on whether the score delta or the context-cost delta matters more for the workload.
• Closed-flagship reasoning at the top end: Claude Opus 4.5+ thinking and full GPT-5 are the anchors to compare against if peak quality on a single reasoning benchmark is the binding constraint.
• You need vision-language data you can rely on today: the Kimi VL data gap on our index is real; if you cannot wait for our coverage to fill in, evaluate against vision-language variants from families with stronger published benchmarks (e.g. the Qwen3-VL surface in our index, when it ships).

Sources worth reading

• Moonshot AI: authoritative pricing and product information
• Kimi on Hugging Face: model cards, weights, and per-variant license terms
• Kimi K2 model card: primary source for K2 architecture and intended use