Moonshot AI

AI Coding Agents Report: July 2026 · Updated 31 July 2026 · Version history

Executive Summary

What it is: Moonshot AI's Kimi is a consumer AI platform with a dedicated coding agent (Kimi Code) available via CLI, IDE extensions, and web. On July 17, 2026 the platform launched Kimi K3, a 2.8T-parameter flagship model with a 1M-token context window, at $3.00/$15.00 per MTok (input/output) with a $0.30/MTok cache-hit rate. K3 is released open-weight under the "Kimi K3 License" (more restrictive than K2.7 Code's Modified MIT). The consumer chat app now defaults to K3, up from K2.6 last month. The dedicated coding model K2.7 Code ($0.95/$4.00) remains available, as do K2.6 ($0.95/$4.00) and K2.5 ($0.60/$3.00). K3 is roughly 3x the per-MTok price of K2.7 Code but still about 40% cheaper than Anthropic's Opus 4.8 ($5/$25) and publishes competitive scores on independent benchmarks (DeepSWE 67.5 vs GPT-5.5's 67.0). Source: https://platform.kimi.ai , https://www.kimi.com/blog/kimi-k3

What to watch out for: K3 always thinks with no disable option (like K2.7 Code), defaults to max reasoning effort, and emits reasoning tokens on every call with no separate pricing for thinking tokens. On July 19, Moonshot suspended new consumer subscriptions due to K3 demand overwhelming capacity, and all plan purchase buttons on the pricing page now say "Join Waitlist" instead of "Subscribe." The pricing page also announces "New Membership Plans Coming Soon" that will separate Kimi and Kimi Code benefits, meaning current plan structures may change. Agent credit multipliers changed from the old 1x/5x/15x/30x scheme to a new "more"/2x/5x/10x scheme, still with no defined base unit. The May billing complaints (double-charging, 429s) have no announced resolution, though Moonshot did add an invoice system and daily spending limits to the platform. Enterprise readiness remains near zero (no SSO, SLA, or IP indemnity).

Bottom line: K3 is a genuine frontier-class open-weight model that closes much of the benchmark gap to GPT-5.5 and Opus 4.8 on independent coding benchmarks (DeepSWE, Terminal-Bench 2.1), and it brings a 1M context window that K2.7 Code lacks. But at 3x the per-MTok cost of K2.7 Code, it is no longer the bargain that drew developers to Kimi in the first place. For cost-sensitive coding, K2.7 Code remains the better API value. For teams that need the absolute best open-weight coding model and can tolerate the license terms, K3 is a strong choice. For any team needing governance, billing reliability, or guaranteed availability, Kimi remains a poor fit given the subscription suspension and absent enterprise features.

Key Terms

  • Kimi K3 - Moonshot's 2.8T-parameter flagship model, launched July 17, 2026, with a 1M-token context window. Mixture-of-Experts activating 16 of 896 experts (104B activated parameters), built on Kimi Delta Attention (KDA) and Attention Residuals. Released open-weight under the Kimi K3 License. Source: Hugging Face – Kimi K3
  • Kimi Delta Attention (KDA) - K3's hybrid linear attention mechanism, combined with Attention Residuals (AttnRes), designed to improve information flow across long sequences and deep models. K3 uses 69 KDA layers and 24 Gated MLA layers across its 93 total layers. Source: Kimi – Kimi K3
  • Reasoning effort - K3's configurable thinking depth via the top-level reasoning_effort request field, supporting low, high, and max (default max). K3 always thinks and cannot have thinking disabled, but this field controls how much reasoning it does per call. Source: Kimi – Kimi K3 Quickstart
  • Kimi K3 License - K3's open-weight license. Unlike K2.7 Code's Modified MIT, the K3 License reportedly exempts internal use (use that does not make outputs available to third parties) from attribution requirements, but commercial use above approximately $20M/year revenue or 100M users may require a license fee. Source: Hugging Face – License , News – Item
  • MXFP4 quantization - K3's native quantization format, using MXFP4 weights with MXFP8 activations via quantization-aware training from the SFT stage onward. The full model is approximately 1.5TB in this format. Source: Hugging Face – Kimi K3
  • 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
  • MCP (Model Context Protocol) - an open standard for connecting models to external tools and data. K3 supports tool_choice constraints and dynamically loaded tools as new API capabilities. Source: Kimi – Kimi K3 Quickstart
  • Mooncake - Moonshot's disaggregated inference architecture. The official K3 API reportedly achieves a cache hit rate above 90% in coding workloads via Mooncake. Source: Kimi – Kimi K3

Latest Changes

Changes since the 2026-06 report.

June Watch Items Verified

  • Kimi K3 launch (CONFIRMED): K3 launched July 17, 2026, exactly as the June report flagged as a watch item. The consumer app homepage changed from "Kimi AI with K2.6" to "Kimi AI with K3." Full model weights were released on HuggingFace by July 27. The API platform banner reads "Kimi K3 has launched!" Sources: Moonshot AI , Kimi , Hugging Face – Kimi K3
  • Launch top-up rebate (CONFIRMED, ended as scheduled): The K2.7 Code Launch Top-Up Rebate that ran June 11 to July 2, 2026 ended on schedule. The promotion page now redirects to a general account/billing FAQ. No extension was announced and no K3-specific rebate has replaced it. Source: Kimi – Promotion
  • Billing complaint resolution (PARTIAL): Moonshot added an invoice system (Invoice Management at /console/invoice, issuing invoices via Beijing Moonshot AI Technology Co., Ltd. at 6% tax rate) and budget controls (daily spending limits in project settings, account balance alerts). However, the original May complaints about double-charging after cancellation, 429 rate-limit errors, and no self-service cancel link have no announced resolution. Source: Kimi – Promotion
  • Enterprise readiness improvements (FAILED): No SSO, SCIM, audit logs, IP indemnity, SLA, or admin controls have been added. The platform page and blog mention "Kimi Enterprise" and "Enterprise Solutions" with marketing language about "SLA backed reliability," but no concrete enterprise features are documented. Source: Kimi

New Items This Month

  • New model: Kimi K3 launched July 17, 2026 as a 2.8T-parameter flagship with a 1M-token context window at $3.00/$15.00 per MTok (input/output) with $0.30/MTok cache-hit. Released open-weight on HuggingFace under the Kimi K3 License. Source: Hugging Face – Kimi K3 , Kimi – Kimi K3
  • Behavior: K3 always thinks (cannot disable), defaults to max reasoning effort, supports configurable reasoning effort via reasoning_effort (low/high/max). Temperature is fixed at 1.0, top_p is configurable (0.95 for single-step, 1.0 for agentic). Source: Kimi – Kimi K3 Quickstart
  • Subscription suspension: Moonshot suspended new consumer subscriptions on July 19, 2026 due to K3 demand exceeding capacity. All plan purchase buttons on the pricing page now say "Join Waitlist." Source: Twitter – Status , News – Item
  • Plan restructuring announced: The consumer pricing page displays "New Membership Plans Coming Soon" and states "Kimi and Kimi Code benefits will be separated. Existing subscribers are unaffected. You can still buy the current plan before launch." Source: Kimi – Pricing
  • Agent credit multiplier change: Consumer plans now list agent credits as "more"/2x/5x/10x (Moderato/Allegretto/Allegro/Vivace) instead of the old 1x/5x/15x/30x. The base unit remains undisclosed. Source: Kimi – Pricing
  • Annual billing: Consumer plans now offer annual billing at a discount: Moderato $15/mo (annual) vs $19/mo (monthly), Allegretto $31/mo vs $39/mo, Allegro $79/mo vs $99/mo, Vivace $159/mo vs $199/mo. Annual billing saves up to $480/year. Source: Kimi – Pricing
  • Deprecation: Moonshot V1 models (8K/32K/128K) now show "full platform sunset expected on August 31" on the pricing page. Source: Kimi – Chat
  • New API features (K3): K3 adds tool_choice constraints (auto/none/required) and dynamically loaded tools (inject tool definitions via a system message without content mid-conversation), structured output via JSON Schema with strict: true, and Partial Mode (continue from a text prefix). Source: Kimi – Kimi K3 Quickstart
  • Open infrastructure: Moonshot released K3 alongside open-source infrastructure: MoonEP (expert parallelism), AgentEnv, and FlashKDA. Source: News – Item
  • No pricing changes to K2.7 Code / K2.6 / K2.5: Per-MTok prices and rate-limit tiers are unchanged from June. Source: Kimi – Chat K27 Code , Kimi – Limits
  • K3 weights: Full model weights (2.8T params, MXFP4 format, approximately 1.5TB) were released on HuggingFace by July 27, 2026. Downloads last month: 387,822 (as of July 31). Source: Hugging Face – Kimi K3

Plans

Consumer Plans (Kimi.com)

The consumer membership page was extracted live this cycle via browser automation. Key change: all plan purchase buttons now say "Join Waitlist" (not "Subscribe"), consistent with the July 19 subscription suspension. The page also announces "New Membership Plans Coming Soon" with Kimi and Kimi Code benefits being separated.

Plan Price (monthly) Price (annual) Agent Credits Concurrent Tasks Key Inclusions
Adagio (Free) $0 N/A Baseline 1 2 scheduled tasks, 2 projects, 500MB storage
Moderato $19/mo $15/mo ($180/yr) "more" 2 K3 up to 1M tokens, Kimi Code, 10 scheduled tasks, 20 projects, 20GB, Swarm (2 subagents)
Allegretto $39/mo $31/mo ($372/yr) 2x 2 K3 up to 1M tokens, Kimi Code, Goal mode, 15 scheduled tasks, Swarm (4 subagents)
Allegro $99/mo $79/mo ($948/yr) 5x 4 K3 up to 1M tokens, Kimi Code, Goal mode, 20 scheduled tasks, 100 projects, 50GB, Swarm (8 subagents)
Vivace $199/mo $159/mo ($1,908/yr) 10x 4 K3 up to 1M tokens, Kimi Code, Goal mode, 25 scheduled tasks, 100 projects, 50GB, Swarm (8 subagents)

Changes from June report:

  • Annual billing option is new (saves up to $480/year).
  • Agent credit multipliers changed from 1x/5x/15x/30x to "more"/2x/5x/10x.
  • "K3 Extra Long Chat Capacity (Up to 1M Tokens)" is now listed as a paid-tier feature.
  • New features visible: Goal mode, Customizable Dashboard, Dream Memory, Self-evolving Skill, Deployable Kimi Claw.
  • All plans now include 4x speed priority (previously listed as 4x on paid tiers).
  • Swarm subagents quantified: 2/4/8/8 by tier (was listed as "50/120/240 Swarm uses/mo" in prior reports).

    • Verification status: Verified live via browser automation on July 31, 2026. Source: Kimi – Pricing

    Subscription status: New subscriptions are suspended. All purchase buttons say "Join Waitlist." Existing subscribers are unaffected. Source: Twitter – Status

    API Plans (Pay-as-you-go)

    API rate-limit tiers are unchanged from June. Source: Kimi – Limits

    Tier Cumulative Recharge Concurrency RPM TPM TPD
    Tier 0 $1 1 3 500,000 1,500,000
    Tier 1 $10 50 200 2,000,000 Unlimited
    Tier 2 $20 100 500 3,000,000 Unlimited
    Tier 3 $100 200 5,000 3,000,000 Unlimited
    Tier 4 $1,000 400 5,000 4,000,000 Unlimited
    Tier 5 $3,000 1,000 10,000 5,000,000 Unlimited

    K3 requires a minimum $1 top-up to unlock. 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 , Kimi – Kimi K3 Quickstart

    API Pricing

    All figures per 1M tokens, verified live on the pricing pages. K3 is new in July.

    Model Input (cache miss) ($/MTok) Cache Hit ($/MTok) Output ($/MTok) Context Window
    Kimi K3 $3.00 $0.30 $15.00 1,048,576 (1M)
    Kimi K2.7 Code $0.95 $0.19 $4.00 262,144 (256K)
    Kimi K2.7 Code HighSpeed $1.90 $0.38 $8.00 262,144 (256K)
    Kimi K2.6 $0.95 $0.16 $4.00 262,144 (256K)
    Kimi K2.5 $0.60 $0.10 $3.00 262,144 (256K)
    Moonshot V1 8K $0.20 N/A $2.00 8K
    Moonshot V1 32K $1.00 N/A $3.00 32K
    Moonshot V1 128K $2.00 N/A $5.00 128K

    Notes on K3 pricing: K3 uses flat pay-as-you-go pricing with no tiering by context length. The official API reportedly achieves a cache hit rate above 90% in coding workloads via Mooncake's disaggregated inference architecture, meaning effective input cost in coding is closer to the $0.30 cache-hit rate than the $3.00 cache-miss rate. Source: Kimi – Chat K3 , Kimi – Kimi K3

    K3 is approximately 3.2x more expensive on input and 3.75x more expensive on output than K2.7 Code ($0.95/$4.00). However, K3 offers a 1M context window (4x K2.7 Code's 256K) and significantly higher benchmark scores on independent coding benchmarks.

    Deprecation notice: Moonshot V1 models (8K/32K/128K) show "full platform sunset expected on August 31" on the pricing page. Source: Kimi – Chat

    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. However, K3's web search is currently being updated and is "not recommended for use in the near term." Source: Kimi – Chat K3
    • File upload/extract: temporarily free.
    • K3 supports text and image input (native multimodal). Video input is an experimental feature only supported via the official API. Source: Kimi – Kimi K3 Quickstart
    • K3 always thinks (cannot disable), defaults to reasoning_effort=max. max_completion_tokens defaults to 131,072 and can be set up to 1,048,576. Temperature is fixed at 1.0, top_p is configurable. Source: Kimi – Kimi K3 Quickstart
    • K3 supports tool_choice (auto/none/required), dynamically loaded tools, structured output (JSON Schema), and Partial Mode as new API capabilities. Source: Kimi – Kimi K3 Quickstart
    • Context caching is automatic. A new request can hit the prefix cache only when the previous request's prompt tokens exceed 256. Source: Kimi – Kimi K3 Quickstart

    Model Performance / Benchmarks

    K3's benchmark table is a major improvement over K2.7 Code's. Where K2.7 Code only published vendor benchmarks (Kimi Code Bench v2, Kimi Claw 24/7 Bench), K3 publishes scores on independent benchmarks including DeepSWE, Terminal-Bench 2.1, FrontierSWE, SWE-Marathon, and ProgramBench. All K3 results use reasoning effort set to "max." K3 was tested via the Kimi Code harness.

    Coding Benchmarks

    Benchmark Kimi K3 (max) Claude Fable 5 (max) GPT-5.6 Sol (max) Claude Opus 4.8 (max) GPT-5.5 (xhigh) GLM-5.2 (max)
    DeepSWE 67.5 70.0 73.0 59.0 67.0 46.2
    Terminal-Bench 2.1 88.3 88.0 88.8 84.6 83.4 82.7
    FrontierSWE 81.2 86.6 71.3 66.7 64.9 67.3
    SWE-Marathon 42.0 35.0 39.0 40.0 14.0 13.0
    ProgramBench 77.8 76.8 77.6 71.9 70.8 63.7
    MLS-Bench-Lite 48.3 49.9 46.2 42.8 35.5 40.4
    SciCode 58.7 60.2 56.1 53.5 56.1 50.5
    PostTrainBench 36.6 41.4 34.6 34.1 28.4 34.3
    Kimi Code Bench 2.0 (in-house) 72.9 76.9 64.8 71.7 69.0 64.2

    Reading the table: K3 is now competitive on independent coding benchmarks. On DeepSWE (67.5), K3 matches GPT-5.5 (67.0) and significantly outperforms Opus 4.8 (59.0) and GLM-5.2 (46.2), though it trails Claude Fable 5 (70.0) and GPT-5.6 Sol (73.0). On Terminal-Bench 2.1, K3 (88.3) is essentially tied with Fable 5 (88.0) and GPT-5.6 Sol (88.8). On SWE-Marathon, K3 (42.0) leads all competitors including Fable 5 (35.0) and GPT-5.6 Sol (39.0). On FrontierSWE, K3 (81.2) is second only to Fable 5 (86.6).

    Note: Kimi K3 attains 67.3 on DeepSWE with the mini-SWE-agent harness (vs 67.5 with Kimi Code). K3 also attains 73.7 on Kimi Code Bench 2.0 with the Claude Code harness (vs 72.9 with Kimi Code). Source: Hugging Face – Kimi K3

    K2.7 Code Benchmarks (for comparison, from June)

    K2.7 Code was benchmarked only on vendor benchmarks. For reference, K2.7 Code scored 62.0 on Kimi Code Bench v2 (vs K3's 72.9 on Kimi Code Bench 2.0) and had no published independent benchmark scores.

    K3 Architecture

    Mixture-of-Experts with 2.8T total parameters, 104B activated per token (16 of 896 experts selected, plus 2 shared experts), 93 layers (69 KDA + 24 Gated MLA), 1,048,576-token context window. Built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) with Stable LatentMoE framework. MoonViT-V2 vision encoder (401M params). Native MXFP4 weights with MXFP8 activations via quantization-aware training. SiTU-GLU activation function. Approximate 2.5x scaling efficiency improvement over K2. Source: Hugging Face – Kimi K3 , Kimi – Kimi K3

    K3 Documented Limitations

    Moonshot explicitly documents three limitations:

  • Sensitivity to thinking history: K3 requires the complete assistant message (including reasoning_content and tool_calls) to be passed back in multi-turn conversations. If the harness fails to do this, or if a session is switched from another model to K3 mid-conversation, quality may become highly unstable.
  • Excessive proactiveness: K3's emphasis on long-horizon tasks means it may make unexpected decisions when encountering minor issues or ambiguous intent. Moonshot recommends explicit behavioral constraints in system prompts or AGENTS.md.
  • Frontier gap: Moonshot acknowledges a "noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol."

    • Source: Kimi – Kimi K3

    Latest News

    Kimi K3 Launch (July 17, 2026)

    Moonshot released Kimi K3, a 2.8T-parameter flagship model with a 1M-token context window, open-weight under the Kimi K3 License. It is built on Kimi Delta Attention (KDA) and Attention Residuals with Stable LatentMoE (16 of 896 experts activated). K3 is described as the "world's first open 3T-class model." Case studies include GPU kernel optimization, building a Triton-like compiler (MiniTriton) from scratch, chip design in a 48-hour autonomous run, and game development with vision-in-the-loop. The HackerNews launch thread reached 1,376 points and 544 comments, the highest-engagement Kimi thread ever tracked. Sources: Hugging Face – Kimi K3 , Kimi – Kimi K3 , News – Item

    Kimi K3 Technical Report (July 27, 2026)

    The K3 technical report was released as a PDF on GitHub alongside the model weights. The HackerNews thread reached 391 points and 181 comments. Sources: GitHub – K3 Tech Report.Pdf , News – Item

    Moonshot Suspends New Subscriptions Due to K3 Demand (July 19, 2026)

    Moonshot announced it was temporarily pausing new consumer subscriptions because "demand has pushed close to the limits of our current capacity." Existing subscribers are unaffected. The HN thread reached 284 points and 113 comments. The consumer pricing page now shows "Join Waitlist" on all plan purchase buttons. Sources: Twitter – Status , News – Item

    Consumer Plan Restructuring Announced

    The consumer pricing page now displays "New Membership Plans Coming Soon" and states "Kimi and Kimi Code benefits will be separated." This suggests Kimi Code may become a separately purchasable product. Agent credit multipliers changed from 1x/5x/15x/30x to "more"/2x/5x/10x. Annual billing was introduced with savings up to $480/year. Source: Kimi – Pricing

    Moonshot V1 Sunset (August 31, 2026)

    The API pricing page now shows Moonshot V1 models (8K/32K/128K) with "full platform sunset expected on August 31." These legacy text generation models have been listed since before the K2 series. Source: Kimi – Chat

    Open Infrastructure Released with K3

    Moonshot released open-source infrastructure alongside K3: MoonEP (expert parallelism training), AgentEnv (agent environment), and FlashKDA (fast KDA implementation). Source: News – Item , GitHub – Moonep

    Rest of World Coverage on US/China AI Dynamics

    An article titled "The U.S. wants to contain China's AI. Silicon Valley keeps using it" covers the tension between US export controls and Silicon Valley's adoption of Kimi models. The HN post reached 12 points. (The article itself returned HTTP 403 and could not be fetched for content verification.) Sources: Restofworld – China Siliconvalley Ai Moonshot Kimi , News – Item

    Cursor Remains Listed as Platform Partner

    Cursor is listed as a partner on the Kimi API platform homepage, alongside Vercel, OpenRouter, Windsurf, Kilo Code, and others. Cursor's Composer models remain confirmed RL fine-tunes of Kimi K2.5. Source: Kimi

    Community Signals

    K3 launch thread: frontier-level open weights, massive engagement

    The K3 HN thread (1,376 points, 544 comments) is the highest-engagement Kimi thread ever tracked in this report series, reflecting the significance of a 2.8T open-weight model. Discussion focused on self-hosting costs, data sovereignty, and whether open weights change the competitive landscape.

    • NitpickLawyer: "Since it's going to be mxfp4 native, it'll take ~1.5TB of VRAM to host this, which is juuust at the limit of 8xb200s (but realistically you'll need 16x for context / throughput optimisation). Won't be cheap to host, but at least we should get some range of $/MTok for a 3T model." News – Item
    • FuriouslyAdrift: "We're running Kimi 2.8 on a $107k server and getting around 50k tokens per second or better on most things. We've already saved money compared to last years token cost on Claude/Gemini." (Confirmed as 4x MI300A GPUs.) News – Item
    • GodelNumbering: "If you are a large enough company that spends million+ on inference a month, it makes sense to buy a GB300 rack ($6M)... you get less than 60 cents (USD) per million output token, for a frontier model with plenty of capacity to share, all your data never leaving premises and well over an order of magnitude cheaper!" News – Item
    • m_ke: "Also open sourced a bunch of infra to go with it. Anyone who claims open source and open weights models are 'decel' needs to get their head checked." (Links to MoonEP, AgentEnv, FlashKDA.) News – Item

    Subscription suspension: praise for transparency, concern about capacity

    The subscription suspension thread (284 points, 113 comments) generated discussion about capacity reliability and how Moonshot handled the overload compared to competitors.

    • Alifatisk (quoting Moonshot's announcement): "Over the past 48 hours, demand has pushed close to the limits of our current capacity. To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and prioritizing compute for current members. Existing subscribed users are not affected." News – Item
    • vblanco: "Ive been pleasantly surprised with the quality of this model. Its really good at code review and PR review. But right now due to them being so overloaded + this being a big model, its SO slow. It takes forever to do a relatively simple code review." News – Item
    • charcircuit: "The rate limits on the $20 plan go so fast that it's easy to use my weekly budget in only a couple days. I can't imagine also trying to use K3 on the plan. Back in January with K2.5 it felt like I could use it a lot more." News – Item
    • cube00: "A breath of fresh air to pause new subscriptions rather than the Google approach of quietly nerfing the limits and hoping you don't notice you're getting less value for your monthly/annual subscription." News – Item
    • kburman: "They could've simply cut the usage limits, removed K3 from the plan, added a weekly cap, and called it 'fair usage.' Anthropic has been running that playbook for a while now, and they're a trillion-dollar company." News – Item
    • KronisLV: "Currently using their 100 USD subscription, considering going up to their 200 USD subscription, because the model is approximately as capable as Opus (maybe sometimes better, sometimes worse) but less annoying to use in practice." News – Item
    • himata4113: "I wonder if anthropic and openai will remain relevant simply due to the fact that they're the only ones that are able to handle this much demand for the forseeable future?" News – Item

    Data sovereignty and self-hosting as K3's key enterprise angle

    A significant subthread in the K3 discussion focused on whether open-weight models like K3 enable data-sovereign deployment scenarios that closed APIs cannot, particularly for European and Canadian organizations concerned about US government data access under the CLOUD Act.

    • walrus01: "There are a number of use cases where sending the contents of your context and prompts (and the resulting output) to a 3rd party service is off the table as an option, and people will compromise speed for data sovereignty." News – Item
    • Taunt4: "Yes, its the US cloud act risk EU companies run up against on hyperscalers like MS/AWS... Even for EU companies running open weights on EU stacks LLM inference on the GPU must process plaintext and I can't find any EU provider with NVIDIA H100/H200/Blackwell CC mode plus SEV-SNP or TDX." News – Item
    • Sanzig: "I am in Canada. I cannot legally host defence workloads in the United States, even if I wanted to (I most certainly don't)." News – Item

    K3 download tool reflecting geopolitical anxiety

    A developer built a 68KB binary tool (kimi-fetch) to download the full 1.5TB K3 from HuggingFace with restart capability, motivated by geopolitical concerns about model availability.

    • marsninja: "I created this because I couldn't find a focused tool just for this type of thing, and I want to get this done with then next 24 hrs, because there is no telling what the government might do." News – Item

    Failed sources

    • Rest of World article (Restofworld – China Siliconvalley Ai Moonshot Kimi) returned HTTP 403 and could not be fetched for content. The HN submission (12 points) was used as the source reference instead.
    • The consumer membership/pricing page was JavaScript-rendered but was successfully extracted via browsermcp automation this cycle.

    Enterprise Readiness

    Feature Available? Details
    SSO (SAML) No Not mentioned anywhere on the platform or in the K3 documentation. Kimi remains a consumer and API product.
    SSO (OIDC) No Not mentioned.
    SCIM No Not mentioned.
    Audit logs No Not mentioned.
    IP indemnity No Not mentioned. The K3 License carries no indemnity.
    Data residency Partial K3 is open-weight and self-hostable. Berget AI offers European-hosted K2.6 for GDPR compliance; no K3 EU hosting announced yet. Moonshot's FAQ mentions "data compliance and privacy protections" for business customers but provides no specifics. Sources: Berget – Berget Code Launch En , Kimi
    HIPAA No Not mentioned.
    Air-gapped / on-prem Partial K3 is open-weight and runnable via vLLM, SGLang, or TokenSpeed. Requires approximately 1.5TB VRAM in MXFP4 format (8x B200 minimum, 16x recommended for throughput). No official on-prem enterprise product. Source: Hugging Face – Kimi K3
    SLA No No published SLA. The docs state Moonshot "may take temporary measures to adjust the rate limits" when load peaks. The subscription suspension on July 19 demonstrates capacity constraints. Source: Kimi – Limits
    Admin controls (RBAC) No No admin controls documented. API tiers are single-user. The platform mentions "Organization Management" and project-level daily spending limits, but no RBAC. Enterprise custom limits exist only via email (api-service@moonshot.ai). Source: Kimi – Promotion

    Transparency Gaps

    Gap Details Severity
    K3 thinking cost not separately priced K3 always thinks (defaults to max reasoning effort) and emits reasoning tokens on every call, but there is no separate line item for thinking tokens versus output tokens. Reasoning tokens appear to be billed at the output rate ($15.00/MTok). With max reasoning effort, a significant fraction of output tokens may be reasoning tokens. High
    Subscription suspension with no ETA New consumer subscriptions have been suspended since July 19 with no announced restoration date. Prospective buyers cannot sign up. The "Join Waitlist" buttons provide no timeline. High
    Agent credit multiplier base unit still undisclosed Consumer plans changed from 1x/5x/15x/30x to "more"/2x/5x/10x, but the base unit (what "1x" or "2x" means in tokens or tasks) is still not defined. The multiplier scheme itself changed without explanation. High
    Upcoming plan restructuring undefined "New Membership Plans Coming Soon" with "Kimi and Kimi Code benefits will be separated" is announced but with no details on pricing, timing, or what the new structure will look like. High
    May billing complaints still unaddressed No July announcement resolves the May complaints (double-charging after cancellation, 429 errors, no visible cancel link). An invoice system and budget controls were added, but the original issues are not acknowledged. High
    Cache hit rate claim is vendor-reported Moonshot claims "a cache hit rate above 90% in coding workloads" for the official K3 API, but this is a vendor figure with no methodology or third-party verification. If accurate, effective input cost is near $0.55/MTok (90% at $0.30 + 10% at $3.00), but buyers cannot verify this before spending. Medium
    K3 web search broken K3's web search tool is documented as "currently being updated" and "not recommended for use in the near term," with no timeline for fix. This is a regression from K2.6/K2.7 Code where web search was functional at $0.005/call. Medium
    HighSpeed capacity The K2.7 Code HighSpeed variant is still described as "limited" resource with "slightly fluctuate" performance, with no availability or throughput guarantee. Medium
    No batch API No batch API or discounted async processing tier documented. Low
    No refund policy No refund policy published. Account deletion is "not currently supported" per the FAQ. Low
    Cursor/xAI ownership status The "SpaceX acquired Cursor" claim from June remains unverified. Cursor is listed as a Kimi API platform partner. Cursor's Composer models remain confirmed as K2.5 RL fine-tunes. Low

    Previous month: Moonshot AI 2026-june

    ← Back to AI Coding Agents Report: July 2026