AI Model Daily - 2026-07-28 | QName AI

30-Second Summary

Kimi K3 is today's clearest new-model story: Moonshot/Kimi released open weights for K3, a 2.8T MoE model with native vision and a 1M-token context window. Official materials say it is available through Kimi.com, Kimi Work, Kimi Code, the Kimi API, and Hugging Face weights; the license is open-weight but commercially constrained.
Artificial Analysis now marks Kimi K3 as the leading open-weight model with an Intelligence Index score of 57. Arena gives a matching signal: Kimi K3 Max is the top open-weight model in Agent Arena and #2 overall in Frontend Code Arena, behind only Claude Opus 5 Max.
Arena had two other important changes today: Claude Opus 5 Max topped Frontend Code Arena, while GPT-5.6 Terra/Luna xHigh entered Agent Arena at roughly #15 / #17. Both are Arena appearance / leaderboard updates, not new model launches today.
OpenAI expanded GPT-Live into ChatGPT Voice for Business, Edu, and Enterprise. OpenAI describes it as a new-generation full-duplex voice model that can listen and speak at the same time and delegate complex work to GPT-5.5.
Microsoft released MAI-Cyber-1-Flash inside MDASH as a lower-cost cyber agent model/system. Cohere launched North Automations. NVIDIA's long-term SSI partnership, the Open Secure AI Alliance, and Anthropic's open-weight-model statement are today's main policy and safety threads.

Must Watch

1. Kimi K3 weights released: 2.8T MoE, native vision, 1M context

Event type: Official release / X first signal / open-weight release.
First / official / available time: @Kimi_Moonshot released Kimi K3 weights and the technical report on 2026-07-27; the official blog and API docs say full weights were released by July 27.
Capability change: Kimi K3 is a 2.8T-parameter MoE model with roughly 104B activated parameters, 896 experts, and 16 selected experts per token. It supports text/image input, native visual understanding, a 1,048,576-token context length, KDA, and Attention Residuals. Kimi says scaling efficiency improved by roughly 2.5x versus K2.
Availability: Kimi.com, Kimi Work, Kimi Code, Kimi API, and Hugging Face weights. Third-party access through OpenRouter, Together AI, Fireworks, Baseten, DigitalOcean, Nebius, and others has started or is in day-one rollout.
Impact: This is Moonshot's largest open-weight milestone since Kimi K2. K3 combines long context, vision, agentic/coding capability, and open weights in one model, creating direct pressure on Qwen, GLM, DeepSeek, Llama, Gemma, and Mistral's open-weight lines. The license should not be treated as unconstrained open source: Model-as-a-Service businesses with more than $20M in revenue over the past 12 months need a separate agreement.
Sources: Kimi K3 official blog, Kimi K3 API quickstart, Hugging Face model card, Kimi K3 license, Kimi X.

2. Kimi K3 leaderboard signal: AA leads open weights, Arena coding/agent strength

Event type: Arena appearance / third-party benchmark verification / X first signal.
First / official / available time: Artificial Analysis reported the Kimi K3 weights release and updated its model page on 2026-07-27; @arena posted Kimi K3 Max performance across Arena tracks the same day.
Capability change: Artificial Analysis gives Kimi K3 an Intelligence Index score of 57 and marks it as the current leading open-weight model. The model page records text+image input, 1M context, 2.8T total parameters, 104B active parameters, and pricing around $3/$15 per 1M input/output tokens. On Arena, Kimi K3 Max is the top open-weight model in Agent Arena; the public Frontend Code Arena page shows kimi-k3-max at 1682 Elo, #2 overall behind claude-opus-5-max at 1725.
Availability: Artificial Analysis model page, Arena Agent / Code / Text leaderboards, Kimi, and third-party inference providers.
Impact: Kimi K3 is not relying only on launch messaging today; it has both Artificial Analysis and Arena signals. It should move into the "must test" open-weight set, especially for coding agents, visual understanding, long-context retrieval, tool use, and mixed Chinese/English tasks.
Sources: Artificial Analysis Kimi K3, Artificial Analysis X, Arena Agent leaderboard, Arena Code WebDev, Arena X - Kimi K3.

3. Arena snapshot: Claude Opus 5 Max, GPT-5.6 Terra/Luna, and Kimi K3 compete directly

Event type: Arena appearance / leaderboard change.
First / official / available time: Arena public leaderboards and @arena updated on 2026-07-27.
Capability change: Frontend Code Arena shows claude-opus-5-max at 1725 Elo in #1, kimi-k3-max at 1682 in #2, and claude-opus-5-high at 1670 in #3. In Text Arena, Claude Fable 5 remains #1, Claude Opus 5 Max is #5, and Kimi K3 Max is #11. The public Agent Arena page shows #1 Claude Fable 5 High at 12.60%, #2 GPT-5.6 Sol xHigh at 10.11%, and #3 Kimi K3 Max at 9.75%; the newly visible GPT-5.6 Terra xHigh / Luna xHigh sit around #15 / #17.
Availability: Arena public leaderboards; Terra/Luna can also be tried through OpenRouter, which announced a 50% discount on the same day.
Impact: Arena's picture is clearer than yesterday. Opus 5 Max has a confirmed lead on frontend coding. Kimi K3 Max has reached #2 overall and is ahead of many closed models. GPT-5.6 Terra/Luna are new agentic ranking signals, but they are not yet top-tier leaderboard leaders and need more sample-size stabilization.
Sources: Arena Code WebDev, Arena Text leaderboard, Arena Agent leaderboard, Arena X - Terra/Luna, OpenRouter Terra, OpenRouter Luna.

4. OpenAI GPT-Live expands into ChatGPT Voice for commercial and education plans

Event type: Official release / availability update.
First / official / available time: @OpenAI announced on 2026-07-27 that GPT-Live in ChatGPT Voice is globally available for Edu, Business, and Enterprise; OpenAI's official introduction page and Business release notes provide details.
Capability change: GPT-Live is described as a next-generation voice model that supports full-duplex conversation, meaning it can listen and speak at the same time. For deeper reasoning, web, or complex work, it delegates to a background frontier model; OpenAI's introduction says the launch configuration uses GPT-5.5 in the background. GPT-Live-1 and mini are rolling out globally, with API access planned later.
Availability: ChatGPT Voice for Edu / Business / Enterprise; API availability remains pending.
Impact: This is not a new text/code foundation model, but it can materially improve voice-agent usability. The key things to watch are latency, interruption behavior, long-conversation stability, enterprise controls, and voice coordination with Work/Codex agents.
Sources: OpenAI - Introducing GPT-Live, ChatGPT Business release notes, OpenAI X.

5. Microsoft MAI-Cyber-1-Flash: lower-cost cyber agent model inside MDASH

Event type: Official release / media coverage / security model.
First / official / available time: Microsoft AI released MAI-Cyber-1-Flash inside MDASH on 2026-07-27; Axios covered it the same day.
Capability change: Microsoft says MAI-Cyber-1-Flash is a compact code-heavy security model evolved from the MAI-Thinking-1 lineage. It handles roughly 90% of cyber tasks in MDASH, while GPT-5.4 is reserved for the hardest 10%. The combined system reaches around 96% on CyberGym while cutting cost by about 50%. Axios reports that access will be controlled through Azure AI Foundry and paired with Project Perception's Red/Blue/Green agents.
Availability: MDASH; Azure AI Foundry availability requires vetting and GPU provisioning, and the exact public-preview timing still needs confirmation.
Impact: This is the vertical model/agent release most likely to be overshadowed by Kimi K3 today. It shows large labs placing specialized smaller models inside multi-agent security systems, using cost and latency advantages for most tasks and routing the hardest work to larger models.
Sources: Microsoft AI - MAI-Cyber-1-Flash, Axios coverage.

6. Cohere North Automations: enterprise workflow orchestration enters North

Event type: Official release / agent workflow capability update.
First / official / available time: @cohere and Cohere's official blog released North Automations on 2026-07-27 and said it is available to all North customers.
Capability change: North Automations lets users turn complex enterprise workflows into controlled steps using natural language. It supports scheduled runs, loops, branching, plan mode, versioning, testing before production, guardrails, analytics, roles, per-step model choice, first-party integrations, custom MCP, and the SDK.
Availability: Cohere North customers.
Impact: This is not a foundation-model launch, but it shows enterprise agent products moving from one-off Q&A or single-agent execution toward governed workflow automation. It should be compared against OpenAI Work, Microsoft Copilot/MDASH, and Anthropic Claude for Enterprise.
Sources: Cohere blog - North Automations, Cohere X.

7. SSI and NVIDIA: compute expansion is a future-model signal, not a model launch

Event type: Official release / X first signal / media coverage / future model signal.
First / official / available time: NVIDIA Newsroom published the long-term strategic partnership between SSI and NVIDIA on 2026-07-27; @xiaohu tracked the WSJ/NVIDIA signal the same day.
Capability change: NVIDIA says it will invest in Safe Superintelligence Inc. and use NVIDIA Vera Rubin systems to expand SSI's compute by an order of magnitude. SSI/X messaging says compute can increase 10x over the next 12 months. NVIDIA also says it received rare access to SSI's closely guarded research.
Availability: No model, API, or weights are available.
Impact: This is not an available model today, but it matters for frontier-model competition in late 2026 and 2027. If SSI can scale training compute quickly, it may move from a highly secretive research organization into a frontier lab that the market can more seriously benchmark.
Sources: NVIDIA Newsroom - SSI partnership, Ilya Sutskever X, xiaohu X, xiaohu WSJ source.

8. Anthropic and Open Secure AI Alliance: open-weight policy keeps heating up

Event type: Official statement / media coverage / policy signal pending verification.
First / official / available time: @AnthropicAI said on 2026-07-27 that there had been heavy speculation about its open-weights position and linked to "our position on open-weights models." NVIDIA announced the Open Secure AI Alliance the same day.
Capability change: Anthropic's X post is an official entry point, but the linked Anthropic page did not load reliably during this run, so it should remain "official X plus page pending verification." NVIDIA's Open Secure AI Alliance includes NVIDIA, Hugging Face, Microsoft, IBM, LangChain, Cloudflare, Palantir, ServiceNow, SpaceXAI, and others. Its goal is to share open AI security tools, models, harnesses, and operational experience. NVIDIA's blog also says SpaceXAI plans to open source Grok line weights.
Availability: Anthropic X / Anthropic web page pending recovery; NVIDIA Open Secure AI Alliance and NOOA GitHub.
Impact: Open weights are now not only a release strategy but also a security, regulatory, geopolitical, and enterprise procurement issue. Today, Anthropic's signal should be labeled as "official X + page pending verification," while NVIDIA's alliance should be labeled as an official release, not a model launch.
Sources: Anthropic X, NVIDIA blog - Open Secure AI Alliance, The Verge coverage, Techmeme X.

Arena Watch

Artificial Analysis / Intelligence Index: Kimi K3 is scored at 57, and Artificial Analysis X calls it the leading open-weight model. The AA model page records text+image input, 1M context, 2.8T total parameters, 104B active parameters, $3/$15 per 1M input/output tokens, sample output speed around 33.3 tok/s, and TTFT around 144.34s. Source: Artificial Analysis Kimi K3.
Arena / Agent: The public page shows #1 Claude Fable 5 (High) at 12.60%, #2 GPT-5.6 Sol xHigh at 10.11%, and #3 Kimi K3 Max at 9.75%. Kimi K3 Max is also the top open-weight model and has a strong Confirmed Success signal. GPT-5.6 Terra xHigh / Luna xHigh newly appear around #15 / #17. Sources: Arena Agent leaderboard, Arena X - Terra/Luna.
Arena / Frontend Code: claude-opus-5-max is #1 at 1725 Elo, kimi-k3-max is #2 at 1682, and claude-opus-5-high is #3 at 1670. This leaderboard supports both Opus 5's closed-model coding lead and Kimi K3's open-weight strength. Sources: Arena Code WebDev, Arena X - Opus 5.
Arena / Text: Claude Fable 5 remains #1, Claude Opus 5 Max is #5, Claude Opus 5 High is #7, Kimi K3 Max is #11, and GPT-5.6 Sol xHigh is #13. Source: Arena Text leaderboard.
OpenRouter availability signal: GPT-5.6 Terra / Luna are marked on OpenRouter as Jul 9 releases with 1M context, and OpenRouter announced a 50% discount on 2026-07-27. Kimi K3 is live on OpenRouter with Balanced/Nitro/Exacto routing. Sources: OpenRouter X - Terra/Luna, OpenRouter Kimi K3.
Conclusion: Kimi K3 should be upgraded to a priority open-weight model today. Claude Opus 5 Max continues to lead on frontend code. GPT-5.6 Terra/Luna are new Arena/access signals, but not yet top leaderboard models.

X Watch

@Kimi_Moonshot: Released Kimi K3 weights and the technical report, then amplified Together AI, DigitalOcean, Nebius, Fireworks, Baseten, and other day-zero / early availability signals. It also released PerceptionBench, a benchmark that decomposes visual perception into atomic capabilities. Sources: Kimi K3 weights, PerceptionBench, Together availability.
@ArtificialAnlys: Confirmed that Kimi K3 weights were released and said Kimi K3 is the leading open-weight model on the AA Intelligence Index with a score of 57. It also linked to the GitHub technical report, Hugging Face weights, and AA model page. Sources: AA X, AA links.
@arena: Posted leaderboard signals for Kimi K3 Max, Claude Opus 5 Max, and GPT-5.6 Terra/Luna. Today's Arena X posts are higher confidence because the public leaderboard pages cross-check them. Sources: Kimi K3 Arena, Opus 5 Arena, Terra/Luna Arena.
@OpenAI: Announced GPT-Live in ChatGPT Voice for Edu, Business, and Enterprise globally. Source: OpenAI X.
@AnthropicAI: Posted its open-weight position link, but the page did not load reliably in this run. Treat it as "official X published, page pending verification." Source: Anthropic X.
@cohere: Released North Automations, Cohere North's enterprise AI workflow automation capability. Source: Cohere X.
@HuggingFace / @ClementDelangue: Hugging Face amplified Kimi K3; Clement said Kimi K3 became the top trending model on Hugging Face. Hugging Face also joined the Open Secure AI Alliance. Sources: Hugging Face X, Kimi K3 HF.
@xiaohu: Mainly tracked the SSI/NVIDIA investment and compute expansion, which is useful as a Chinese-language auxiliary signal. It is not a model launch, but it matters for future frontier-lab competition. Sources: xiaohu SSI, xiaohu WSJ source.
@OpenRouter: Posted Kimi K3 availability, GPT-5.6 Terra/Luna discounts, and signals that Kimi K3 Fast providers are coming. Sources: OpenRouter Kimi K3, OpenRouter Terra/Luna.
@togethercompute / @FireworksAI_HQ: Both posted Kimi K3 access and hosting/performance information. Sources: Together X, Fireworks Kimi K3.
@SakanaAILabs: Posted Fugu-Ultra v1.1 and the Claude Code interface for Fugu near the edge of the window. The signal is smaller than Kimi K3/Microsoft/Cohere today, so it remains a secondary item. Sources: Sakana X, Fugu release notes.
@GoogleDeepMind, @GoogleAI, @GeminiApp, @AIatMeta, @MistralAI, @Alibaba_Qwen, @TencentHunyuan, @deepseek_ai, @runwayml, @LumaLabsAI, @Kling_ai, @Alibaba_Wan, @karpathy, @gdb, @sama: No high-confidence new official model launch was seen in this window. @sama replying to the idea that OpenAI could stop shipping new models is a sentiment/hint signal, not a release. Kling/Runway/Luma were mainly posting examples or event content.

Pending Confirmation

Kimi K3: Continue verifying the full technical report's evaluation tables, tool-use behavior, visual-task performance, Chinese long-context stability, MoE inference cost, provider-level speed differences, and the practical SaaS/API constraints of the Kimi K3 License.
Kimi K3 on providers: Together, Fireworks, Baseten, OpenRouter, DigitalOcean, Nebius, and other entrances need hands-on checks for context length, ZDR, region, rate limits, pricing, and real latency. OpenRouter routing may produce different results from a single provider.
PerceptionBench: Kimi released the visual perception benchmark, but the dataset, scoring method, reproducibility, and third-party adoption by Arena/AA or others still need confirmation.
Arena: Kimi K3, Opus 5 Max, and GPT-5.6 Terra/Luna sample sizes are still moving. Terra/Luna rankings are especially new and should be checked again after 48-72 hours.
GPT-Live: Confirm exact availability across ChatGPT Voice plans, API timing, GPT-Live-1/mini model IDs, pricing, latency, interruption handling, background GPT-5.5 delegation boundaries, and enterprise data policies.
MAI-Cyber-1-Flash: Wait for actual Azure AI Foundry / MDASH access details, vetting conditions, whether the model can be called independently, benchmark reproducibility, and the Project Perception public-preview date.
Cohere North Automations: Confirm whether it is only for existing North customers, whether external MCP servers are supported, the granularity of approvals/audit logs, the model choices available per step, and pricing.
Anthropic open-weight position: Official X is live, but the linked page was not reliably accessible in this run. Continue verifying the original text, whether Anthropic clearly opposes a ban, and what boundaries it proposes around export control, testing, and open weights.
Open Secure AI Alliance: Track whether the NOOA GitHub opens core models/weights/harnesses. NVIDIA's blog statement that SpaceXAI plans to open source Grok line weights needs further confirmation from SpaceXAI/xAI.
SSI/NVIDIA: There is no model or API entrance. In the short term, track compute, hiring, papers/safety frameworks, and any first public model evaluation.
XiaoHu AI Daily auxiliary source: Today's run mainly used @xiaohu X posts. For future citations from XiaoHu's site, open individual posts and verify the publication date, source chain, and whether old content has been resurfaced.