Moonshot AI

Key Terms

• Kimi K2.7 Code - Moonshot's coding-focused agentic model built on K2.6, released June 12, 2026. Mixture-of-Experts, 1T total / 32B activated parameters, 256K context. Reduces thinking-token usage ~30% versus K2.6 and reports +10% agentic performance. Source: Hugging Face – Kimi K2.7 Code
• Forced thinking / preserve_thinking - K2.7 Code cannot run with thinking disabled (it throws an error) and always retains full reasoning content across multi-turn calls. This differs from K2.6/K2.5, which support a non-thinking mode. Source: Kimi – Kimi K2 7 Code Quickstart
• HighSpeed variant - kimi-k2.7-code-highspeed, the same model as K2.7 Code but served faster (~180 tokens/s, up to 260 tokens/s in short context), priced at 2x the standard model. Resources are described as "limited" and performance may fluctuate. Source: Kimi – Chat K27 Code
• Modified MIT License - K2.7 Code's open-weight license. Like MIT but adds an attribution clause requiring products built on the model to disclose that use (e.g. in an "About" section). This is the clause that Cursor's Kimi-derived Composer models reportedly did not honor. Source: Hugging Face – Kimi K2.7 Code
• Token-based billing - charges based on input, output, and cached tokens processed. One token is roughly 3-4 English characters. Kimi automatically applies context caching. Source: Kimi – Chat
• Kimi Code Bench v2 / Kimi Claw 24/7 Bench - Moonshot's in-house benchmarks for coding agents and long-horizon agentic tasks respectively. Because they are vendor-authored, they should be read alongside independent benchmarks. Source: Hugging Face – Kimi K2.7 Code
• MCP (Model Context Protocol) - an open standard for connecting models to external tools and data. Two K2.7 Code agentic benchmarks (MCP-Atlas, MCPMark-Verified) measure tool use over MCP servers. Source: Hugging Face – Kimi K2.7 Code

Latest Changes

Changes since the 2026-05 report.

• New model: Kimi K2.7 Code launched June 12, 2026, branded as Kimi's "most capable coding model to date," at $0.95/$4.00 per MTok (input/output) with a $0.19/MTok cache-hit rate over a 256K context window. Released open-weight on HuggingFace. Source: Hugging Face – Kimi K2.7 Code , Kimi – Chat K27 Code
• New variant: K2.7 Code HighSpeed (kimi-k2.7-code-highspeed) launched the same day at 2x pricing ($1.90/$8.00 per MTok, $0.38 cache hit), targeting ~180 tokens/s output. Source: Kimi – Chat K27 Code
• Behavior change: K2.7 Code forces thinking mode and preserve_thinking on; disabling thinking throws an error. Temperature is fixed at 1.0 and top_p at 0.95, neither can be changed. K2.6/K2.5 still support non-thinking mode. Source: Kimi – Kimi K2 7 Code Quickstart
• Efficiency claim: Moonshot reports K2.7 Code reduces thinking-token usage ~30% versus K2.6 and improves agentic performance ~10%, though these are vendor figures from "external benchmark evaluations." Source: Hugging Face – Kimi K2.7 Code
• Promotion: K2.7 Code Launch Top-Up Rebate runs June 11 to July 2, 2026. A single API top-up of at least $100 earns a tiered voucher (20% for $100-299, 25% for $300-999, 30% for $1,000+), max $4,000, valid 90 days, one reward per org. Source: Kimi – Promotion
• Product (consumer): Moonshot launched the "world's first AI-native credit card" on June 30, 2026, embedding Kimi membership benefits into a credit card. It requires a Chinese mainland 11-digit mobile number, so it is China-only. Source: Kimi – Aicard , News – Item
• No pricing changes to existing models: K2.6 ($0.95/$4.00) and K2.5 ($0.60/$3.00) per-MTok prices and the API rate-limit tiers (Tier 0 through Tier 5, $1 to $3,000 cumulative recharge) are unchanged. Source: Kimi , Kimi – Limits
• Unresolved (flagged in May): No announcement was found resolving the May billing/reliability complaints (double-charging, missing invoices, 429 rate-limit errors, no visible cancel link). These remain open as far as public sources show. Source: Kimi
• Ecosystem: Cursor's Composer 2.x models remain confirmed RL fine-tunes of Kimi K2.5 (licensed via Fireworks.ai), and the K2.7 Code HN thread shows continued debate over whether that arrangement honored Moonshot's Modified MIT attribution clause. Source: News – Item
• Cursor/xAI claim, NOT verified: A claim that "SpaceX acquired Cursor in June" could not be confirmed from any fetched source. No acquisition of Cursor by SpaceX or xAI appears in the platform pages, HN search results for Kimi/Moonshot, or Cursor-related coverage found this cycle. Treated as unverifiable and not reported as fact.

Plans

Consumer Plans (Kimi.com)

The consumer membership/pricing page at Kimi – Pricing is JavaScript-rendered and could not be machine-extracted via WebFetch this cycle (browsermcp fallback was unavailable in this environment). The consumer chat homepage at Moonshot AI still shows "Kimi AI with K2.6" as the default consumer model, indicating the consumer app has not yet moved to K2.7 Code. The plan structure below is carried forward from the May report; any June changes to consumer plan prices or quotas could not be verified live and should be re-checked manually.

Verification status: consumer plan prices and inclusions above are carried from the 2026-05 report. Source (May): Kimi – Pricing . The only live-confirmed consumer fact this cycle is that the app homepage still defaults to K2.6. Source: Moonshot AI

API Plans (Pay-as-you-go)

API rate-limit tiers are unchanged from May and were verified live. Source: Kimi – Limits

Minimum recharge $1. At $5 cumulative recharge users get a $5 voucher (vouchers do not count toward cumulative recharge). Enterprise custom limits via api-service@moonshot.ai. Source: Kimi – Limits

API Pricing

All figures per 1M tokens, verified live on the pricing pages. K2.7 Code and K2.7 Code HighSpeed are new in June.

Notes on K2.7 Code pricing: the HighSpeed variant is exactly 2x the standard K2.7 Code price on all three columns. K2.7 Code's cache-hit rate ($0.19) is slightly higher than K2.6's ($0.16), but its input/output match K2.6 exactly, so headline cost-per-MTok is unchanged versus the prior flagship. Source: Kimi – Chat K27 Code , Kimi – Chat K26 , Kimi – Chat V1

Discontinued (removed May 25, 2026): Kimi K2 (0905), K2 Turbo, K2 Thinking, K2 Thinking Turbo, K2 (0711).

Other API behavior:

• Web search tool: $0.005 per successful tool call (unchanged). Source: Kimi – Tools
• File upload/extract: temporarily free.
• K2.7 Code, K2.6, and K2.5 support text, image, and video input. Source: Kimi – Kimi K2 7 Code Quickstart
• K2.7 Code forces thinking (cannot disable), forces preserve_thinking, fixes temperature at 1.0 and top_p at 0.95, and tool_choice can only be "auto" or "none". Source: Kimi – Kimi K2 7 Code Quickstart
• Moonshot V1 models (8K/32K/128K) remain listed for legacy text generation. Source: Kimi – Chat

Model Performance / Benchmarks

Moonshot published a benchmark table on the K2.7 Code HuggingFace model card comparing K2.6, K2.7 Code, GPT-5.5 (Codex, xhigh), and Claude Opus 4.8 (Claude Code, xhigh). K2.7 Code and K2.6 were tested via Kimi Code CLI with thinking enabled. These are vendor-reported and include two in-house benchmarks, so they favor Moonshot.

Reading the table: K2.7 Code is a clear step up over K2.6 on every row (notably +11.1 on Kimi Code Bench v2 and +8.3 on MCPMark-Verified), but it still lands behind GPT-5.5 on every single benchmark and behind Opus 4.8 on all but MCPMark-Verified. Source: Hugging Face – Kimi K2.7 Code

Independent community benchmarkers caution that vendor and public coding benchmarks are "gamed" (answers are public and in training data) and that real-world gaps to Sonnet/Opus are larger than the numbers suggest. See Community Signals. No independent SWE-Bench Verified / DeepSWE numbers for K2.7 Code were published by Moonshot, so cross-supplier comparison on the standard harnesses is not yet possible.

Architecture: Mixture-of-Experts, 1T total parameters, 32B activated per token, 61 layers, 256K context, MLA attention, MoonViT vision encoder (400M params), native INT4 quantization. Source: Hugging Face – Kimi K2.7 Code

Latest News

Kimi K2.7 Code Launch (June 12, 2026)

Moonshot released Kimi K2.7 Code, a coding-focused model built on K2.6, open-weight under a Modified MIT license. It forces thinking and preserve_thinking, reports ~30% fewer thinking tokens than K2.6 and ~10% better agentic performance, and launched with a same-priced HighSpeed variant. The HackerNews launch thread reached 463 points and 240 comments. Sources: Hugging Face – Kimi K2.7 Code , News – Item

K2.7 Code Launch Top-Up Rebate (June 11 to July 2, 2026)

A time-limited promotion gives API users a 20-30% voucher on a single top-up of at least $100, capped at $4,000 and valid 90 days. This is effectively a discount for larger spenders during the launch window. Source: Kimi – Promotion

Moonshot Launches "AI-Native Credit Card" (June 30, 2026)

Moonshot announced what it calls the world's first "AI-native credit card," embedding Kimi membership benefits into a credit card, paired with prompt-engineering courses and AI events. It requires a Chinese mainland 11-digit mobile number, restricting it to the China market. Sources: Kimi – Aicard , News – Item

Cursor Composer 2.x Confirmed as Kimi K2.5 Fine-tune (background, re-surfaced June)

HN discussion in the K2.7 Code thread re-confirmed that Cursor's Composer 2.x models are RL fine-tunes of Kimi K2.5, licensed through Fireworks.ai, and that the attribution requirement of Moonshot's Modified MIT license was the source of the earlier "shame Cursor into disclosure" friction. Sources: News – Item , TechCrunch – Cursor Admits Its New Coding Model Was Built On Top Of Moonshot Ais Kimi

No SpaceX/xAI Acquisition of Cursor Found (claim, unverified)

A reported claim that "SpaceX acquired Cursor in June" could not be confirmed. It does not appear in the platform pages, HN search results for Kimi K2 or Moonshot AI, or the Cursor-related coverage found this cycle. The only adjacent signal is an unverified HN comment that "if you do have SuperGrok, they are giving you Composer 2.5 in Grok Build," which is about xAI's Grok Build reportedly routing to Composer 2.5, not an acquisition. Source: News – Item (treated as unverified, not reported as fact)

Community Signals

K2.7 Code launch thread: cost advantage vs. real-world quality gap

The K2.7 Code HN thread (463 points, 240 comments) split between enthusiasm for the price/performance ratio and skepticism that vendor benchmarks overstate how close Kimi is to frontier models. The clearest cost framing came from a developer comparing list prices.

• dnstje: "I was wondering how does Anthropic and likes keep competitive when Opus is ($5 / $25) 5x times more expensive compared to Kimi K2.6 ... while being only marginally better." (Note: the actual K2.6 price is $0.95/$4.00 per MTok.) News – Item
• DCKing: "There's no question Kimi models are very good for a lot of code tasks. They're the best quality open weight model. But to get similar overall outcomes as on Sonnet/Opus, on average you'll spend many more tokens and will have to do more managing of the model. You shouldn't look at price per token, you should look at how much you pay for the entire process." News – Item
• DCKing (on benchmarks): "on not-yet-gamed benchmarks like DeepSWE, Kimi K2.6 is beaten soundly by Claude Sonnet 4.6 ($3 / $15) and even slightly by GPT 5.4 Mini ($0.75 / $4.50)." News – Item

Kimi needs a harness tuned to its quirks

A recurring theme is that Kimi performs well inside its own CLI (and Kimi Code CLI specifically) but worse in generic harnesses, because the CLI contains model-specific workarounds.

• vidarh: "Kimi works great in their CLI, but their CLI has a number of workarounds for quirks of their models, including detecting when the model gets into a loop, and reverting to a checkpoint but letting the model compose a 'message' to its past self (search their CLI for 'BackToTheFuture'...) It doesn't work so well in a harness that doesn't take those quirks into account." News – Item
• regularfry: "the Kimi K2.5 and K2.6 models will comment out failing tests rather than fix a problem they just caused (mistaking them for 'pre-existing failures'), so you need to specifically make commented-out tests break the build. I've not personally had that problem with any of the Anthropic or OpenAI models." News – Item
• solarkraft: "For some reason I never had a good experience with Kimi (via OpenRouter) in OpenCode. It would only take a few turns for it to run off and mess something up. Terrible instruction following I'd say." News – Item

Early positive K2.7 Code hands-on report

At least one developer reported a successful large real-world task with the new model shortly after launch.

• pizlonator: "I just had Kimi K2.7-code rebase my Fil-C OpenSSL patch from 3.3.1 to 3.5.7 with quite bare bones instructions and it seems to have worked. 177KB patch, so it's not a small change." News – Item

Failed sources

• The consumer membership/pricing page (Kimi – Pricing ) returned only the page title via WebFetch (JavaScript-rendered content), and the browsermcp fallback was unavailable in this environment, so June consumer plan prices and inclusions could not be verified live. They are carried from the May report and flagged in Transparency Gaps.
• Reddit JSON endpoints were not polled this cycle; HN coverage above was used as the primary community source.

Enterprise Readiness

Transparency Gaps