Skip to content

← X AI Daily

X AI Daily

Date20 October 2026 · TuesdayWindow2026-10-19 02:00 to 2026-10-20 07:56CoverageComplete

Translation

English translation of the Chinese original.

1. AI models, products, and applications

MiniMax M3.1 Flash released: Space Bunny goes official, aimed at agent cost

  • What happened: MiniMax released M3.1 Flash on 10-19 at 19:56, confirming that Space Bunny — the model that had been anonymously flooding OpenCode and OpenRouter for free — was its early version, and that the anonymous model had cumulatively processed more than 200 trillion tokens. The official selling points: agent task cost 70% lower than M3, online inference nearly 3 times faster, flat API pricing of 0.15 dollars per million input tokens, 0.60 dollars output, 0.015 dollars cached reads, with weights to be released on October 23. The same account posted a teaser at 17:21, then early on 10-20 said M3.1 Flash would be free on OpenCode for 3 days and free on OpenRouter for 72 hours.
  • Why it matters: the Flash tier is becoming the main tier for agents, and the competitive point is shifting from "first on the leaderboard" to "the cost of completing one multi-step task." Space Bunny running volume under an anonymous identity first and only then claiming it is also a rare form of product validation: let developers use it in real workflows first, then disclose the identity and the price.
  • Key posts: MiniMax release post (10-19 19:56, 889K views, 6,472 likes, 566 reposts, 371 replies, 554 quotes); teaser post (10-19 17:21, 802K views, 6,557 likes); OpenCode free post (10-20 04:37, 5,889 views, 153 likes); OpenRouter free post (10-20 04:57, 3,982 views, 81 likes). Chinese summary @NFT_Chen (13K views, 80 likes), price comparison @xiangxiang103 (3,843 views, 16 likes); usage notes @MATKA_r (3,092 views, 36 likes).
  • Verification: against the MiniMax official blog, it is confirmed that Space Bunny processed over 200 trillion tokens and is an early version; inference "nearly 3 times faster"; agent usage cost down 70%; flat pricing 0.15 / 0.60 / 0.015 dollars; weights "released later" (the post says October 23); the scores listed on the official blog are DeepSWE v1.1 73.7%, Agents' Last Exam 29.6, OSWorld V2 61.7, CyberGym 82%, BankerToolBench 76.2, output 200+ tokens/s, P50 170+. Of the X post's 74%, 30%, 62% and 91%, the first two are roundings of the blog numbers; Automation-Bench 62% has no numeric value given in the blog body, and BabyVision 91% does not appear in the blog body, so neither could be checked on a first-party page. 1M context and image/video input are not in the blog body; the llm-stats model page and the OpenRouter free page show 1M input, 524K output, text/image/video input; total parameters of 428B with 23B active appear only in secondary reports such as MarkTechPost, the official blog gives no parameter count, and this article does not adopt them. The Chinese post's "1 yuan input and 4 yuan output per million tokens" is roughly equivalent to the dollar price at exchange rates; "same as DeepSeek's off-peak price" is the poster's own cross-comparison, and DeepSeek's price list for the period was not checked line by line.
  • Assessment: M3.1 Flash's positioning is clear — it is not chasing the strongest model, but pushing the per-task cost of agents down to a level where they can be run in bulk. Two things need watching: the official scores mix two sets of numbers, from the blog and from the post, and some items have no published values; the weights do not arrive until October 23, so until then "open source" is only a promise. Space Bunny's leaderboard first place is vendor narrative; what is actually verifiable is the weights after October 23 and third-party reproduction.

OpenAI reports progress on COBOL, Fortran and Visual Basic, and open-sources MultiLangBench

  • What happened: OpenAI posted on 10-20 at 06:30 that its latest models are stronger at building and migrating low-resource programming languages such as COBOL, Fortran and Visual Basic, with the focus on understanding decades of business logic and preserving existing behavior during modernization, and it invited enterprise customers to suggest the next language and workflow to tackle.
  • Why it matters: this splits "writing code" into two questions that are closer to enterprise budgets: building from scratch (writing a SQLite-compatible engine in 20 languages) and migration (moving COBOL to C++, Java, Kotlin, Python, Rust with unchanged behavior). The latter maps directly onto the installed systems of banks, governments and insurers, and is also where AI coding has the hardest time earning production trust.
  • Key posts: @OpenAI (10-20 06:30, 206K views, 2,478 likes, 169 reposts, 237 replies, 116 quotes).
  • Verification: against the OpenAI official blog post (published 10-19), the design and scores of both evaluations are confirmed: MultiLangBench requires models to build a SQLite-compatible database engine from scratch, using the database file as the cross-language interface; GPT-6 Astra averages an 87.4% test pass rate across 20 languages, while GPT-5.6 Sol is 26.8%; by language, COBOL rises from 10.9% to 72.1%, OCaml 25.6%→85.8%, Fortran 21.4%→83.5%, Visual Basic 27.5%→89.4%, with Astra above 83% on 19 of 20 languages, COBOL still the hardest. GPT-6 Luna averages 40.3%, with an estimated API cost of 0.77 dollars per sample versus 3.48 dollars for GPT-5.6 Sol (about 78% lower). The COBOL migration evaluation requires byte-identical output and all hidden tests passing to count as success: GPT-6.1 Sol at low reasoning effort scores 86.6%, GPT-5.6 Sol 5.4%, at about 61% of the latter's cost (the blog's opening writes "39% cheaper," which is consistent); at high reasoning effort it is 97.9%. The evaluation set and per-language results are in openai/MultiLangBench. The X post itself has no numbers; "800 billion lines of COBOL" comes from CryptoBriefing's coverage, not from OpenAI's own text, and is not cited here.
  • Assessment: this treats "low-resource languages" as an independent capability dimension rather than waiting for general coding ability to spill over; COBOL migration uses hidden tests and byte-level consistency, close to real compliance requirements, which is more meaningful than "it compiles." But all scores are vendor self-reported, and OpenAI itself says production migration still needs expert review; what really matters is whether enterprises can use this evaluation to reproduce their own codebases, not the average pass rate.

Meta publishes the Muse Glimmer technical report: 30B open weights, trained inside a "Sims-like" life

  • What happened: Meta announced on 10-20 at 03:33 the release of the Muse Glimmer technical report. It is a 30B open-weight model aimed at agent tasks on consumer hardware, able to run on a single consumer GPU after 4-bit quantization, and the company says it leads same-size open models in tool use and multi-step planning. The report discloses the architecture, the distillation recipe from a larger model, quantization and speculative decoding.
  • Why it matters: the model itself shipped back in August; what is new here is the technical report, and the value is in the training method: Meta trains the agent on "simulated digital lives" with 10–14 contacts and weeks of email, calendar, document and message history, and removes hallucinations, false successes and unverified trajectories from the data. That says more about how to build a "local personal agent" than another agent benchmark score.
  • Key posts: @AIatMeta (10-20 03:33, 68K views, 721 likes, 91 reposts, 73 replies, 63 quotes); analysis @kimmonismus (28K views, 241 likes); team member @arreqe_ai (303 views, 6 likes).
  • Verification: against the Hugging Face model card, it is confirmed as Apache 2.0, about 29.6B dense parameters (including about 1.8B vision encoder), 131,072+ context, knowledge cutoff 2026-01-04, runnable in 24GB / 32GB of VRAM; K-Quant-17GB on an RTX 5090 (Meta's own testing) reaches 74.9 tokens/s without speculative decoding and 233.4 tokens/s with DFlash speculative decoding (3.1 times), and 37.8 on M4 Max, 50.2 on M5 Max. Against the Meta technical report page and its PDF, section 6.2's synthetic data pipeline is confirmed: each "universe" is a pre-populated digital life containing 10–14 contacts, weeks of correspondence, calendar, documents and message history, with tasks spanning multiple days; post-processing removes failed, duplicate and fourth-wall-breaking items, plus any trajectory containing hallucinations, false successes or unverified results. kimmonismus's account matches the report. On independent evaluation, Artificial Analysis gives Muse Glimmer 35 an Intelligence Index of 35, below Qwen3.6 27B's 38 and above Gemma 4 31B's 30; GDPval-AA v2 is only 953 Elo, below Qwen3.6 27B's 1141; AA-Omniscience hallucination rate is 82% versus Qwen3.6 27B's 49%. Meta's claim of "leading at the same size" mainly holds for tool calling and some agentic items, not across the board.
  • Assessment: the interesting part of Muse Glimmer is not "30B can fight frontier models" but the engineering combination of "30B + quantization + speculative decoding + distillation from simulated environments" that puts an agent on local hardware. The weaknesses exposed by independent evaluation are also clear: weak knowledge work and calibration, high hallucination rate. It fits as a local execution layer, not as a final arbiter for decisions requiring highly reliable facts. The data-filtering rules in the technical report are more worth copying by other teams than the leaderboard rankings.

Inception releases Mercury 3: a diffusion language model running at 1,068 tokens/s

  • What happened: Inception released Mercury 3 on 10-20 at 01:37, calling it the strongest diffusion language model to date: more than 40 times larger than Mercury 1, 3 times the intelligence on standard benchmarks, still 1,000+ tokens/s, 1M context, and the company says it beats GPT-6 Luna and Claude Haiku 5.5 on multiple benchmarks. It went live the same day on the Inception API, Baseten and OpenRouter.
  • Why it matters: diffusion models generate in parallel and then refine iteratively, which naturally suits voice agents, real-time tutoring and customer service — scenarios where being one second slow ruins everything. Mercury 3 wants to show that the diffusion route keeps its speed advantage after scaling 40 times, without having to choose between "fast" and "smart."
  • Key posts: @_inception_ai (10-20 01:37, 36K views, 603 likes, 45 reposts, 56 replies, 43 quotes); co-founder @volokuleshov (5,537 views, 106 likes); co-founder @adityagrover_ (4,148 views, 104 likes); media coverage @TheRundownAI (3,603 views, 17 likes).
  • Verification: against the Inception official blog, it is confirmed as 1,068 tokens/s (NVIDIA GPU), 1M context (131,072 output cap), more than 40 times larger than Mercury 1, about 3 times the intelligence, list price 0.20 dollars per million input / 0.70 dollars output, at half price 0.10 / 0.35 during the launch period, cached input at one tenth of the input price (i.e. 10% of list), default reasoning effort low, with instant / medium / high also available. Key gap: the blog contains no benchmark scores at all; "beats GPT-6 Luna and Claude Haiku 5.5" is a comparative claim with no margin, benchmark name or per-item results; the only structured item is a 50% completion rate on 50 Mind2Web navigation tasks on its own harness, which the company says is significantly higher than GPT-6.1-Sol with the same token volume and step count. Third-party OrcaRouter's analysis notes that 1,068 gives no hardware model, batch size or reasoning effort, and that the default low setting distorts comparisons; Artificial Analysis had no Mercury 3 page as of October 20, and the previous generation Mercury 2.5 is about 706 tokens/s under AA's methodology. The Baseten, OpenRouter and Vercel launches match the blog; the X post omits Vercel.
  • Assessment: this is a substantial scaling demonstration for the diffusion route, aggressive on all three of speed, context and price; but "smarter" currently rests only on the vendor's comparative sentence, with no reproducible numbers. Businesses that are latency-sensitive and can fall back can run a week of shadow evaluation first; for tasks where errors are costly, wait for third-party harnesses. Note also that 0.10 / 0.35 is launch-period pricing; the list price is what belongs in a cost model.

Ant's Ling open-sources five Ming-3.0 speech models, with 0.5B taking first among open weights for controllable voice

  • What happened: the Ant Ling team open-sourced five Ming-3.0 models on 10-19 at 23:54, across three speech families: Ming-3.0-tts includes a 0.5B model for voice design and paralinguistic control plus a Tiny 7.9B-A1.3B model for long audio; Ming-3.0-audio includes Flash 126B-A6.5B for audio understanding and tool calling plus Tiny 9.3B-A2.7B for efficient conversation; Ming-3.0-asr is Tiny 8.6B-A2B, focused on dialects and context.
  • Why it matters: the 0.5B model simultaneously does zero-shot cloning, text-based voice design, paralinguistic tags such as laughter, sighs, coughs and whispers, and covers Chinese dialects; the 126B audio-flash turns "understanding long audio, music and acoustic scenes" into a separate tier. For an open-source family, the size range and the license are more useful than a single SOTA.
  • Key posts: @AntLingAGI (10-19 23:54, 384K views, 640 likes, 89 reposts, 40 replies, 32 quotes); ModelScope (10-19 22:37, 6,427 views, 135 likes); TeksEdge (10-20 01:15, 3,814 views, 128 likes).
  • Verification: all five Hugging Face repositories are MIT, with sizes and active parameters matching the post (0.5B dense, tts-tiny 7.9B / 1.3B, audio-tiny 9.3B / 2.7B, audio-flash 126B / 6.5B, asr-tiny 8.6B (about 9B) / 2B). The timeline needs correcting: the 0.5B weights were created on September 22 and uploaded on September 24, while the GitHub inference code was only created on October 19, so "open-sourced today" holds for the 0.5B only as "a runnable version"; the model itself predates the window. Artificial Analysis added this checkpoint on October 19: Controlled Voice 1,069 Elo (range 1,053–1,085), first among open weights; Provider Voice 1,188 (1,173–1,203), third among open weights, behind Fish Audio Drama 3 Preview's 1,245 and Breeze TTS 2's 1,220; median 42.1 characters/s on AA hardware. ModelScope's "first among open weights" holds only on the Controlled Voice sub-leaderboard, and TeksEdge's 1,069 matches AA. The official post says "you can serve it with vLLM-Omni today," but in reality it is an unmerged PR (#9119) that only serves tts-tiny, not audio-flash, audio-tiny or asr; "day-zero support" needs a discount. The 0.5B's WER, speaker similarity and other tables are all Ant's self-reported numbers, with no third-party reproduction yet.
  • Assessment: what really carries weight is the 0.5B controllable TTS (cloned voices first among open weights) and the 126B audio understanding, not the count of "five models open-sourced together." The license is permissive and the family is complete, which suits building local voice products; but deployment has to be checked checkpoint by checkpoint, audio-flash needs a machine on the order of eight H800s, and "open source" should not be mistaken for out-of-the-box.

Dream Labs comes out of stealth: Dream-1 controls a dual-arm robot it has never seen, 90 million dollar seed round

  • What happened: Dream Labs came out of stealth on 10-20 at 01:13 and released the robot foundation model Dream-1, which it says can control dual-arm robots it has never seen, perform tasks it was not taught, and work in environments it has not visited, with the Dream API open the same day at 3 dollars per hour. The same day it announced a 90 million dollar seed round led by B Capital and Kindred Ventures.
  • Why it matters: robot foundation models have long been stuck on "swap in a new machine and you have to retrain." Dream-1 learns from human video and controls across embodiments zero-shot, and the API is priced directly at 3 dollars per hour — selling robot intelligence as a cloud service rather than selling whole machines or project-based delivery.
  • Key posts: @dreamlabs_ai (10-20 01:13, 85K views, 965 likes, 105 reposts, 93 replies, 95 quotes); Robocurve (7,340 views, 69 likes); Techmeme (2,917 views, 2 likes); clankrmedia (332 views, 6 likes).
  • Verification: the press release, Techmeme and multiple reprints confirm 90 million dollars, led by B Capital and Kindred, with participation from A.Capital, Unusual Ventures, SV Angel, Wing VC and Scribble VC; the company was founded by former NVIDIA robotics researchers and works on GR00T, DreamZero and World Action Models; Dream-1 learns from human video and initially supports dual-arm robots with parallel grippers for tabletop pick-and-place and clothes folding, with more than 20 hardware partners, robots priced at roughly 2,000–10,000 dollars; the Dream API is 3 dollars per hour and the lightweight Dream-1-flash 2 dollars per hour, each call returning 1.6 seconds of motion; the company says it tested on 27 robots it had not trained on, and Unite.AI adds that this was in simulation, with 300 rollouts per embodiment. Robocurve's RoboDojo-RC Tier 1 leaderboard gives: Dream-1 59.3%, Gemini 4 Argon (high) 48.2%, GPT-6 Astra (high) 42.5%, GPT-6.1 Sol (high) 41.7%, Opus 5.5 (high) 38.2%, with Dream-1 at about 0.15 dollars per attempt versus 0.66–3.48 dollars for the competitors; "first to exceed 50%" holds. The reprint posts write Opus 5.5 as 48 and GPT-6 Astra as 47, which does not match that leaderboard — 48.2 is Gemini 4 Argon. The 3 dollars per hour is the API price and is not the same as per-task cost; Robocurve estimates 0.15 dollars per attempt.
  • Assessment: the third-party leaderboard score is the most valuable part of this story (Robocurve's evaluation methodology and independence were not examined in this issue), turning "cross-embodiment" from a demo into a comparable score; but 27 robots is still simulation, real-world contact, lighting and object variation are not covered, and long-task reliability is unproven. The 90 million dollar seed round shows capital is willing to bet on an "Android for robots"; what to watch next is whether real customers stick with it on non-tabletop tasks.

Skyfall AI releases SKY-001, calling it the first engineering world model

  • What happened: Skyfall AI released SKY-001 on 10-20 at 01:58, calling it the first engineering world model: on CAD and BIM tasks it is faster, cheaper and has higher intersection-over-union (IOU) than GPT-6 Astra and all tested frontier models — specifically 51–125 times faster and 150–210 times cheaper. It does not produce static 3D meshes; instead it decomposes text or 2D sketches into executable modeling steps, breaking parts down layer by layer and reassembling them in 3D, outputting designs that open directly in Autodesk Fusion. The post says the team comes from the Maluuba deep learning lab acquired by Microsoft and worked on this for 18 months.
  • Why it matters: this is an attempt to move the "world model" from generating video toward engineering geometry and spatial reasoning. LLMs are good at retrieval and code but poor at the hierarchy of physical objects, assembly order and parametric feature trees; if SKY-001 really outputs buildable CAD trajectories directly, the impact lands on engineering outsourcing and the front end of manufacturing, not on content creation.
  • Key posts: @spisallyouneed (10-20 01:58, 883K views, 3,499 likes, 403 reposts, 317 replies, 279 quotes); official account @skyfallai (4,834 views, 65 likes, 14 reposts, 13 replies, 4 quotes); coverage @testingcatalog (8,360 views, 98 likes).
  • Verification: the press release confirms SKY-001 is a hierarchical, compositional visual world model that plans complete CAD trajectories from text or 2D sketches, recursively decomposing down to 2D sketch strokes and 3D feature extrusions, supports exploring multiple designs in parallel, and can be opened in Autodesk Fusion; it is currently a research preview, with commercial deployment targeted for early 2027. But the press release's wording is "100x faster" and "150x lower cost," with comparisons written as GPT-4, Gemini, Claude Opus, Grok and leading open models — not GPT-6 Astra; the X post's 51–125 times and 150–210 times do not match the press release and are the founder's more granular framing. There is no paper, model card or evaluation protocol, and no specific IOU value or test set is given, so third parties cannot reproduce it. "First engineering world model" is marketing; the post writes that the company was acquired by MSRResearch, when the accurate statement is that it was acquired by Microsoft (in 2017) and the team joined Microsoft Research.
  • Assessment: the direction is sound — replacing the LLM's token-by-token generation with geometric and spatial priors really could bring down the inference cost of CAD tasks; but right now there are only demos, a press release and self-reported numbers, with no reproducible benchmark. It is better treated as a product preview for the "engineering world model" category, and judged after commercial deployment in 2027 and third-party evaluation.

Cloudflare releases Clairic: no generation, just typed values back

  • What happened: Cloudflare released Clairic on 10-20 at 02:06, calling it a System One extraction model that can extract, compute, estimate and predict numeric values from unstructured text or images, in a single forward pass without generating tokens, outputting typed results with confidence / probability. It is already in preview on Workers AI as @cf/cloudflare/clairic, with open weights "soon."
  • Why it matters: a large share of business calls are really just "put the amount, date and order number from this email into a table," and doing that with a generative LLM is both slow and billed by output token. Clairic turns these tasks into regression problems: no options to choose from, direct estimation of continuous values, plus p10–p90 intervals, which engineering-wise is closer to an auditable extractor.
  • Key posts: @Cloudflare (10-20 02:06, 53K views, 897 likes, 106 reposts, 27 replies, 44 quotes); @michellechen (95K views, 1,005 likes); @kevin_flansburg (32K views, 372 likes).
  • Verification: against the Cloudflare documentation, it is confirmed as 65,536 token context, 0.10 dollars per million input tokens, no output token billing, support for text and up to 4 images, field types including text / number / date / list / estimate, returning confidence, p_null and intervals; requests can split fields across multiple parallel passes of 8,192 tokens. Against PPC Land's item-by-item breakdown of Cloudflare's own testing, the official three-stage claims pipeline is confirmed: Clairic + Clef + GLM-5.3 end to end in 19.3 seconds, versus 61.7 seconds for GLM-5.3 throughout (about 3.2 times faster); cost per million claims 6,032 dollars versus 11,740 dollars (49% saved); extraction plus decision together 0.68 seconds, versus 38 seconds for GLM-5.3 on the same two steps. Note: the blog's "higher accuracy" has no general benchmark behind it, only one claims case and a 21-day forecasting test, with the number of runs undisclosed; the open weights have no date; kevin_flansburg says "based on regression rather than classification," while the blog frames it as System One / decision model without giving regression formula details.
  • Assessment: this is a clear example of splitting decisions and extraction out of generative models. For invoices, claims, routing and risk control — businesses with fixed fields and high call volume — the cost and latency gains are real; but at preview stage, with a single pipeline self-test, the right move is shadow-traffic comparison first, not directly replacing production decisions.

Three labs queue up new models: Anthropic, Gemini 4 Carbon, OpenAI Terra

  • What happened: The Information reported on 10-19 that Anthropic is preparing to release a new model that better understands user intent, with stronger reasoning and intuition than existing commercial models; Google is internally testing Gemini 4 Carbon, positioned one tier above the just-released Argon; and OpenAI is expected to release mid-tier models in the Terra family within weeks, with higher performance and a lower price than GPT-6.1 Sol. Over the past month the three companies have released more than 10 new models.
  • Why it matters: the center of competition is shifting. Anthropic sells "it understands even when you don't say it clearly," Google sells catching up on vision and coding, OpenAI sells "a cheaper mid-tier" — none of them is simply stacking the highest scores. For developers, the coming month is a window to re-compare prices and switch default models.
  • Key posts: @Mr_Salio (10-19 22:22, 22K views, 308 likes, 17 reposts, 19 replies, 5 quotes); @AIStockSavvy (8,069 views, 43 likes); Chinese summary @NFT_Chen (14K views, 87 likes); @WesRoth (10-20 07:00, 1,078 views, 15 likes).
  • Verification: the original The Information article is behind a paywall and was not read directly. Against DigitalToday's detailed summary, it is confirmed that: Anthropic's new model is stronger than existing models in reasoning and intuition and can guess the real intent when the user does not state the goal clearly; the example is that when fixing one bug it recognizes a class of underlying problems and proactively asks whether to fix them together; it is not known whether it belongs to the Fable or Opus line. Google's Carbon is one tier above Argon, and Argon did not surpass the top models from Anthropic and OpenAI, though developers consider Gemini stronger on visual tasks such as video analysis and website inspection. OpenAI's new model is in the mid-tier Terra line, higher performance and cheaper than GPT-6.1 Sol, with a strategy of using price to hit Anthropic. Business Insider also confirmed from internal documents that Carbon is being tested on the internal coding platform Jetski, with an employee saying "coding feels like Opus 5.5," but Google declined to comment and does not guarantee a release. Mr_Salio's claim that "this could be Fable 5.5, launching next week" is personal speculation with no source and cannot be treated as fact.
  • Assessment: the news value here is in the "cadence" and the "positioning," not the specific model names. All three are making differentiation out of "intent understanding" and "price band," which shows that benchmark scores alone can hardly drive migration anymore. All names and dates should defer to official releases; before launch, Carbon and Terra are only internal codenames.

Other

  • Developer traffic bought by "intelligence advantage" is declining: The Information's 10-14 tally (earlier than this window) says a 10-point intelligence advantage now yields only about a 20% increase in developer traffic, versus close to 69% at the end of 2025. Cognition's Russell Kaplan said on 10-19 that this matches internal data, and that many software engineering tasks are already "intelligence-saturated," so once any model can do them well, cost and speed are everything (@russelljkaplan, 108K views, 927 likes). On the OpenRouter side, 84% of tokens go to non-frontier models, which offer about 77% of frontier performance at about 2.5% of the price (Tunguz analysis, relayed by TokenPost and CryptoBriefing); @stretchcloud's long post strings these together into an argument that routing commoditizes models (260 views, 1 like). The Information's original tweet confirms the 20% / 69% numbers; 84% / 2.5% is third-party analysis for a specific week under a specific methodology and should not be extrapolated to the whole market.
  • Artificial Analysis releases AA-PersonalityBench: launched on 10-20 at 02:27, it uses 900 questions (30 personality traits × 30 questions) to have an LLM judge pairwise comparisons, aggregated into 8 indices; a further 21 communication features (sentence length, number of tables, em dashes, emoji and so on) are counted programmatically (@ArtificialAnlys, 29K views, 526 likes). The official page confirms the design of 900 questions, 8 indices and 21 features, and RuntimeWire says the first round covers 36 models. The representative models for each index (Fable 5.1, GLM-5.3, Muse Wonder and others) and specific values such as "Qwen3.8 27B averages about 1,030 words, GPT-6 Luna 270 words" appear only in the post and were not checked item by item on the public page. It suits style selection, not capability ranking.
  • NVIDIA open-sources the Kimodo human motion diffusion model: 700 hours of optical motion capture (the BONES Rigplay dataset), text plus kinematic constraints (full-body keyframes, sparse joints, 2D paths and so on) to generate 3D human and robot motion, with seven checkpoints for SOMA, Unitree G1 and SMPL-X (@HuggingPapers, 10-19 05:09, 32K views, 726 likes). The NVIDIA project page confirms 700 hours, a two-stage denoiser, and that G1 motion can be exported to ProtoMotions / MuJoCo to train policies; the weights went up on Hugging Face and GitHub in mid-October, and the X post is second-hand propagation within the window. There is no third-party benchmark; treat it as a research release.

2. AI compute and industry

Meta discloses Prometheus: a million-GPU target, 85% goodput and a 99% cut in interruptions

  • What happened: Meta disclosed details of Prometheus on 10-20 at 02:14 — a gigawatt-scale supercluster in New Albany, Ohio. The first phase is complete and has trained the Muse family of foundation models (including Muse Wonder); pretraining goodput exceeds 85% and reinforcement learning exceeds 90%, and during Muse Wonder training the mean time between interruptions was about 47K GPU hours, later raised above 54K through firmware fixes. The cluster is currently several hundred thousand GPUs across multiple buildings, targeting about 1 million, and will introduce Vera Rubin NVL72 on the San Miguel platform.
  • Why it matters: this is a public lesson in translating "compute scale" into operational metrics. For very large training, one bad card, switch or link can stall an entire batch of jobs; 85% goodput means about 15% of effective compute is still eaten by failures, recovery and waiting. The 30–40% power transients on the electrical side — tens of megawatts swinging within tens of milliseconds — are problems other teams will hit sooner or later.
  • Key posts: @AIatMeta (10-20 02:14, 147K views, 2,329 likes, 182 reposts, 109 replies, 63 quotes); NVIDIA AI Infra (5,074 views, 121 likes); detail thread @choblin29 (2,307 views, 24 likes).
  • Verification: against the Meta engineering blog, it is confirmed that the first phase is complete and has trained Muse Wonder, the target is about 1 million GPUs, and the current several hundred thousand GPUs span multiple buildings; 85%+ pretraining goodput, 90%+ RL goodput, about 47K GPU hours MTBI (later 54K+); in mid-2025 peaks approached 1,000 interruptions per day per 8,000 GPUs, cut 99% within a year; to keep 32,768 GPUs training continuously, it actually configures 523 NVL72 racks, 37,656 GPUs (about 13% headroom, of which 11% comes from the 64/72 slicing geometry); power transients of 30–40%, with Dynamo Dimmer doing cluster-level power capping alongside rack batteries, capacitor trays and energy storage; the network BAG spans 2–100 km. choblin29's details are basically accurate, but "plans to expand to about 1 million and add Vera Rubin" is information Meta had already made public, not his exclusive; NVIDIA's "cutting interruptions by 99%" matches Meta. A distinction is needed: 47K / 54K is MTBI (mean time between interruptions), not the same metric as "failure rate down 99%" — one looks at the interval, the other at the reduction.
  • Assessment: these numbers show that once scale goes up, reliability engineering is itself part of model capability. Spare-part ratios, fault-domain size and power-transient management all turn directly into training cost and iteration speed. The lesson for peers is not "buy more cards" but to treat goodput and MTBI as first-class metrics from the start; 85% also reminds us that there is a considerable gap between nominal and effective compute.

SemiAnalysis forecasts about 50GW of new data center capacity in 2027

  • What happened: SemiAnalysis published the report "Datacenters 2027: The Year of Acceleration" on 10-19, forecasting about 50GW of new global data center capacity in 2027, close to the cumulative global build before 2024; it also says more than 24GW across 40 campuses is already in late-stage construction, most of it delivering in 2027. Slightly under 1 trillion dollars of binding, non-cancellable leases have been signed year to date in 2026.
  • Why it matters: if the forecast holds, then on the 25GW figure for 2026 cited in the X post, new capacity in 2027 would more than double, directly pulling GPUs, HBM, networking, cooling and power equipment. The harder part is not the forecast but the "24GW already in late-stage construction," an existing stock that satellite imagery can cross-check.
  • Key posts: @MelvinInvests (10-19 21:20, 17K views, 149 likes, 19 reposts, 15 replies, 5 quotes); @l_m_s15 (312 views, 2 likes).
  • Verification: against the SemiAnalysis public summary, the headline is confirmed as "+50GW in a Single Year"; >24GW across 40 campuses in advanced stage of construction, most over half complete with external cooling and electrical equipment installed; slightly under 1 trillion dollars of binding leases year to date in 2026; US hyperscaler self-build starts at a record high; modularization compressing build times to within 10 months. The three precise numbers the X post gives — 51.2GW, 25GW for 2026, 15.6GW for 2025 — cannot be found in the public summary and presumably belong to the paid report body, so they could not be checked item by item. The coverage says the methodology excludes China. Note also that SemiAnalysis itself discusses power, labor, supply chain and financing constraints, and lists "the next bottleneck may shift back to semiconductors and memory" as a risk.
  • Assessment: this is one of the most aggressive capacity forecasts around, but "24GW under construction" is more credible than "51.2GW in 2027," because the former is backed by satellite imagery and construction progress. The 1 trillion dollar lease figure shows demand is locked in, but that does not mean on-time delivery; grid interconnection, labor and modular delivery are the variables that decide whether 2027 materializes. It is good news for the supply chain, and a counterargument to the claim that "compute will immediately oversupply."

The Information: DeepSeek and Huawei co-design chips, trialing Ascend 950DT pretraining by year-end

  • What happened: The Information reported on 10-19 that DeepSeek is working with Huawei to design future AI chips, with Huawei engineers stationed in DeepSeek's office feeding DeepSeek's feedback on compute partitioning and inter-processor data transfer into the design; DeepSeek plans to start using Ascend 950DT for model pretraining as early as the end of 2026. Huawei also plans to allocate most of its next-generation Ascend 960DT in the first quarter of 2027 to DeepSeek as a priority.
  • Why it matters: this moves from "buying cards" to "co-defining chips." Model companies know best where the compute graph, communication patterns and memory bottlenecks are, and getting involved early in chip design can yield more from software-hardware co-design than simply switching suppliers. For domestic compute, the training side has always been the hardest link; this pushes the trial from inference to pretraining.
  • Key posts: Techmeme (10-20 01:05, 4,214 views, 12 likes, 3 reposts, 4 replies, 2 quotes); MilkRoadAI (2,158 views, 11 likes); Newsforce (2,222 views, 2 likes).
  • Verification: the original The Information article is behind a paywall. Against the Investing.com / Yahoo summary, it is confirmed that: Huawei engineers are on site and involved in designing compute and inter-chip data transfer; DeepSeek will pretrain with 950DT as early as the end of 2026; most of the 960DT (expected in the first quarter of 2027) goes to DeepSeek as a priority; DeepSeek has not abandoned NVIDIA and is still seeking its advanced processors; Epoch AI research says NVIDIA B300's raw arithmetic performance is about 7 times Huawei Ascend 950's, and training a frontier model could require more than 10,000 Huawei chips running continuously for weeks. The 950DT specs (144GB HiZQ 2.0, 4TB/s bandwidth, 1 PFLOPS FP8 / 2 PFLOPS FP4) come from earlier reporting and are not new here. A conflict in framing should be flagged: Bloomberg's September report said 160,000 950DT chips would go mainly to Ulanqab in Inner Mongolia for inference, with training still relying on NVIDIA (as relayed); this time The Information says they will be used for pretraining by year-end, which may reflect different batches or different stages of the plan — neither is officially confirmed. MilkRoadAI says "last month DeepSeek open-sourced Ascend programming tools and optimized a supernode"; the tools part is supported by Reuters' September 30 report (a Huawei-compatible version of TileLang), while the supernode part could not be verified.
  • Assessment: the industry implication is that competition in domestic chips is shifting from "can it run" to "customized for a specific model." But the 7x raw compute gap, the system reliability of coordinating more than 10,000 chips in training, and HBM supply are all hard constraints; in the short term NVIDIA's position in training will not be shaken. What is really worth tracking is whether that pretraining run at the end of 2026 finishes and at what cost, not what partnership was signed next.

3. Chinese AI ecosystem

NetEase LobsterAI publishes its second office agent report: bigger tasks, WeChat as the entry point, bills driven by cache

  • What happened: NetEase LobsterAI published its "China Office Agent User Behavior Report" (September edition) on 10-19, the second in the series, with additional statistics for the National Day holiday. The report says average token consumption per task in September was 202K, 3.2 times March's 63K; daily active users grew 83.1% month over month in September, the fastest since May; general programming / IT accounted for 40.1% of tasks and design / film / creative work 17.2%; 77% of users assigning work via IM used WeChat; and nearly 70% of ordinary users used only one model.
  • Why it matters: this is a rare domestic agent report based on real product user behavior rather than a questionnaire. Its direction is very concrete: task volume is growing, the entry point is moving from the desktop to WeChat, and cost differences are driven mainly by the cache-hit price rather than the hit rate. For teams building office agents, that has more decision value than "the model gained a few more points."
  • Key posts: @dotey (10-19 22:48, 14K views, 63 likes, 6 reposts, 32 replies, 1 quote).
  • Verification: the full public text of the report itself was not found, but 10-19 coverage by Science and Technology Daily and Sina Finance confirms the core numbers: DAU +19.8% month over month in June, +25.1% in July, +30.6% in August, +83.1% in September; per-task tokens from 63K to 202K (3.2 times); general programming / IT 40.1%, design / film / creative 17.2%; WeChat 77%; nearly 70% of users on a single model, with heavy users averaging 2.4; the top 5 model combinations all use the DeepSeek Flash series as the base, with "+GLM 5.3 Flash" at 14.0% and "+V4 Pro" at 12.4%; September token share month over month was roughly +166.7% for GLM, -0.3% for DeepSeek, -18.2% for Qwen, -66.7% for MiniMax, -78.6% for Kimi; image generation Seedream 57.2%, video generation Seedance 54.4%. Cache hit rates (DeepSeek V4.1 Flash at 97.3% and so on) and "actual cost as a proportion of list price" (4.6% to 35.1%) appear only in dotey's post and attached images, with no full report text or third-party review found; they are cited as per the post and not treated as independently verified data. DAU is given only as month-over-month growth with no absolute numbers; dotey estimates "September is about 3.6 times May," which the report itself does not say.
  • Assessment: the three things most worth copying from this report are: tasks are getting bigger (token consumption up 3.2 times in 5 months), the entry point is migrating (WeChat at 77%, with work being assigned as early as the morning commute), and cost is set by cache pricing (hit rates are all around 90%, yet bills differ sevenfold). But remember it is a single-product sample, and vendor shares are relative changes, not absolute values; that cache table reflects each vendor's pricing strategy, not model capability.

Two additions from Chinese discussion

  • Price comparison for MiniMax M3.1 Flash: @xiangxiang103 converts the official dollar price into 1 yuan input and 4 yuan output per million tokens and sets it alongside DeepSeek's off-peak price. That conversion roughly matches the official 0.15 / 0.60 dollars, but "exactly the same as DeepSeek's off-peak" is a cross-comparison made at posting time, prices change, and it should be re-checked against current pricing before citing; the OpenCode free access, zero data retention and no training on data mentioned by @MATKA_r are terms-of-use information, while the official post only confirmed the free duration. Model details already covered in section 1 are not repeated.
  • Ming-3.0's "first" needs to be read per leaderboard: ModelScope says Ming-3.0-tts-0.5B took open-weight SOTA on Artificial Analysis, and TeksEdge gives 1,069 Elo. Against AA's two leaderboards, that first place holds only on the Controlled Voice sub-leaderboard (with a unified cloned reference voice); on Provider Voice (each vendor using its own voice) it ranks only third among open weights. The Chinese coverage does not mention two points: the 0.5B weights were uploaded in September and the code was only completed on October 19, and "day-zero support" is still an unmerged PR. Model details already covered in section 1 are not repeated.