开源权重争议与Kimi K3发布

Everyone say hi to Richard MacManus, our new Head of Editorial!The current debate about Open Weights is the kind that creates a lot of grandstanding on a topic, while they wait for a very small set of players that will actually decide how things go (in either direction); this is not very conducive for those of us trying to focus on high signal to noise.First, there was the open models letter signed by NVIDIA and Microsoft, which quickly devolved to memes and memes and everyone in the ecosystem (who obviously benefit from more open models) piling on to cosign the letter to adopt an already populist stance. Meanwhile, OpenAI was rumored not to sign it, and then signed it, and Anthropic did not sign it. All very predictable, and all somewhat exhausting.Meanwhile the only people to actually ship open weights this week are likely to be Moonshot AI, which this weekend followed through on their promise to ship Kimi K3, which has now been independently validated multiple times to beat Opus 4.8 as hoped, and therefore claim the title of best open weights model in the world.If you don’t make law, make chips, or make models, we recommend reading the Kimi K3 tech report rather than 50 tweets of low-perplexity invective by the commentariat to the proletariat. AI News for 7/25/2026-7/27/2026. 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 RecapMoonshot’s Kimi K3 Open-Weights Release and the New 3T-Class Open FrontierKimi K3 is the day’s dominant release: Moonshot released Kimi K3 weights, report, and supporting infra as an open-weights package: a 2.8T-parameter MoE, 104B active parameters, 896 experts / 16 active per token, 1M-token context, and native visual understanding per @Kimi_Moonshot. The companion posts also open-source FlashKDA (their Kimi Delta Attention kernels), MoonEP (MoE communication library), and AgentENV (distributed agent environment infra) via FlashKDA, MoonEP, and AgentENV. This is more than a model drop; it is a fairly complete recipe for large-scale agentic post-training and serving.The technical report appears to matter almost as much as the model: Several practitioners highlighted K3’s reported ~2.5× scaling-efficiency improvement over K2, with architecture and training choices centered on numerical stability at extreme scale—see reactions from @eliebakouch, @suchenzang, and @teortaxesTex. Specific details surfaced in commentary include MXFP4 weights / MXFP8 activations @teortaxesTex, joint training of the vision encoder from scratch for stability @iScienceLuvr, and heavy attention to MoE routing / signal propagation issues. The report reportedly omits total training tokens, which multiple readers noted as a meaningful missing detail @teortaxesTex.Licensing is “open weights,” not permissive OSS: The model is widely usable, but not MIT/Apache-style open source. Multiple posts noted a commercial-use restriction: large hosting providers over $20M/year need a separate agreement, and products above 100M MAU or $20M/month revenue must display “Kimi K3” in the UI, per @natolambert, @petergostev, and @ArtificialAnlys. This is a useful signal for where frontier “open” may be settling: source-available / open-weight with business carve-outs rather than OSI-style licensing.Distribution was immediate and broad: K3 was available day 0 via vLLM @vllm_project, Baseten @baseten, Modal @modal, Fireworks @Kimi_Moonshot, Nebius @Kimi_Moonshot, Together @Kimi_Moonshot, DigitalOcean @Kimi_Moonshot, Cursor @cursor_ai, Cognition/Devin @cognition, Ollama Cloud @ollama, and Dell Enterprise Hub @jeffboudier. That breadth underscores that open-weight frontier launches are now supply-chain events, not just research announcements.Open AI Security, Open Weights Politics, and Anthropic’s PositionNVIDIA formally launched the Open Secure AI Alliance: Jensen Huang framed the core thesis starkly: attackers already have strong AI, so defenders need an ecosystem spanning open and closed frontier models, plus shared tooling and research. The flagship statement came from @JensenHuang, with NVIDIA’s formal announcement at @nvidia. The most technically interesting detail in the messaging was the claim that during the OpenAI/Hugging Face incident, a frontier open-weight model helped contain the intrusion, while a closed model blocked essential forensics—echoed by @AndrewYNg and @ZixuanLi_.The alliance quickly accumulated credible infra and tooling members: Confirmed participants posting publicly included Hugging Face @huggingface, LangChain @LangChain, Nous Research @NousResearch, and support from voices across the open ecosystem such as @UnslothAI and @Yuchenj_UW. The argument is not “open is automatically safer,” but that defensive capability and auditability require open access to models, harnesses, and traces.Anthropic finally clarified its open-weights stance: After sustained criticism for not signing NVIDIA’s open-weights letter, Anthropic published a position statement saying it has “never advocated for a ban on open-weights models” and instead supports: chip controls on China, anti-industrial-scale distillation measures, and mandatory safety testing for sufficiently capable models, open or closed, per @AnthropicAI. Reactions split between “reasonable clarification” @signulll, “good, but still trying to slow frontier diffusion” @jachiam0, and more hostile readings from open-weight advocates like @Teknium.Policy pressure is intensifying around pre-release review: Separate reporting suggested the US government may seek up to 30 days of pre-release access to frontier systems for evaluation by agencies such as NSA and CAISI, with open-vs-closed treatment still unresolved, via @kimmonismus and @leomschwartz. Together with Anthropic’s statement and OpenAI’s Washington briefings, the direction is clear: frontier model release is becoming a governance interface, not just a product launch.Benchmarks, Evals, and Agent ReliabilityK3’s early evals are strong, especially for agents/coding: On Agent Arena, Kimi K3 Max reportedly ranks #1 among open-weight models with +9.75% net improvement, leading across multiple signals including confirmed success and steerability @arena. It also took #1 overall in Frontend Code Arena among all models in a later post @arena. Cognition said K3 is the first open-source model they tested that “approaches frontier-level performance” on FrontierCode 1.1, scoring 58.2% with 63.6% pass rate @cognition.Claude Opus 5 also posted strong leaderboard numbers, but practitioner feedback was mixed: Arena reported Opus 5 Max at #1 in Frontend Code Arena and Text Arena with factuality on @arena, while WeirdML numbers from @htihle put Opus 5 high/max at 91.6% / 91.8%, roughly tied with Fable 5 max. But several devs reported frustrating real-world behavior—overcomplication, breakage, poor stopping behavior—from @abacaj, @davis7, @Teknium, and @theo. As usual, public eval gains and harness-specific production utility are diverging.New eval work focused on sequential degradation and hidden regressions: @_philschmid highlighted EvoCode, an eval built around 26 tasks / 227 sequential rounds in a persistent container, measuring whether agents can follow evolving requirements without breaking earlier behavior. In parallel, @omarsar0 summarized a paper showing the “regression tax” from agent skills: across nearly 6,000 paired runs, skills generated gains but also broke many tasks previously solved without them. That is a practical warning against naïvely stuffing more procedural skills into context.Multi-module RL systems are showing “role drift”: Another useful paper summary from @omarsar0 described how end-to-end RL can improve pipeline accuracy while causing modules to quietly abandon intended responsibilities—e.g. a decomposer embedding the answer rather than structuring the problem. This feels increa

原文超出正文长度上限，此处截断——上游还有内容，完整版见上方「原文 ↗」。