Kimi K3 Opens 2.8T Weights as NVIDIA Backs SSI and the Open Secure AI Alliance Lands

The most important change on 2026-07-27 PT was that a Chinese lab published a 2.8T-parameter MoE flagship directly as weights plus a technical report for the first time. After 8 pm PT, Moonshot AI’s Kimi K3 rolled out weights, PerceptionBench, SGLang/Miles day-0 support, and AgentENV, a distributed agent environment built together with kvcache-ai. The same day, NVIDIA announced an equity investment in Ilya Sutskever’s SSI and a 10× compute expansion within 12 months, alongside the formal launch of the Open Secure AI Alliance and the choice by OpenAI and Google not to sign. The three events converge in the same direction across the supply chain: open-sourcing the frontier moat of closed models, locking in compute, and redrawing the camp boundary around the claim that “defense requires open weights.” The secondary storyline is Anthropic CEO Dario Amodei’s position paper stating that the company “has never advocated banning open weights,” which sits in direct opposition to both the NVIDIA-led alliance and the Kimi K3 release. This edition carries many topics, but the evidence boundaries deserve caution: the alliance roster and vendor readings come from relayed accounts on X, most K3 performance numbers come from first-hand posts, and the OpenAI/Google non-signature comes from community discussion rather than a formal statement by either party.

Theme 1: Kimi K3 opens a 2.8T MoE in full, with the technical report and base model released together

What happened. Starting at 8 pm PT, Kimi progressively released the K3 model weights, a technical report, high-performance attention kernels, an MoE communication library, and an accompanying large-scale agent training environment. K3 is an MoE model with 2.8T total parameters and 104B active parameters, natively supporting text, image, and video along with a 1M-token context, and aimed at long-horizon coding, deep research, and agent tool calling. The official claim for the new architecture is a “2.5× improvement in intelligence per unit of compute,” drawn from a training-recipe comparison in the technical report; the technical report is open-sourced at the same time.

Why it matters. At 2.8T, this is the largest open-weight model disclosed to date, and publishing the technical report on the same day means both the reproduction path and the training recipe are released. The SGLang team, writing on the lmsys.org blog, confirmed roughly 113 tok/s for single-GPU batch-1 decoding on SGLang, reaching roughly 423 tok/s with DSpark speculative decoding. Modal trained a custom DFlash speculator for it on day 0. On the same day, PerceptionBench decomposed visual perception into 10 atomic abilities and built 3,000 verified questions that force frontier models to answer using a single perceptual ability, without cheating via reasoning or external knowledge — a more restrained way of evaluating every vision model.

Evidence boundary. The performance figures (113 tok/s, 423 tok/s, the 2.5× intelligence gain, 42/42 on the IMO) come from the vendor and first-hand community posts and have not been aligned with independent third-party benchmarks. Several accounts described the license as “modeled on MIT, but model-as-a-service providers with more than $20 million in annual revenue must separately sign a commercial agreement” — this is a community paraphrase rather than the full text, and the implementation details are not public. Frontier Code Arena ranked K3 Max first among open-source models and first overall, but that leaderboard is a single platform’s relative ranking.

Sources:

Theme 2: NVIDIA bets heavily on SSI, scaling its compute 10× within 12 months

What happened. After two years of silence, Ilya Sutskever posted “Time to scale that SSI” at 7:23 pm PT, after which multiple accounts relayed that NVIDIA and SSI had reached a long-term strategic partnership: NVIDIA takes an equity stake in SSI and opens its next-generation Vera Rubin computing platform to the lab, raising SSI’s compute 10× over the next 12 months. Ilya’s only public comment was: “We have research that is worthy of scaling up.” SSI had previously relied mainly on Google TPUs and now formally joins the NVIDIA camp.

Why it matters. This is the first time SSI has publicly acknowledged that its closed-door research has passed internal validation and reached the point where it needs large-scale compute to be tested. NVIDIA’s decision to bet after seeing strictly confidential research amounts to acquiring the real requirements that next-generation models place on chips, interconnect, memory, and system architecture — which can in turn feed the roadmap back into the Vera Rubin design. SSI had already raised roughly $3 billion at a valuation of about $32 billion (community-relayed figures) without publishing any model or demo. This 10× compute step is the most substantial move SSI has made since its founding, and it is also a step in which NVIDIA continues to use investment to bind customers beyond OpenAI and Anthropic while pulling one of Google TPU’s most significant customers back into its own ecosystem.

The direct implication of this thread for the frontier research landscape is that the “small but elite frontier lab” path has been actively chosen by NVIDIA for the first time. SSI has run for two years with no public product and no model — NVIDIA placed its bet only after seeing the research, which means a compute giant is willing to pay real money for the judgment that “this direction is worth scaling.” In a retweet at 7:30 pm PT, Emad Mostaque added: “At the current $5 billion-a-year level of GPU delivery cycles, plus the time window for a single training run, you can back out the IlyAGI timeline.” That is community-level extrapolation, not the official cadence from NVIDIA or SSI.

Evidence boundary. Ilya’s one-line “worthy of scaling” does not amount to having found a new paradigm beyond pretraining; whether continual learning has been solved, and whether any new method has been proposed, remain undisclosed. The “10×” compute figure describes the NVIDIA–SSI partnership and is not deployment already completed. The $32 billion valuation and $3 billion cumulative funding come from Chinese-language relays on X, not from a formal announcement by either party. The $5 billion-a-year GPU delivery cycle Emad cited is his own estimate and has not been reconciled with NVIDIA’s announcements.

Sources:

Theme 3: The Open Secure AI Alliance launches, and OpenAI and Google choose not to sign

What happened. NVIDIA, Microsoft, Hugging Face, IBM, Databricks, Anthropic, Mistral, Cloudflare, LangChain, OpenClaw and dozens of other organizations jointly founded the Open Secure AI Alliance, arguing for building an auditable, customizable AI security defense system through open models, tools, and frameworks. In the original post, NVIDIA CEO Jensen Huang used the Hugging Face incident as an example: “Attackers have frontier AI, and defenders need a frontier AI ecosystem… open-weight frontier models helped contain the intrusion during the incident.” Retweeting during the PT evening, Andrew Ng stressed: “Stop believing the PR that closed models are safer — that is regulatory capture.” LangChain, Hugging Face, Anthropic, Cloudflare, OpenClaw, Databricks, and Mistral confirmed their participation in public retweets, with the NVIDIA blog serving as the main entry point.

Why it matters. On the same day the alliance launched, MaxForAI and several other accounts pointed out directly that OpenAI and Google had not signed the open-weights open letter that NVIDIA co-signed. This is a clear cross-section of how AI camps split in 2026: the side betting on open weights and positioning open models as “infrastructure required for defense” (NVIDIA, Hugging Face, Databricks, Anthropic, Mistral, LangChain, Cloudflare) stands in contrast to the non-signing side (OpenAI, Google). NVIDIA is using the alliance to institutionalize the “open equals defense” narrative, and in doing so makes it harder for the closed-source camp to argue back — because the argument now rests on empirical evidence from a specific incident (the Hugging Face incident Jensen Huang described).

Evidence boundary. “OpenAI and Google withdrew from signing the formal document” is currently a community-discussion-level description; neither party has formally stated that it will not sign. The alliance member list comes from relays by several official accounts plus the NVIDIA blog entry point, and Anthropic’s and Mistral’s subsequent participation is supported by explicit statements from Arthur Mensch and Peter Steinberger, but the complete roster and each member’s specific contributions remain fragmentary. No public technical post-mortem has been released on the details of the “Hugging Face incident” Jensen Huang mentioned, and “open-weight models contained the intrusion” is Huang’s qualitative characterization rather than a published incident report.

Sources:

Theme 4: Anthropic publishes its “open-weights position,” forking directly from the NVIDIA camp

What happened. At 3:26 pm PT, Anthropic CEO Dario Amodei published “Our position on open-weights models,” whose core message is that “Anthropic has never advocated, and does not advocate, banning open-weight models.” He simultaneously proposed three alternative measures: restricting the flow of advanced chips to China while cracking down on smuggling and export-control circumvention; cracking down on industrial-scale model distillation (naming Chinese companies that can replicate capabilities by making large volumes of calls to US models); and mandatory safety testing for every model that reaches a capability threshold, with exemptions for weaker small-company and academic models. Dario explicitly listed “mandatory safety testing for all sufficiently powerful models” as a public priority, covering both open and closed source.

Why it matters. This is an active act of separation by Anthropic against the dual backdrop of the NVIDIA camp’s co-signed open letter and the Kimi K3 open-sourcing on the same day. Its actual argument is “do not draw a blanket line based on open versus closed, but mandate testing based on a capability threshold, while continuing to restrict Chinese chips and distillation” — a framing whose beneficiaries happen to be closed-source vendors that charge per token. The community reading is that Anthropic is using “safety” language to limit competitors’ distillation capability and raise the compliance cost of frontier open source, which puts it in commercial conflict with the purely open-source camp. Debates over Steipete, and over whether Dario represents the open ecosystem, kept building through the PT evening; Peter Steinberger directly retweeted heated community phrasing such as “this burns Alexandr,” and David Sacks went further by contrasting Anthropic’s copyright position with this one.

Evidence boundary. Dario’s statement comes from Anthropic’s official site and relays by several accounts, but the specific implementation details behind claims such as “industrial-scale distillation comes mainly from China” and “mandatory safety testing for all models” are not public. MaxForAI pointed out that Anthropic’s official statement differs from two earlier versions circulating in the community — forged versions have already been rejected by the company. The real point of disagreement between Anthropic and the NVIDIA camp is not “whether to be open,” but “by what standard and under whose regulation.”

Sources:

Theme 5: Alibaba’s “Qwen Office” enters closed beta, with DingTalk’s Chen Yusen leading the integration of three agent lines

What happened. Alibaba has consolidated three agent product lines — QoderWork, Wukong, and MuleRun — into a unified “Qwen Office,” led by DingTalk’s new CEO Chen Yusen. The desktop client and the DingTalk version are already live, with a web version to follow for consistency across all three surfaces. Qwen Office differentiates itself by having the agent generate HTML directly, bundled with a domain, database, and hosting, which extends “building a web page” from agent output all the way to deployment.

Why it matters. This is the first systematic integration attempt in China’s WorkAgent race. Kazik’s analysis divides office products into three generations: local Office, cloud collaboration (DingTalk), and now “selling a kind of intelligence” — an agent that builds the tool best matched to your scenario for you. Alibaba simultaneously holds three assets: the model (Qwen), the infrastructure (Alibaba Cloud, Bailian), and organizational data (the messages, documents, workflows, and permissions accumulated in DingTalk), making it one of the few companies in China that has all four required elements in this race at once — intelligence, capability, availability, and a security floor.

Evidence boundary. “Qwen Office” is currently in a low-key closed beta, and public hands-on reports are limited mainly to Kazik’s first-hand trial, with no large-sample user feedback yet. The specific capabilities, pricing, and functional boundary with DingTalk after the three product lines are merged have not been officially disclosed. Whether the integration can hold up externally depends on the execution of Chen Yusen’s team; for now this is an organizational move rather than product validation.

Sources:

Theme 6: Agent evaluation and harness assessment enter a refined “mechanism-level” phase

What happened. Three kinds of new work related to agent evaluation appeared that day. First, EvoCode, recommended by Philipp Schmid: 26 tasks and 227 sequential rounds (5–15 rounds per task), with tests accumulating continuously in a single container, specifically to see whether an agent breaks existing behavior across multi-round evolving requirements — the common failure is not “cannot build it” but “changed something else.” Second, a Role Drift paper from Harvard/MIT: when end-to-end RL improves the accuracy of a composite system, modules maintain overall performance by taking shortcuts (the decomposer stuffs the answer into sub-questions, the reader falls back on parametric memory), and this invisible “role drift” costs 86% of the RL improvement. Third, NVIDIA research showing that AdamW has a scaling ceiling as batch size grows, and giving the location of that upper bound. These three point in the same direction as GEPA Optimize_anything (pluggable pipelines) retweeted by Berkeley and DAIR AI’s autoresearch with coding agents — all of them put questions beyond “which model to pick” (supervision, drift, training ceilings) on the table. All three lines of work point at the same thing: beyond the model layer, evaluating the “mechanism layer” carries more information than average success rate.

Why it matters. Agent evaluation is evolving from “success rate plus average score” toward “regression rate, mechanism decomposition, and cumulative consistency,” which means quality measurement for enterprise agents is shifting from surface metrics to interpretable, specific failure modes. Philipp Schmid’s summary is that “how well a skill was added shows up in how many new regressions it created”; EvoCode and Role Drift give that regression capability structure. A paper cited by omarsar0 notes that skills often degrade agent performance through three mechanisms — description bleed, grounding shift, and verification shift — which means the comparison between “adding a skill” and “removing a skill” does not only reflect the skill itself, but also the skill’s perturbation of how the agent reads input and checks its own work. This is a direct methodological input for every team working on enterprise agent evaluation.

Evidence boundary. All three pieces of work are previews or summaries: EvoCode came via Schmid’s recommendation, but the full paper link needs further tracing; the Role Drift paper was recommended by omarsar0 and the specific citation is still pending; the AdamW ceiling is a relay of NVIDIA research, and neither the paper title nor the conference venue appears in the captured sources. The specific effects of the “three-mechanism set” for skills come from a secondhand citation of a paper, and the original source needs further verification.

Sources:

High-value briefs

OpenAI Codex Voice expands globally to Edu/Business/Enterprise: OpenAI’s official account announced that GPT-Live in ChatGPT Voice is live, covering the education, business, and enterprise tiers. Source: https://x.com/OpenAI/status/2081794871795589485
Anthropic releases Claude Opus 5: AI Valley reported that Claude Opus 5 launched simultaneously in its app, Claude Code, and the API, positioned as approaching Fable 5 intelligence at roughly half the price, and achieving a perfect score across 42 questions at the 2026 International Mathematical Olympiad. Source: https://www.theaivalley.com/p/anthropic-s-opus-5-is-almost-fable-5
Ling-3.0-flash quietly lands on OpenRouter: 124B total parameters, 5.1B active, listed directly with no launch event, output quality close to some flagships, token cost roughly half of Claude’s, 256K context with tool calling fully enabled, and free for now. Source: https://x.com/AYi_AInotes/status/2081746074675388816
GitHub Copilot’s “Harness” workflow and the Copilot app go live: Harness packs prototyping, planning, implementation, and code review into a single AI tool; the Copilot app is upgraded into a multi-agent workspace, where /create-canvas allows previewing in the browser and clicking to make edits. Source: https://github.blog/ai-and-ml/github-copilot/the-harness-is-all-you-need-mostly
Cursor adds Kimi K3: inference on the US side via Fireworks, Together, and Baseten, with support for Zero Data Retention. Source: https://x.com/cursor_ai/status/2081848014444876166
Cloudflare open-sources pvcli: an OHTTP debugging CLI built on Cloudflare’s engineering experience handling millions of requests per second, able to execute a complete OHTTP request in a single command. Source: https://x.com/Cloudflare/status/2081727564356182171
Lilian Weng leaves Thinking Machines: after 20 months at Thinking Machines she announced her departure, citing seven months of continuous illness that she attributed to stress and workload overload, and said she hopes to continue AI work in an environment with clearer boundaries. Source: https://x.com/shao__meng/status/2081934534694879291
Google AI Overviews appearance rate rose from 15% to 43% in a year: AI Mode monthly visits grew from 126 million to 279 million, and user search length increased significantly. Source: https://techcrunch.com/2026/07/27/googles-ai-search-is-rapidly-becoming-the-default-new-data-shows
On the Chartography benchmark, Fable raised chart-reading scores from 29% to 73%: produced in collaboration between Surge AI and Anthropic, demonstrating how much frontier models have improved at professional chart reading. Source: https://x.com/echen/status/2081777626801229835
OpenAI and Anthropic accuse each other of “blocking rivals”: Steipete and BeffJezos retweeted accusations that Anthropic is “burning books” through lobbying, and Santiago (svpino) publicly announced he is moving his main repository off Claude Code, on the grounds that he does not want to keep paying Anthropic while watching it lobby the government to restrict competitors. Source: https://x.com/svpino/status/2081767307215351978
Hugging Face built a dedicated countdown page for Kimi K3: a sign that the K3 open-sourcing is treated as an important milestone by the international open-source community. Source: https://x.com/op7418/status/2081665633411092793

🕐 Selected hourly signals

PT time	Signal	Why it is worth remembering
00:00	AI Valley’s morning digest starts circulating Claude Opus 5	The flagship price war enters a “close to Fable 5, half the price” narrative
02:00	LangChain / NVIDIA / Anthropic simultaneously retweet the Open Secure AI Alliance	Coordination posture ahead of the alliance launch
03:00	SGLang/Miles publish Kimi K3 day-0 support, inference stack ready	K3 is not just model weights, but a complete inference stack open-sourced in step
07:00	Anthropic publishes its long-form “open-weights position”	Directly opposed to the NVIDIA open letter and the K3 release on the same day
08:00	Ilya Sutskever posts “Time to scale that SSI,” community relays the NVIDIA strategic partnership	His first outward move after two years of silence
09:00	Peter Steinberger publicly announces using Claude Code’s Robobun flow to have two agents fix each other’s bugs	An engineering sample of agent self-collaboration
11:00	Andrew Ng retweets Jensen Huang’s open letter and endorses it	Academia and industry converge on open weights
12:00	Anthropic internal sources reveal it is joining the alliance alongside Mistral and others	The open-weights camp expands beyond NVIDIA alone
13:00	Cursor and Fireworks / Together / Baseten add K3 in step	The speed at which K3 became commercially usable on day 0
14:00	Yangqing Jia and other accounts retweet the “2.5× training-recipe improvement” from the Kimi K3 technical report	Reproducibility on the training side becomes a focus of discussion
17:00	AYi, Max For AI, and other accounts converge on discussing the impact of “execution-layer cost” on agents	The agent era enters a “tiered model” narrative
19:00	Topview Film Studio launches its “director’s desk”	AI video moves from generation toward controllability

Editorial conclusion

Today’s main thread is a set of convergences pointing the same way: frontier models open-sourced, controllable compute locked in, and defense requiring open weights — three things that happened within the same time window and corroborate one another. Kimi K3 provides the sample showing that “even a 2.8T MoE can be released in full on day 0”; NVIDIA extends the logic of “using compute to bind customers” through its investment in SSI; and the Open Secure AI Alliance turns “open weights equals defense infrastructure” from an assertion into a camp boundary. Anthropic’s “position paper” is the day’s counter-pull — shifting the discussion from “open or not” to “capability thresholds and who regulates,” which is language aimed at regulators rather than developers. Secondary but not negligible are the domestic consolidation in the WorkAgent race and the fact that agent evaluation is starting to enter a refined “mechanism-level” phase.

Sources and method

That day covered three named sources with substantive content — aihot-morning.md, aivalley.md, and hubtoday.md — plus 19 hourly captures. Among the named sources, chrome-dev, claude-blog, openai-blog, google-research, cline-blog, and xiaohu-ai were empty or failed to capture that day and did not participate in theme selection. Community posts make up the majority; performance figures and the alliance roster come primarily from first-hand vendor and executive accounts. Community relays (such as the SSI valuation, Kimi’s active parameters, and OpenAI/Google not signing the alliance) have been marked in place, and single-source conclusions stay close to the original posts.