X AI Daily¶

Date11 November 2026 · WednesdayWindow2026-11-10 02:00 to 2026-11-11 07:51CoverageComplete

Translation

English translation of the Chinese original.

1. AI Models, Products and Applications¶

Gemini 4 Carbon opens to all users: Google says it tops three benchmarks — CursorBench 4.0, WANDR and AutomationBench¶

On November 10 Google opened Gemini 4 Carbon to developers, enterprises, researchers and ordinary consumers. The product positioning is explicitly split from Argon, released a few weeks earlier: Argon remains the default choice for most software engineering and knowledge work, while Carbon is for tasks that "push model capability to its limit," including complex multi-step coding and debugging and open-ended scientific research, consuming more tokens for deeper exploration, reasoning and longer verification loops. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at less than half the cost per task of the next-best external model; cybersecurity is its strongest domain and requires applying to the Fairwind program to obtain it. Pricing matches Argon's early-access price: $2 per million input tokens and $10 per million output tokens, with the early-access period ending November 30, after which it becomes $4/$20. The same day Google previewed /swarm in Antigravity, splitting a single prompt into a set of agents that can be tracked in real time, letting Carbon take the most reasoning-hungry roles such as architecture design, coordination and auditing, with the rest of the work handed to lighter models and adversarial verifiers.

Carbon is not a new model that only appeared on November 10. The official timeline is that it first opened on October 28 to cybersecurity defenders in the Fairwind program, and two weeks later expanded to all paid developers, enterprises and Ultra subscribers; individual users need a Google AI Ultra subscription to use it in the Gemini App and AI Mode in Search, the enterprise side goes through Gemini for business, Android Studio, Antigravity and CodeMender, and researchers can apply for 6 months of free quota. Artificial Analysis says Carbon tops Intelligence Index v5 with a score of 59, ranking first in 4 of 9 sub-evaluations.

Why it matters: this is the most important frontier model release of this window, with the three official posts from Google, Pichai and Google AI together exceeding 1.87 million views. The more practical change is price: Carbon costs the same as Argon, while on Cursor's public leaderboard Carbon's cost per task is only about a quarter of Claude Fable 5.2 Max.

Key posts: Google's launch post, about 1.157 million views, 7,818 likes, 794 reposts; Sundar Pichai's rollout note, about 475,000 views, 6,197 likes; Google AI's capability note, about 241,000 views, 4,370 likes; Logan Kilpatrick, about 250,000 views, 3,568 likes.
Verified: Google's official blog confirms the November 10 opening, the same $2/$10 early-access price as Argon plus the formal $4/$20 price after November 30, first place on CursorBench 4.0, WANDR and AutomationBench at less than half the cost of the next-best external model, the /swarm preview, the rollout scope for each channel, and the safety figures (99% of adaptive red-team attack trajectories blocked, single-attempt breach rate down from Argon's 3.1% to 0.8%); Cursor's CursorBench leaderboard confirms Gemini 4 Carbon High at 60.4% and $9.22 per task, Medium at 60.1% and $8.49, and Fable 5.2 Max at 60.0% and $32.79, consistent with the comparison posts on X; Artificial Analysis's announcement confirms first place on Intelligence Index v5; the model card is the official description of capabilities and limitations.
Source claim: the 53.8% pass rate on SWE-Marathon v1.1, ahead of Muse Wonder Full at 53.1%, Opus 5.5 at 46.3% and GPT-6 Astra at 42.5%, plus the claim that it is "88% cheaper per solved task than Opus 5.5," appear only in a post from the benchmark owner Abundant Labs, and this time no Carbon row could be verified in their public leaderboard snapshot. Before launch, a tester said that in a brand-new Google Cloud project the global endpoint returned gemini-4-carbon with a matching model version field; this claim likewise rests only on the post.
Assessment: most of the leaderboard results come from Google and its partners and reflect vendor framing; the only thing that can genuinely be cross-checked is the "slightly higher score at much lower cost" on Cursor's own leaderboard, and Google still designates Argon as the default for most tasks, making Carbon look more like a premium substitute for high-difficulty tasks than a wholesale replacement.

Anthropic moves the workbench into Claude: Whiteboards and Sheets enter beta, Dashboards is renamed Data and goes GA¶

On November 10 at its San Francisco event Anthropic released a batch of office-oriented features: Claude Whiteboards and Claude Sheets launched in beta, the former Claude Dashboards was renamed Claude Data and became generally available, and all three are open to every paid plan. Whiteboards is a shared canvas where teams brainstorm with Claude, able to turn a rambling spoken plan into a timeline with owners and dates, or sort everyone's notes by theme at a retrospective and draft action items; selecting any region lets you ask follow-up questions, and when done the whiteboard can become a document or slide deck, or be carried into Lucid, Miro or Vani by Zoho for further editing. Sheets can generate a spreadsheet with worksheets, formulas, formatting and dropdowns from a single prompt, lists each change for easy tracking, and for now only covers simple spreadsheet functions, requiring export to xlsx or CSV when full Excel capability is needed. Data adds certified queries, data storytelling and real-time sharing on top of the existing dashboards, supports caching and change tracking, and its partners include Amplitude, Mixpanel, Omni, PostHog and Sigma. The event also previewed Claude Meetings, where Claude can join a Google Meet call and start follow-up work while the meeting is still running. Interactive connectors were also upgraded into full applications that open beside the conversation or full screen, covering Adobe, Atlassian, Box, Notion and others.

Why it matters: the official post has about 5.14 million views, the most widely spread of the new releases in this window, and it is a clear move by Anthropic to expand from "a tool for programmers" toward an entry point for knowledge work, directly confronting Microsoft's and Google's spreadsheet and dashboard ecosystems, on the same day as its IPO roadshow. The enterprise defaults are worth noting: on the Team plan all three are on by default, on the Enterprise plan whiteboards and sheets are off by default, and Claude Data turns on by default on November 19.

Key posts: Claude's official launch post, about 5.145 million views, 37,600 likes, 2,502 reposts; testingcatalog's rollout note, about 9,300 views, 191 likes; The Rundown AI, about 4,700 views, 18 likes.
Verified: Anthropic's official article confirms that whiteboards and sheets are beta, that Data is Dashboards renamed and now GA, availability on all paid plans, the feature details, export and migration paths, and the enterprise default switches with the November 19 timing; the Claude Meetings preview and event details appear in TechCrunch's report and Fortune's report.
Assessment: of the three products only Data is genuinely generally available; whiteboards and sheets remain simplified versions by Anthropic's own description, and the near-term goal looks more like keeping users inside Claude and reducing leakage through exports to Office; whether it can shake up existing office suites depends on how fast the collaboration and permissions systems are filled in.

Gemini Robotics 2.5 released: evaluated on 200-plus tasks, 96% sorting success in Mercedes-Benz production pilot¶

Google DeepMind released Gemini Robotics 2.5 on November 10, calling it its strongest vision-language-action model to date, while broadening both the range of skills robots can learn and how fast they learn. Google evaluated it on more than 200 tasks, including serving meals with tongs, tying a bow on a gift box and pipetting liquid, and says success rates are 30% to 100% higher across task types versus the prior baseline, with completion 20% to 30% faster. The most concrete number comes from the production line: on Apptronik's Apollo 2 humanoid, Gemini Robotics 2.5 controls the hands and arms while Gemini Robotics ER tracks the process, performing parts sorting and feeding on Mercedes-Benz lines in Kecskemét, Hungary and Sindelfingen, Germany, with a 96% success rate. On safety, the model uses inference-time steering to avoid physical risks in real time, and Google says safe completion rates rise 33% around obstacles and 43% when people are nearby; it also open-sourced ASIMOV-Jailbreak, a benchmark for robot jailbreaks, and says Gemini 4 Argon resists jailbreaks on that benchmark 99.5% of the time versus 82% for the prior Gemini Robotics ER 2 baseline.

Why it matters: this is the latest version wiring frontier language-model capability directly into physical execution, and Google explicitly says Gemini 4 Argon serves as a simulation engine within it — generating thousands of virtual practice runs from a single camera viewpoint and reviewing its own failure footage to find causes, compressing training cycles that used to take months. The iteration speed of robot models therefore depends in part on the iteration speed of general models.

Key posts: Google DeepMind's launch post, about 278,000 views, 3,384 likes, 404 reposts; the Argon training-acceleration note, about 23,000 views, 295 likes; Carolina Parada, about 2,300 views, 58 likes; testingcatalog, about 12,000 views, 223 likes.
Verified: DeepMind's official blog confirms the 200-plus tasks, the 30% to 100% success gains and 20% to 30% speedup, the 96% sorting success rate at the two Mercedes-Benz plants, the 33% and 43% safety improvements, the open-source ASIMOV-Jailbreak benchmark with 99.5%/82% jailbreak resistance, and Argon's role as a simulation engine; the ASIMOV-Jailbreak dataset is a publicly checkable benchmark page; Ars Technica's report provides an independent third-party description.
Assessment: 96% is only a proof-of-concept result for a single sorting-and-feeding task on a real production line, not the same as stable output over long cycles and many tasks; the 200 tasks are also Google's own selection, and cross-embodiment generalization still depends on what partners disclose later.

Unreal Engine 5.9 Preview: Epic says this is the first engine version built mainly by agents¶

Epic posted Unreal Engine 5.9 Preview on its official forum on November 9, ahead of the Epic AI MegaJam running November 12 to 15. The key line in the official notes: this is the first Unreal Engine version built mainly with agentic workflows — a small number of people directing batches of agents, porting the AI toolchain from the unfinished UE6 development branch and adding Web as a target platform; the Web platform code was rewritten by agents starting from the retired UE4-era HTML5 platform, with engineers directing, reviewing and testing. WASM thus becomes an export target again, though still experimental. On features, Unreal MCP starts automatically with the editor and extends to rendering, version control, profiling, debugging, playtesting, geometry, audio, UI and level creation; a new Fab toolset lets agents search the catalog, browse asset libraries and install free or paid assets into a project; and the Unreal Assistant panel can connect directly to coding agents such as Claude Code, Codex, Cursor or OpenCode. The known-issues list centers on Web: save data is lost after reload in C++ Web games built by the launcher, Wasm crashes when memory exceeds 2GB, and the precompiled Mac and Linux builds lack some Web support files; also, new projects have the MCP server on by default and the toolset plugins off by default, so without manually enabling "All Toolsets" the agent has no tools for editing levels.

Why it matters: it is rare for Epic to publicly admit the engine itself is mainly produced by agents, and that sentence matters more to the game industry than any specific feature — it moves the "AI writes code" debate from the application layer to the engineering infrastructure layer. Tim Sweeney's phrasing is "a new era of game development."

Key posts: Tim Sweeney, about 91,000 views, 2,650 likes, 190 reposts; 80 Level, about 15,000 views, 210 likes; GameFromScratch, about 3,800 views, 75 likes.
Verified: Epic's official forum post (November 9, 2026) confirms that 5.9 Preview was made for MegaJam, that it is the first version built mainly with agentic workflows, that the Web platform code was rewritten by agents starting from UE4 HTML5, the specific capabilities of Unreal MCP and Unreal Assistant, the Fab toolset, and known issues including Web save loss, 2GB Wasm crashes and missing files in the Mac and Linux builds; 80.lv's article and Inven Global's article restate the same official notes.
Assessment: "built mainly by AI" is Epic's own account with no third-party audit of the code's provenance, and this preview, which only serves MegaJam, has plenty of known issues on the Web target, so it should not be treated as a production option in the near term.

AI Futures Project updates its timeline: all six authors agree a "superhuman AI researcher" arrives in 2027¶

The AI Futures Project published a Q3.5 timeline update on November 10. The trigger was METR's estimate that Anthropic is currently advancing AI at roughly twice its historical pace (range 1.3 to 3 times), with "about 80%" of the acceleration coming from AI itself. On that basis the team concludes AI-driven acceleration is about 5 months ahead of the AI 2027 scenario. All six authors now agree a superhuman AI researcher will appear at some point in 2027 (assuming no substantial deliberate slowdown), and they put 35% to 65% probability on AI capable of automating all white-collar work appearing before 2028. The article also acknowledges the model may be wrong.

Why it matters: AI 2027 is one of the most influential AI timeline narratives of the past two years, and its authors collectively moving forecasts earlier directly affects safety discussion and investment narratives. Kokotajlo himself wrote in the launch post, "I am scared, sad, and angry," and that post has about 397,000 views.

Key posts: Daniel Kokotajlo, about 397,000 views, 2,842 likes, 456 reposts; Andrew Curran's summary, about 129,000 views, 1,669 likes.
Verified: the AI Futures Project original confirms the roughly 2x acceleration and 80% attribution from METR, the 5 months ahead of AI 2027, the six authors' unanimous prediction of a superhuman AI researcher in 2027, the 35% to 65% TED-AI probability, and the authors' own caveat that the modeling may be wrong; the Techmeme entry records the November 10 publication.
Assessment: this is an update to a forecasting model rather than an observed capability jump; the whole chain of reasoning starts from METR's estimate based on Anthropic internal data, the authors marked their own error bars, and treating "5 months" as a precise countdown is not appropriate.

Other research and model releases¶

Geodesic released NovaAtom-Lite on November 10, an all-atom structure prediction model for drug discovery that can predict the structures of proteins, nucleic acids and the small molecules they bind, and how they fit together. The results given in the post: on FoldBench antibody-antigen structure prediction (DockQ ≥ 0.23) the success rate is 79.07% (n=172), above IsoDDE's 75.58% and AlphaFold 3's 47.90% (n=167), with evaluation using 5 random seeds × 5 samples and taking top-1 by the model's own ranking score. The model is available through the Geodesic API and NovaDDE. What has been verified is the positioning: Geodesic's site describes NovaAtom-Lite as "a lighter, faster version previewed before full-model training is complete"; the FoldBench antibody-antigen task is indeed one of the hardest categories in that benchmark, and the FoldBench paper notes that most prior methods failed more than 50% of the time on such tasks. The specific scores rest only on the official post and reposts, and should be treated as a source claim. Key post: Geodesic, about 243,000 views, 459 likes.

MBZUAI's Institute of Foundation Models released PAN 2, an open foundation world model for physical intelligence that uses generative latent prediction: an LLM first writes in text "what will happen and how the camera moves," and a 29B video diffusion model trained from scratch renders it into video of no more than 20 seconds, letting agents simulate consequences before acting. The comparison given officially and in reposts: without the LLM planner, multi-step tasks reach only 8.7% of steps correct with no fully completed plan; with both parts together, 75.1% of steps are correct and 34.0% of plans are completed in full, the highest among the models IFM has tested. Weights, training and inference code are open under Apache 2.0, with first-day support in vLLM-Omni. What has been verified is the open-source and architecture description: IFM's PAN 2 project page is the official entry point for the model, and daily.dev's summary of IFM's technical blog confirms the two-stage design of an LLM plus a 29B diffusion model, the 20-second duration and the Apache 2.0 license; the performance numbers come from the vendor and secondary restatements and should be treated as a source claim. Key post: IFM, about 286,000 views, 367 likes.

NVIDIA released Alpamayo 2 Edge on November 10, a 2B-parameter reasoning vision-language-action model designed for real-time in-vehicle inference: adaptive reasoning generates a driving trajectory and provides causal explanations when needed, aimed at scenarios from yielding to pedestrians to giving way to oncoming buses. Technically it builds on the Cosmos 3 Edge VLM backbone plus a diffusion-based action decoder, is the first to be post-trained with closed-loop reinforcement learning in AlpaGym, introduces AlpaJudge for human-calibrated reasoning evaluation, and ships with quantization and inference-acceleration recipes. Verified: the Hugging Face model card confirms 2B parameters (a 2.45B backbone plus a 0.2B action expert), the Cosmos 3 Edge base, weights released under OpenMDW-1.1, Apache 2.0 code, and deployment targeting DRIVE AGX Thor; NVIDIA's technical blog explains that BF16 weights can run directly on supported GPUs but that real-time performance on Thor still needs inference optimization. Key posts: NVIDIA, about 111,000 views, 425 likes; José Alvarez, about 1,500 views, 27 likes.

Lila Sciences' Open-Endedness team published the company's first foundational research on November 10: one blog post plus four preprints, covering how to sidestep the "creativity paradox" during training, whether deliberately keeping a model "naive" in some domain confers an advantage, a new regularization method called UFO, and why networks with unified factored representations are naturally resistant to catastrophic forgetting. Verified: two of the preprints were located this time, Creative Priming and Unified Factored Organizers, both public on arXiv with authors affiliated with Lila Sciences and topics matching the post's description; the other two were not individually located. Key post: Kenneth Stanley, about 94,000 views, 305 likes.

2. AI Compute and Industry¶

Google releases TPU 8tx and 8i Max: four eighth-generation TPU systems in one year¶

On November 9 Google expanded its eighth-generation TPU lineup with two new systems. TPU 8tx targets secure and confidential inference, using the same chip as 8t with fifth-generation AMD EPYC hosts, building encryption and isolation into the hardware layer so sensitive data and model weights are never decrypted even in use, and reusing the low-latency Boardfly network designed for real-time inference; Google says latency is up to 2x lower than TPU 8t with near-zero performance loss, and the same system can also do private fine-tuning, distillation and reinforcement learning without moving sensitive data to another cluster. TPU 8i Max addresses the memory wall: two 8i chips in a single package, doubling compute, SRAM and HBM bandwidth versus a single 8i with no change in rack footprint, which Google says means more compute deployable per square foot and per watt. The four systems divide the work as 8t for large-scale training, 8i for low-latency inference, 8tx for confidential computing and 8i Max for ultra-long context and maximum rack density. The same batch announced open-sourcing of TorchTPU and TPU Sync, and general availability of Cloud Pathways and GKE dynamic slices. Google says the TPU business now has nearly 100 customers, including frontier labs, trading firms, retailers and universities.

Why it matters: Pichai makes "one chip generation, four systems in a year" the core selling point — hardware teams co-designing with Google DeepMind and responding to customer needs faster through packaging and networking, rather than waiting for a semiconductor cycle that turns every few years.

Key posts: Sundar Pichai, about 358,000 views, 5,334 likes, 357 reposts; Decrypt's report post, about 2,800 views, 41 likes.
Verified: Google's official blog confirms the November 9 release, four systems in one year, 8tx's confidential computing and 2x lower latency, 8i Max's dual-chip package with doubled compute, SRAM and HBM bandwidth, nearly 100 customers, and the progress on TorchTPU, TPU Sync, Cloud Pathways and GKE dynamic slices; the Google Cloud blog is the cloud-side version of the same content.
Assessment: Google gave no absolute performance numbers, prices or customer delivery timelines, so price-performance still awaits third-party benchmarks, and the nearly 100 customers claim has no breakdown by industry.

Mercor closes a $750 million Series D at a $20 billion valuation¶

AI data and evaluation company Mercor announced a $750 million Series D on November 10 at a $20 billion valuation, up from $10 billion at its Series C a year earlier. Investors include NVIDIA, Benchmark, General Catalyst, Felicis, an Apollo-affiliated fund, Salesforce, Zoom, XTX Ventures, StepStone and Glade Brook, and the round also included secondary sales of employee shares. The company says its network includes more than 5 million experts training frontier models with their own domain expertise, that it has expanded to New York and London this year, and it reiterated that its earlier acquisition of Deeptune this year was to provide agents with simulated application environments for practicing enterprise workflows. The money goes to evaluation and post-training infrastructure and to more environments where agents learn complex workflows.

Why it matters: a doubled valuation in a year shows that human expert data and evaluation environments remain scarce inputs for frontier model training. Mercor's pitch has also expanded from "supplying data to frontier labs" to "letting enterprises own their own intelligence," targeting enterprise post-training demand.

Key posts: Molly O'Shea's interview and figures, about 110,000 views, 366 likes; IPO Newsroom's summary, about 2,100 views, 5 likes.
Verified: Mercor's official blog confirms the $750 million Series D, the $20 billion valuation, the employee secondary sale, the investor list (including NVIDIA, Salesforce, Zoom and XTX Ventures), the network of more than 5 million experts, the Deeptune acquisition and the New York and London offices; the company press release (Business Wire, November 10) confirms the amount and valuation; the doubling comparison comes from the prior $10 billion valuation mentioned in TNW's report.
Assessment: a doubled valuation in one year shows human expert data and evaluation environments are now priced as scalable infrastructure, but the company disclosed no revenue, so operating figures such as 5x growth can only be taken as company framing.

Anthropic launches its IPO roadshow: 825 million shares at $70 to $78, ticker ANT¶

Anthropic began its initial public offering roadshow on November 10. The plan is to offer 825 million Class A common shares, of which 801.35 million are issued by the company and 23.648 million come from existing shareholders selling down, with the company receiving no proceeds from the secondary portion; underwriters have a 30-day overallotment option for up to 123.75 million additional shares. The price range is $70 to $78 per share, and the company has applied to list on the Nasdaq Global Select Market under the ticker ANT. The company explicitly states that the registration statement has not yet become effective and that no securities may be sold nor purchase offers accepted before it does. The same day the company also announced Claude's office features in San Francisco.

What spread more widely the same day was a video of Dario Amodei. In the video he says cancer, Alzheimer's, heart disease and mental illness are "things that I think will be in the past in less than a decade," and that "if we get this right, we can create an era of miracles where humans solve problems that have plagued us for thousands of years," while stressing that only companies that handle risk well will earn social trust and receive operating licenses.

Why it matters: this is the first time a leading AI lab has entered the public-market pricing process, and the share count and price range in the roadshow filings are checkable hard facts; the "cure disease within a decade" line circulating the same day is more of a narrative expression, and the two differ completely in reach: likes on the video quote post are a hundred times those on the roadshow flash.

Key posts: First Squawk quoting the IFR roadshow flash, about 27,000 views, 35 likes; firesidealpha's Amodei quote post, about 140,000 views, 1,594 likes, 152 reposts; ns123abc's restatement, about 226,000 views, 5,229 likes.
Verified: Anthropic's official roadshow announcement confirms the November 10 roadshow launch, the 825 million share structure and secondary proportion, the $70 to $78 range, the Nasdaq Global Select Market and ticker ANT, the overallotment option, and the note that the registration statement is not yet effective.
Source claim: the timing of the roadshow launching "Tuesday morning" was first reported by IFR and restated on X as a flash; the original IFR text was not obtained this time. The full source and publication time of the Amodei video could not be verified, so the main text only transcribes it as quoted in the post, and Manton Reece's blog records the same quote circulating that day.
Assessment: the hard facts of the roadshow all match the official announcement and represent substantive progress, but the pricing date and first trading day depend on the registration becoming effective, while "disease in the past within a decade" is only a personal prediction, not a company commitment and with no clinical evidence.

Fireworks AI incident widens: customer billing data, datasets and partial model weights accessed¶

Fireworks AI updated its October security incident notice on November 9: attackers entered the environment where its R&D machines sit through a third-party cloud provider, gained access on October 23 and ran unauthorized workloads on October 24, affecting more customer data. The earlier October 10 notice only said environment variables used by evaluation tasks (including API keys and tokens) had been obtained in a small number of accounts, with no models, datasets or personal information found affected. The updated scope includes resource metadata (resource identifiers, display names, descriptions, creation times), billing information (billing and shipping addresses, invoice and payment history, account balances, credit usage), and for some customers partial model weights and subsets of datasets. The company says it does not store payment card numbers, so card numbers were unaffected, and that it has notified affected customers and completed credential revocation and rotation, system isolation, tightened access controls and enhanced logging and monitoring. The company attributes the incident to a cybercrime group profiting from selling access to large models, consistent with LLMjacking-style attacks. Notion also disclosed that fewer than 1% of workspaces were affected by the incident, and that follow-up analysis found it involved Notion AI chat records used for the Insights feature, most of which cannot be attributed to a specific user or workspace.

Why it matters: this is a case that escalated from "credential leak" to "model weights and training data leak," directly touching the two asset classes customers care most about, with the entry point in a supplier's third-party infrastructure. For teams using hosted inference and fine-tuning services, this is a real-world lesson on key rotation and third-party hosting review.

Key posts: The Cyber News, about 4,900 views, 35 likes; hiroppy's restatement of the Notion notice, about 2,000 views, 3 likes.
Verified: The Hacker News report confirms the first notice on October 10, entry into the R&D environment via a third-party cloud provider on October 23, unauthorized workloads on October 24, the affected data categories, that the company does not store card numbers, the attribution to a crime group profiting from LLMjacking, and Notion's fewer-than-1% affected workspaces and AI chat records; Fireworks' official security bulletin is the original disclosure; Notion's security advisory confirms the wording about a third-party AI vendor incident.
Assessment: the incident has been disclosed over multiple rounds and the basic facts are credible, but the list of affected customers and the specific meaning of "partial weights" are not public, so what customers can do is still to rotate keys and audit third-party hosting paths.

3. Chinese AI Ecosystem¶

MiniMax open-sources H3-Image: a single 4K image in under 10 seconds, weights already on Hugging Face¶

MiniMax released the image model MiniMax H3-Image on November 10, built on its earlier omni-modal model H3 and supporting text-to-image, reference-image generation and image editing. The company says it can generate a 4096 × 4096 4K image, about 16 million pixels, in under 10 seconds, producing native output at any aspect ratio without upscaling; up to 16 reference images can be used, and it emphasizes photorealistic detail and prompt adherence, claiming to be the strongest instruction-follower among open-source models. The weights are open under a community license and the model is also available through the MiniMax API, MiniMax Code and 海螺 AI, with standard mode 1.5 to 2 times faster than mainstream options and turbo mode 3 to 4 times faster, at about 50% to 70% of the price of comparable mainstream options, plus an extra 50% off 4K and higher resolutions in the launch week.

Why it matters: the Chinese-language community spread it as "an open-source 4K image model," and the buzz comes mainly from the open-sourcing itself rather than benchmark results. A domestic image model pressing on price and speed by opening its weights is a real option for studios that need local deployment and model modification.

Key posts: MiniMax's official launch post, about 389,000 views, 2,775 likes, 244 reposts; Chinese-language restatements such as VeknQAQ, about 950 views, 9 likes, and daveylan2, 300 views, 3 likes; the hands-on comparison post by underwoodxie96, about 2,000 views, 10 likes, finds that with the same prompt H3-Image's layering is a bit flatter than GPT Image 2.5.
Verified: MiniMax's official blog confirms the H3 base, 4K in under 10 seconds, up to 16 reference images, open weights, availability through the API plus MiniMax Code and 海螺 AI, the relative speeds of standard and turbo modes, and the 50% off 4K and above in launch week; the Hugging Face model card confirms a 33.1B-parameter single-stream DiT distilled to 6-step sampling, about 16 million pixels per aspect ratio (4096 × 4096 at 1:1, up to 6272 × 2688), the text encoder carried over from H3's Qwen3-VL, and weights released under the MiniMax-H3-Image community license.
Assessment: the open weights are certain, but "strongest instruction-following among open-source models" has no public third-party evaluation behind it, Chinese-language discussion mostly stops at reciting parameters, and actual image quality, editing stability and the commercial terms of the community license all need more independent reproduction and line-by-line checking.