Moonshot AI が世界最大級オープンモデル「Kimi K3」発表

Z.ai GLM has been getting a bit too much love recently, so it’s time for Kimi K3 to fight back! It’s hard to put the scale of today’s open model release in perspective, so thankfully Moonshot AI did it for us:

Their vibe reel was entirely edited by Kimi K3 and worth a watch:

You can read SimonW and Arena for standard takes and rankings, none of which will be particularly unexpected given the large size of the model, but this pic best summarizes the K2.5 to K3 jump:

AI News for 7/15/2026-7/16/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

Moonshot AI launched Kimi K3 as a frontier-class open-weights model, with official claims that place it near top closed models and above prior open competitors.

Moonshot officially introduced Kimi K3 as “Open Frontier Intelligence” with 2.8T total parameters, 1M-token context, native multimodal input, Kimi Delta Attention (KDA), and Attention Residuals, and said the model is live on Kimi.com, Kimi Work, Kimi Code, and API, with open weights promised by July 27, 2026 @Kimi_Moonshot

Moonshot also highlighted product positioning around long-horizon agentic coding and self-evolving workflows, plus “vision in the loop” coding/game-building workflows that iterate between code and screenshots @Kimi_Moonshot

Before the formal announcement, multiple accounts circulated leaked or app-sourced details that K3 was 2.8T params, calling it the largest open-weight model ever if weights ship as promised @scaling01, @scaling01, @eliebakouch

The official Kimi blog went live later and was widely shared as the primary technical source @Jianlin_S, @scaling01, @Yulun_Du

Moonshot’s own phrasing acknowledged a limitation: despite being highly competitive overall, K3 still has a “noticeable gap in user experience” versus Claude Fable 5 and GPT-5.6 Sol @scaling01

Arena announced that Kimi K3 entered Agent Arena, plus Text, Vision, Document, and Frontend Code Arena, with community evaluations to follow @arena

Arena then reported a major early result: Kimi K3 became #1 in Frontend Code Arena with 1679 points, surpassing Claude Fable 5 and jumping from #18 (K2.6) to #1, ranking #1 in 6 of 7 frontend domains and #2 in Gaming @arena

Arena later added that K3 has a 76% pairwise win rate in Frontend Code Arena, versus 63% for Fable 5 and 58% for GPT-5.6 Sol @arena

In Text Arena, K3 landed at #9 with 1486 points, a jump from #38, with top-10 placements in creative writing, coding, and instruction following, and #1 in several occupation slices @arena

Artificial Analysis published an independent evaluation placing K3 at 57 on the AA Intelligence Index, calling it comparable to Opus 4.8 and GPT-5.5, but still behind Fable 5 and GPT-5.6 Sol overall @ArtificialAnlys

AA also reported K3 at 1668 Elo on GDPval v2, 53% / #1 on AutomationBench-AA, and 1547 Elo on AA-Briefcase, with cost per task of $0.94, about 21% fewer output tokens than K2.6 across the full Intelligence Index run @ArtificialAnlys

The launch immediately triggered strong reaction from engineers and model-watchers who framed K3 as an open-model milestone comparable to earlier DeepSeek moments @kimmonismus, @nrehiew_, @eliebakouch

Technical details

Architecture and systems details

Official specs: 2.8T total parameters, 1M context, native multimodal input (text + images), text output, open weights by July 27 @Kimi_Moonshot, @ArtificialAnlys

K3 uses Kimi Delta Attention (KDA), which Moonshot says enables up to 6.3x faster decoding in million-token contexts @Kimi_Moonshot

It also uses Attention Residuals (AttnRes), claimed to deliver ~25% higher training efficiency at <2% additional cost @Kimi_Moonshot

Community readers of the blog highlighted additional architecture details: LatentMoE / Stable LatentMoE, 16 activated experts out of 896, implying an activation ratio under 2% @nrehiew_, @eliebakouch

More community-extracted details from the blog/report discussion: per-head Muon, QB load balancing / quantile load balancing, and a new activation function called SiTU (Sigmoid Tanh Unit) @eliebakouch

One engineer noted the architecture as notable for combining KDA + LatentMoE + AttnRes while scaling more than 2x over prior Kimi models @teortaxesTex

KDA had a long incubation cycle: design reportedly started in Jan 2025 and took ~1.5 years to reach frontier scale @zxytim

Inference and serving

K3 pricing was reported as $3 / 1M input tokens and $15 / 1M output tokens, with cached input discounted 90% to $0.30 / 1M @scaling01, @ArtificialAnlys

Several posters compared that pricing to Sonnet 5, with some noting Sonnet was temporarily cheaper until end of August, after which prices align more closely @kimmonismus

A blended estimate at 80% input / 20% output came out to $5.40 / 1M tokens, vs $9 for Opus 4.8 and $10 for GPT-5.5 @jaminball

Artificial Analysis estimated $0.94 average cost per Intelligence Index task, versus $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8 @ArtificialAnlys

Early live serving observations: ~28 tok/s via Moonshot API on OpenRouter @scaling01, and another observer saw 26 tok/s, calling it slower than Opus and speculating that speculative decoding wasn’t yet enabled @nrehiew_, @nrehiew_

Moonshot’s blog reportedly recommends deployment on supernode configurations with 64+ accelerators for best inference efficiency @teortaxesTex

vLLM said Moonshot contributed a KDA prefix caching implementation directly to vLLM, with support available day 0 for official release @vllm_project

Moonshot’s KDA contribution was cited as important because KDA breaks assumptions behind conventional prefix caching, so upstream runtime changes were required @vllm_project

Benchmarks and evals

Moonshot’s official benchmarking message, as summarized by others, positioned K3 behind only Claude Fable 5 and GPT-5.6 Sol among tested models, and ahead of Claude Opus 4.8 @scaling01, @Yuchenj_UW

One cited number: 1687 on GDPval-AA v2, above Opus 4.8 and behind GPT-5.6 Sol at 1747.8 in that comparison @scaling01

Artificial Analysis’ independent numbers:

AA Intelligence Index: 57

GDPval v2 Elo: 1668

AutomationBench-AA: 53%, #1

AA-Briefcase Elo: 1547

AA-Omniscience: +18, with accuracy 46% vs 33% on K2.6, but hallucination rate worsening to 51% from 39% @ArtificialAnlys, @ArtificialAnlys

AA also reported 132M output tokens consumed for K3 across the Intelligence Index, versus 166M for K2.6, i.e. 21% reduction while gaining 13 index points @ArtificialAnlys

Arena’s frontend result was especially prominent because it is a pairwise human-preference arena, not just a static benchmark, and K3’s #1 frontend rank became one of the main launch headlines @arena

Community posts also highlighted strong results on kernel optimization tasks, with some saying K3 was matching or beating Fable in certain kernel/codegen settings @nrehiew_, @scaling01

One benchmark caveat came from ProgramBench author Ofir Press, who said Kimi used a metric they do not recommend: averaging implementation percentage rather than counting fully working programs, which can overstate usefulness @OfirPress, @OfirPress

Facts vs opinions

Facts / directly sourced claims

Kimi K3 is officially announced by Moonshot @Kimi_Moonshot

Officially disclosed specs include 2.8T params, 1M context, native multimodal input, KDA, AttnRes, open weights by July 27 @Kimi_Moonshot

Artificial Analysis independently scored K3 at 57 Intelligence Index, with detailed task, cost, token, and benchmark data @ArtificialAnlys

Arena independently ranked K3 #1 in Frontend Code Arena and later reported its 76% pairwise win rate @arena, @arena

vLLM confirmed Moonshot contributed runtime support for KDA prefix caching @vllm_project

Opinions / interpretations

“DeepSeek moment,” “beginning of the US-China AI race,” and “everything changed” are editorial interpretations from observers, not established facts @kimmonismus, @scaling01, @kimmonismus

Claims that K3 “beats GPT-5.6 Sol on 11 of 14 benchmarks” and “Fable on 6 of 14” are aggregated community summaries and should be treated as contingent on the benchmark set and exact methodology @scaling01

Assertions that this implies Dario/Anthropic margin pressure, a geopolitical turning point, or near-term superintelligence are speculative commentary @teortaxesTex, @Jason

Several “distillation” insinuations were explicitly framed as jokes or conjecture rather than evidence @yacinelearning, @dejavucoder

Different opinions

Strongly supportive

Many engineers called K3 a genuine frontier open model, especially because it appears to be better than Opus 4.8 while being priced near Sonnet and planned for open-weight release @kimmonismus, @cline, @nrehiew_

Supporters emphasized that this is no longer “good for open source,” but simply competitive with top public closed models @tokenbender, @TheAhmadOsman

Some framed the release as evidence that open models are now within weeks or a couple months of the frontier @nrehiew_

Others argued this materially raises the odds that future AGI-level systems are open @MaorShlomo

Supportive but technically cautious

Artificial Analysis gave a more restrained view: K3 is comparable to Opus 4.8 and GPT-5.5, but still behind Fable 5 and GPT-5.6 Sol on overall intelligence @ArtificialAnlys

Simon Willison described K3 as significant, but also pointed readers toward nuanced notes and benchmark caveats rather than simple leaderboard hype @simonw

Ethan Mollick’s hands-on impression: very good open-weights model, but not Sol Max or Fable @emollick

One user said K3’s intelligence is strong, but it is slow, sometimes over-checks, and still trails Claude on taste/aesthetics @nrehiew_

Critical / skeptical

Bindu Reddy warned that K3’s benchmark story might be overstated unless validated on hidden / uncontaminated evals like LiveBench, and argued that if the model “thinks forever,” real cost could be less favorable @bindureddy

ProgramBench maintainers objected to Moonshot’s metric choice, saying it can inflate partial-credit performance relative to fully working programs @OfirPress

Artificial Analysis also flagged a real weakness: hallucination rate regressed on AA-Omniscience despite accuracy gains @ArtificialAnlys

Multiple users noted that K3 currently appears to think a lot, preserve long reasoning history, and may require more careful harness support than simpler chat-first APIs @scaling01, @Xianbao_QIAN

Some skepticism focused on economics and deployability: 2.8T open weights is impressive, but practical self-hosting may still be limited to well-funded teams @mbusigin

Political / strategic interpretations

A broad cluster of tweets framed K3 as proof that Chinese labs are no longer far behind and that the US lead is shrinking @tszzl, @kimmonismus, @scaling01

Others counterweighted that K3 still appears to lag the very best Western models in usability / productization, even if raw capability is close @RyanGreenblatt, @scaling01

Some argued that open Chinese models function as economic pressure on US labs by compressing margins and commoditizing capability @francoisfleuret

Others viewed the inevitable next step as more competition on harnesses, products, and deployment systems, not just raw model weights @AravSrinivas, @theo

Context

Why this matters technically

K3 is notable not just for raw size but for scaling a non-standard attention stack into a frontier-class model: KDA + AttnRes + sparse MoE drew repeated attention from technically literate observers @scaling01, @eliebakouch

The launch is also a systems story: long-context serving, prefix caching, KDA runtime support, and deployment on large accelerator supernodes all matter if the weights are to be practically usable @vllm_project, @teortaxesTex

The emphasis on kernel optimization, chip design, agentic coding, and environment simulation suggests Moonshot is optimizing for AI-improving-AI workflows, not just chatbot benchmarks @18jeffreyma, @yong_zhengxin

Why this matters economically

The strongest repeated theme: frontier-ish performance at materially lower price than top closed models, though not at bargain-basement open-model prices @kimmonismus, @cline, @jaminball

Artificial Analysis’ task-cost framing is especially relevant for practitioners: if K3 is near GPT-5.6 Sol cost-per-task and below Opus 4.8, the real question becomes where it slots into agent stacks, coding platforms, and self-hosted infra @ArtificialAnlys

Some noted the paradox that “open weights” does not automatically mean “cheap to run”: a 2.8T model with 64+ accelerator deployment guidance is frontier infrastructure territory @teortaxesTex, @mbusigin

Why this matters geopolitically

Many reactions explicitly tied K3 to export controls, US-China competition, and the narrowing gap between Chinese open labs and US closed labs @scaling01, @tszzl, @kimmonismus

Several commentators argued that K3 weakens the common narrative that Chinese models trail by 6–8 months, because it appears to outperform a closed US model from late May only weeks later @kimmonismus

Others stressed that “capability parity” is not the same as full-stack parity: product reliability, inference scale, deployment margins, and proprietary post-training may still favor US incumbents @RyanGreenblatt

Early hands-on signals

Users reported K3 building impressive web experiences, games, and shader/code artifacts, reinforcing the Frontend Arena result @johnlindquist, @ChrissGPT, @intheworldofai

One user said K3 generated a CS:GO × Portal clone in 3 shots using ~600k tokens, costing $3.24 by API pricing, compared with claimed higher costs on Fable and GPT-5.6 Sol @ChrissGPT

Another reported K3 continuously working for hours over near-1M context to build a web DOS emulator with low human intervention @bigeagle_xd

At the same time, several users noted it can be verbose, slow, and heavily reliant on thinking-history preservation, implying that serving/harness defaults will matter a lot @nrehiew_, @Xianbao_QIAN, @bigeagle_xd

Open-source/open-weights debate

The surrounding discourse included the usual complaint that “open weight” is not “fully open,” but several commenters pushed back that this distinction is often impractical at frontier scale and that inspectable, fine-tunable weights still matter @Dan_Jeffries1, @ClementDelangue

Yulun Du said the delay before weight release was to ensure a smooth rollout with inference partners, signaling that ecosystem readiness mattered as much as the checkpoint itself @Yulun_Du

vLLM maintainers and others treated Moonshot’s upstream contributions as evidence that the launch is not just “marketing open,” but also includes meaningful OSS infra work @vllm_project, @woosuk_k

Benchmarks, contamination, and what to watch next

Several people cautioned that current public benchmark ecosystems saturate quickly, and that hidden evals or stack-level evals will be more informative @bindureddy, @gdb, @WolfBenchAI

Observers specifically asked for follow-up on METR time horizons, cyber ranges, FrontierMath T4, ARC-AGI-2/3, CritPt, token usage, and broader long-horizon agent evals @scaling01

The most credible near-term follow-up points are:

whether the weights ship on time

what third-party serving stacks achieve for throughput/cost

how K3 performs on hidden evals and real production agent tasks

whether Moonshot closes the UX/post-training gap they themselves acknowledged @Kimi_Moonshot, @scaling01, @ArtificialAnlys

Open Models, Inference Stacks, and Retrieval Infrastructure

vLLM and serving ecosystem support landed quickly: vLLM said Moonshot contributed a KDA prefix-caching implementation directly to vLLM, enabling day-0 support once weights drop. This matters because KDA breaks some conventional prefix-caching assumptions. The post underscores that long-context architectural innovation increasingly requires coordinated systems work, not just model release.

NVIDIA shipped a notable open retrieval release: NVIDIA launched Nemotron 3 Embed 8B, claiming #1 overall on RTEB, and partners quickly made it deployable, including Baseten and Turbopuffer. A more detailed community summary by @kimmonismus reports 78.46 NDCG@10 on RTEB and 75.45 on MMTEB Retrieval, with NVIDIA arguing stronger retrieval reduces downstream agent token usage. The release also includes 1B BF16 and 1B NVFP4 variants, with the NVFP4 version reportedly offering up to 2× BF16 throughput on Blackwell while retaining >99% retrieval quality.

LiteParse added a gRPC interface for backend document pipelines: LlamaIndex introduced liteparse-grpc, exposing PDF/Office/image parsing, rendering, and OCR-complexity estimation over gRPC with protobuf definitions and generated clients. This is a practical infra improvement for polyglot microservice stacks where REST isn’t ideal.

Managed vector/search infra also expanded: Weaviate announced Managed Weaviate on DigitalOcean in public preview, running the unmodified open-source engine (v1.37.1 at launch) with HA, autoscaling, backups, forks, and control-plane observability.

Agents, Harnesses, and System Design Becoming the Real Product Layer

Harnesses were a recurring theme across builders: Harrison Chase’s conversation with Factory AI’s Eno Reyes was repeatedly shared as a case for why “the harness matters more than the model” (Harrison, LangChain). Chase later argued teams should “own the harness,” “own the context and memory layer,” and “own model optionality” rather than rent intelligence from a single provider (thread).

There’s growing interest in open standards for memory and knowledge representation: Harrison Chase promoted OKF (Open Knowledge Format) as an “open standard for memory,” while Brace Sproul detailed OpenWiki’s adoption and the benefits for search, retrieval, and codebase memory.

Agent self-improvement and scheduled multi-agent workflows are becoming mainstream topics: @omarsar0 highlighted a survey on self-improving agentic systems, and elsewhere described using an “LLM Council” with recurring scheduled research updates (thread). On the product side, Google AI Studio added a free tier for Managed Agents, plus max_total_tokens for pausing/resuming long runs and native cron triggers.

Perplexity’s infra direction was also notable: NVIDIA AI Infra highlighted Perplexity’s new SPACE secure sandbox platform, with early tests on NVIDIA Vera CPU showing up to 1.9× faster sandbox starts—a reminder that sandbox startup latency is now part of agent throughput engineering.

OpenAI and Anthropic: Safety, Productization, and Developer Workflow Updates

OpenAI acknowledged a dangerous Codex/GPT-5.6 failure mode around file deletion: Thomas Sottiaux said OpenAI investigated rare reports where GPT-5.6 unexpectedly deleted files, most commonly when full access mode was enabled without sandboxing or auto review, and when the model attempted to override $HOME for temp directories but mistakenly deleted $HOME itself. OpenAI says it is updating developer messaging, nudging users toward safer permission modes, and adding harness safeguards, with a detailed postmortem forthcoming.

OpenAI continued to ship workflow features around Codex and PR review: OpenAI Devs added PR Chat and inline code editing in Codex for reviewing and editing pull requests in context. OpenAI also announced Office Hours around GPT-5.6, ChatGPT, and Codex (source).

Anthropic upgraded Claude Code review depth: ClaudeDevs introduced effort levels for /code-review, from low cost/low effort to ultra, where a fleet of reviewer agents reproduces findings independently. Anthropic says low effort beats other code-review