Where it comes from. The rows were first gathered from posts on X and from the makers’ own pages with the grok CLI on 21 September 2026. Each was then checked again against its primary source — the maker’s post, blog, docs or model card, fetched directly — and corrected where the source said otherwise: 11 of the 46 rows then gathered changed in that pass, most often a date, a licence, or whether a model shipped as a preview or generally. Dates are in UTC, and the date of a post on X is also read from its ID. The expectations were checked the same way: every quoted post says what its card says it does, on the day it says. On 4 October a source check covered new evidence of 3–4 October and earlier omissions. Releases were then listed through 4 October; that forecast was a fresh prospective reading for 4–10 October, not a revision of the 1–7 October record. Its weekly long shots were reassessed: no new model-specific day was confirmed, and Kimi K3.1’s reported before-October window has expired. Every earlier ledger row and the numerical probabilities sealed on 29 September remain unchanged. Seven native Grok OAuth topic runs made 57 search queries and 28 page-open calls, followed by independent primary-source checks. Dedicated X search was unavailable, some pages were blocked or partial, and the coverage is not exhaustive; selected Hugging Face catalog slices do not establish absence everywhere. Publication dates without an explicit timezone retain the source’s calendar date, including Kolibri’s maker-stated 3 October release. Only named speech/API models are added; independent trained weights of Microsoft’s Streaming and Voice Flash variants are not established. Targeted update on 6 October: Mistral’s own announcement and model documentation confirm Mistral Large 4 public preview API access, with weights promised for the end of October rather than released. The release cutoff advances to 6 October, but the weekly forecast and its 4 October source-check day remain unchanged; this is not a fresh full-model search. Targeted check on 7 October: Google’s dated changelog and model card document Nano Banana 2.1 as an updated image model released to GA in the Gemini API on 6 October. Earliest public-access chronology and independent trained-checkpoint identity were not established. The release cutoff, future cards, sealed forecast and its 4 October source-check day remain unchanged. Targeted check on 7 October: Anthropic’s dated announcement documents Claude Haiku 5.5 available on its API and cloud platforms that day; availability was documented, not independently exercised. The release cutoff advances to 7 October; the weekly forecast and its 4 October source-check day otherwise remain unchanged. A targeted 8 October check added OpenAI’s 7 October October-version GPT-6 Sol and Luna in ChatGPT from the maker’s dated post; the forecast was not reassessed. Grok passes over 8–10 October found no new release and no new maker date. Source check: 10 October 2026. Releases are listed through 10 October; the forecast is a fresh prospective reading for 8–14 October, not a revision of the 4–10 October record, which stays open in the ledger. Seven grok topic runs on 6–8 October evidence; no maker named a day inside 8–14 October, and Step 5’s weights remain dated 15 October. On 13 October a pass over 12–13 October checked each new release against its maker’s page — Meta’s newsroom, Google’s developer blog and Gemini API model page, OpenBMB’s GitHub and Hugging Face, and Liquid’s blog; the Watermelon codename comes from CNBC, not Meta, and Argon on Vertex AI was not verified. Source check: 13 October 2026. On 15 October seven grok topic runs covered 14–15 October and each find was checked directly: the Hugging Face API for the Step 5 Preview and Ling-3.1-flash repositories (rechecked 15 October 09:30 UTC), xAI’s post and release notes for Grok Voice 3.0, and the models.dev, OpenRouter and Vercel catalogues for the other forecast models. Ling-3.1-flash is dated by its repository’s creation time; when it became public was not established. Source check: 15 October 2026. On 16 October eight grok topic runs covered 13–16 October, all ending normally with no warnings, and each find was checked directly: Meta’s research blog and Hugging Face repository for Muse Spark Open, OpenAI’s GPT-Live-1.1 system card, the makers’ posts for Kling 4.0, Krea’s and Sarvam’s blogs, NVIDIA’s blog, xAI’s release notes, the Hugging Face organisations of the expected makers and LiteLLM’s pull requests for the Azure Fable 5.5 ID. Microsoft’s pre-training post could not be fetched. Source check: 16 October 2026. On 18 October Beam 1.0 was checked on Reflection’s blog, model card and developer docs and through the Hugging Face API, as the Hugging Face pages could not be fetched; no maker price was found. Source check: 18 October 2026. On 19 October grok topic runs covered 12–19 October, all ending normally with no warnings, and each find was checked directly: Z.ai’s blog for GLM-Legal, FastVideo’s Nuva Lab page and post for FastH3-Omni, IBM’s Hugging Face repositories and blog for Granite Switch 4.1, Black Forest Labs’ blog and Hugging Face repository for FLUX 3 Image Klein, Tencent’s Hugging Face repository and Hy Memory site, NVIDIA’s Hugging Face repository for Nemotron Labs Dory, and DeepSeek’s API update log and Hugging Face repository for V4.1-Flash-1017. Ling-3.1-flash is now dated by Ant Ling’s 14 October post. Source check: 19 October 2026. On 21 October each open card was read again against its maker’s page or Hugging Face organisation, and the FastH3-Omni weights were checked through FastVideo’s post and the Hugging Face API. Source check: 21 October 2026. On 22 October each open card was read again against its maker’s page or Hugging Face organisation, Flint was checked through Recursive’s post and the Hugging Face API, Voxtral 3 through Mistral’s announcement and docs, and the M3.1 Flash weights card rests on MiniMax’s 19 Oct blog and release notes. Source check: 22 October 2026. On 23 October each open card was read again against its maker’s page or Hugging Face organisation; the M3.1 Flash weights were checked through MiniMax’s post and the Hugging Face API, Ling-3.0-flash-Sante and Spark-X2.6-4B through their Hugging Face repositories and the makers’ posts, Spark-X2.6 through iFlytek’s model docs, Canva Craft through Canva’s API reference and post, Eleven Realtime v1 through ElevenLabs’ blog, Grok Imagine Video 1.5 Lite through xAI’s model docs, and the Kimi K3.1 teaser frame by frame. Source check: 23 October 2026. On 25 October each open card was read again against its maker’s page or Hugging Face organisation; Bolt E4B was checked through NVIDIA’s research page, NIM reference and the Hugging Face API, Kimi K3.1 Pro through Moonshot’s blog, model list, pricing page and Kimi Code’s what’s-new page, Hy-Nomas and the Atria Dawn weights and preview through their Hugging Face and GitHub repositories and the makers’ posts, Ideogram Vector and FLUX 3 Max through the makers’ posts and model pages, Glyph Cluster through Vercel’s AI Gateway, Nova 2.5 Sonic through AWS’s announcement, and the Mistral Large 4 date through Mistral’s Hugging Face page and its 9 Oct archive. Source check: 25 October 2026. On 27 October each open card was read again against its maker’s page or Hugging Face organisation; GPT-6.1 Terra was checked through OpenAI’s API changelog, model page and Codex changelog and OpenAI’s post, read in the Wayback Machine because openai.com refused direct reads, Ling-3.1-flash-1019, Hy-SimulMT and Nanbeige4.3-33B through the Hugging Face API, their model cards and the makers’ posts, and Hy-SimulMT’s licence through its licence file and GitHub repository. Source check: 27 October 2026. On 28 October each open card was read again against its maker’s page or Hugging Face organisation, Gemini 4 Carbon against Google’s announcement and the Gemini API changelog and models page, and Genie 4 against Google DeepMind’s post. LFM3 was checked through Liquid AI’s blog, the Hugging Face API and its model cards, Nemotron 3.5 Transcribe through NVIDIA’s blog, the Hugging Face API and its model card, MAI-Image-2.7-Preview through Microsoft AI’s page, Qwen-Audio-3.1-Realtime-Flash through Alibaba Cloud’s Model Studio pages and the QwenCloud changelog, and AssemblyAI-TTS-1 and Prism through their makers’ pages and posts, with the Wayback Machine for the date of AssemblyAI’s blog. Source check: 28 October 2026. On 29 October each open card was read again against its maker’s page or Hugging Face organisation, the Mistral Large 4 weights against Mistral’s Hugging Face page at 00:26 UTC, and Gemini 4 Carbon and Genie 4 against Google’s pages and Sundar Pichai’s third-quarter earnings remarks of 28 Oct. Particle-1 was checked through Particle’s blog and docs and its founder’s post, Clover-35B-A3B-Preview and Maya 2 Native through the Hugging Face API, their model cards and the makers’ posts, Coda v2 through Rime’s page, its model and on-premises docs and its post, and Cadence’s Maestro announcement through the Wayback Machine. The week of 22–28 October was scored with tools/ledger.py close and its matches read by hand: Nemotron 3.5 Bolt E4B is matched to its own row rather than to Nemotron 3.5 Transcribe, and the Kimi K3.1 Pro weights card to none, as the weights are not out. Source check: 29 October 2026, and again in a second pass to 08:52 UTC that day, which read each open card against its maker’s page or Hugging Face organisation, the Mistral Large 4 weights against Mistral’s Hugging Face page and the Hugging Face API at 08:51 UTC, the DeepSeek-V5 identifiers in DeepSeek’s deepseek-harness repository and its git history, GPT Image 2.6 in OpenAI’s API reference, API changelog and models page and in the Internet Archive’s copy of the reference of 23 Oct, and each new lead against its maker’s page, post or Hugging Face repository. The pass of 5 November, to 11:30 UTC, read the Hugging Face, OpenRouter and Vercel catalogues, makers’ X posts and changelogs, then each open card against its maker’s page or Hugging Face organisation and each new lead against its maker’s page, post or Hugging Face repository. Source check: 5 November 2026. The pass of 7 November, from 11:30 UTC on 5 November to 09:23 UTC on 7 November, read the Hugging Face, OpenRouter and Vercel catalogues, makers’ X posts and changelogs, then each open card against its maker’s page or Hugging Face organisation and each new lead against its maker’s page, post or Hugging Face repository; a second grok pass over the same window added no release. Source check: 7 November 2026. The pass of 8 November, from 09:23 UTC on 7 November to 11:02 UTC on 8 November, read the Hugging Face, OpenRouter and Vercel catalogues, makers’ X posts and changelogs, then each open card against its maker’s page or Hugging Face organisation and each new lead against its maker’s page or post; makers’ pages that refused the request were read in the Internet Archive’s latest copy; a second grok pass over the same window added no release. Source check: 8 November 2026. The pass of 9 November, from 11:02 UTC on 8 November to 10:08 UTC on 9 November, read the Hugging Face, OpenRouter and Vercel catalogues, makers’ X posts and changelogs; a script opened the maker’s page for each row grok gave and kept a row only where that page carried the model’s name and date, and every other row was checked by hand against its maker’s page, post or pull request; the six cards the last pass left open questions on were read against their makers’ pages or Hugging Face organisations; a second grok pass over the rows that did not match and those questions added no release. Source check: 9 November 2026.
What is left out. Price changes, partner listings, apps and agents built on existing models, and papers are not releases. Neither is the general release of a model that was already out in preview — GPT-Live-1 reaching the API, GPT-Rosalind, Apple’s AFM 3 — nor a new setting of a model already out, such as the max reasoning level Meta opened for Muse Spark 1.3 on 4 Sep, nor GPT-6 Astra Law, which OpenAI describes as GPT-6 Astra with tools and instructions. Two rows are systems of several models sold as one model — Sakana’s Fugu and Unbiased’s Pareto 26.9 — and are listed because that is how their makers offer them. One row nothing public could confirm, a limited DeepSeek beta announced only in a user group, was left off. Small fine-tunes are not tracked, so the list is wide but not exhaustive. Also left off on 23 September: SenseTime’s SenseNova U1 Pro, whose ‘official version’ of 21 Sep follows an invite-only preview of 19 Jul; iFlytek’s Spark-ASR-2.0, announced on 23 Sep in the press only, with the API to open on 24 Sep; and Cohere’s Command A+, added to OpenRouter on 22 Sep but released in May. From 23 to 25 Sep: Gemini 3.8 Live with Live Avatar reached GA in Gemini Enterprise, a feature on an existing model; NV-Reason-CT is a fine-tune of Qwen3.5-4B; SenseNova’s U1.5 8-step LoRA tweaks an existing model. Recraft V4.1 Flash (image, 23 Sep), Ming Image 0.1 Design Layer (inclusionAI, 23 Sep) and Fish Audio Transcribe 1 Pro (24 Sep) are new on OpenRouter, but no page of their makers dates them, so they wait for one. iFlytek’s Spark-ASR-2.0 (23 Sep) is reported by IT之家 and 36Kr but has no iFlytek page yet. From 25 to 26 Sep: Respan’s Span-01 and Span-01 Lite, new on OpenRouter, score agent behaviour and are classifiers, not chat models; Black Forest Labs’ FLUX 3 Action (September) is a robot-control world-action model, a kind the list does not track; Gemini 3.5 Transcribe reached OpenRouter on 25 Sep but was released on 26 Aug; GitHub Copilot’s listing of Opus 5.5, GPT-6 Sol and Luna and Grok 4.7 on 25 Sep is distribution. From 28 to 29 Sep: H Company’s Holo4-27B and Holo4-35B-A3B are computer-use post-trains of Qwen3.8-27B and Qwen3.6-35B-A3B, and AutoTrust’s JEV-27B is a classifier fine-tune of Qwen3.8-27B; Grok 4.7 on Amazon Bedrock, Sonnet 5.5 in Google’s Model Garden and Solar Mini 4 on OpenRouter are distribution; Manus 2.0 is an agent product. Reddit threads that only restate an X post or an article, and a DevDay ‘o’ agent named only in commentary, are not rows. From 30 Sep to 1 Oct: Inception’s Mercury Decide is a decision model and Voyage’s rerank-3 and rerank-3-lite are rerankers; Runway’s Praxis-1 is a world-action model announced for later; ByteDance’s reported Xiaodou assistant is an app. Checked on 3 Oct: Alibaba Cloud lists qwen-image-2.1-pro for International service in Singapore on 2 Oct, but a distinct trained Pro checkpoint or new Pro weights were not established; Qwen-Image-2.1 weights came out on 20 Sep, and this is not Qwen-Image 3.1. OpenAI’s Dots is an agent product powered by GPT-6 Astra, not a separate checkpoint; Finances and the GPT-6 guide are product/docs updates. MiniMax’s 1–7 Oct unlimited M3.1-Flash-Preview use is a promotion for MiniMax Code subscribers, not a new model or API release. Xiaomi’s V2.5 sunset on 21 Oct and xAI’s Transcribe-1.0 end of life on 2 Oct are lifecycle events, not evidence of MiMo-V3 or Grok 4.8 launch dates. Meta’s 2 Oct research and scaling-framework articles do not announce a new checkpoint or release day. Checked on 4 Oct: Cohere’s Embed 5 Pro and Fast are embedding models released on 30 Sep; Parse 5 is a document-parsing VLM released on 27 Aug. Their 8 Oct X event and unnamed behind-the-scenes teaser are not a promised new checkpoint launch. Ai2’s AstaBrief 8B, announced on 2 Oct, is a specialized Qwen3-8B scientific-report SFT/DPO post-train; its repository creation and modification dates do not establish first public weights. Amazon Strands and Inception Decider are framework/orchestration products, not confirmed new checkpoints. On 7 October NVIDIA published Nemotron 3 specialist checkpoints (Nemotron-3-Ultra-CC and IMO SFT/RL checkpoints), fine-tunes of an existing family; Liquid’s d1 and Perplexity’s pplx-embed-v2 that day were seen only on an aggregator and are not listed until a maker page is checked. GPT-6 Sol and Luna (October) are listed despite sharing names with the September models because OpenAI’s system card treats them as separately released versions with their own evaluations, not a setting of the September models. Checked on 16 Oct: Tencent’s Hy-Memory-30B-A3B is a memory-writer model for memory frameworks, Liquid’s d1-ultra a decision model and Mistral’s Timestral a family of time-series forecasters, kinds the list does not track; Microsoft’s MAI-Base-2 is a pre-trained base model announced without access; FastH3-Omni, a FastVideo post-train of MiniMax’s H3, is in an invitation-only preview on Reactor until 20 Oct; TypeSafe’s Jev 1.14 and ScaleDown’s task models are decision and extraction models; Moondream 3.1 reached OpenRouter on 15 Oct but came out in July; Arena leaderboard placements are not releases; GPT-Live-1.1 in the API is only ‘coming soon’. Checked on 19 Oct: Hy-Memory-30B-A3B is now listed after all, as an open-weight language model released with Hy Memory 2.0, and FastH3-Omni, whose preview API opened to developers on 13 Oct, is listed with its weights forecast. Gemini 4 Argon Cyber Permissive is allowlist-only through Fairwind, a configuration of Gemini 4 Argon; Krea 3 leaving beta is a general release of a model already out; MAI-Base-2 is only the end of pre-training, Kimodo v1.2 a checkpoint fix and Decision 3.0 (11 Oct) outside the window; the anonymous aurora-*-alpha models on Arena and Aspen, Cypress and Maple have no named maker; the Azure Fable 5.5 line was withdrawn and its source cannot be opened. Checked on 22 Oct: MiniMax M3.1 Flash on the API (19 Oct) is the general release of the preview listed on 27 Sep; Claude Duet (21 Oct) is an app built on Opus 5.5; OpenAI’s models in Google Cloud Model Garden (21 Oct) are distribution. Checked on 23 Oct: OpenAI’s Decisions API reaching general availability with gpt-6-luna (22 Oct) is a feature on a released model; Claude on Google Cloud managed by Anthropic is distribution; Microsoft’s MAI-DxO orchestrates other makers’ models; Ai2’s OlmoEarth v1.3 and NVIDIA’s Kumo, Nemotron Retriever, StormScope EU and Minerva are satellite-embedding, tabular, retrieval, weather and monitoring models, kinds the list does not track, as is NVIDIA’s Alpamayo 2 Edge, a driving model due later this month; Tencent’s WeVisOne-VL is a perception post-train of Qwen3-VL; Kling 4.0 in Luma, OpenArt and Picsart and PixVerse V6 in QwenCloud are partner listings; Ling-3.0-tiny-UBTree and Liquid’s LFM2.5-24B-A2B-DSpark are speculative-decoding drafters; Ming-3.0-audio-flash and Ming-3.0-tts-tiny are updates of September repositories; Gradium’s Phonon (weights posted 20 Oct) is a small text-to-speech family built on Kyutai’s Pocket TTS. ByteDance’s Seedance 2.5 4K is in a trial for some enterprise customers, reported only by IT之家, and iFlytek’s SparkAudio 1.0 only by Pandaily. Inception’s Mercury 3, Elorian’s Glimpse and Xingchen’s Xing4.0-29B-A4B are new on OpenRouter (16–17 Oct), but no page of their makers was checked. The anonymous Kibitz and straw models on Arena have no named maker. Checked on 25 Oct: OpenAI’s ‘upcoming GPT‑6 Astra Cyber model’ (23 Oct) has no date and may be a version of GPT‑6 Astra, so it is noted on OpenAI’s card rather than given one; Microsoft’s MAI-Code-1.1-Flash, whose page was archived on 11 Aug, gained a local download on 7 Oct, a new way to run a model out before the window; Atria Dawn on Shanghai AI Laboratory’s API (15 Oct) is the general release of Atria Dawn Preview, listed with its weights; InternScience’s Sage-A1.5 is a scientific-agent post-train of Qwen3.6-35B-A3B; TII’s Falcon-Emirati, Falcon-ASR and Falcon-OCR-Arabic (6 Oct) had no public access that could be found, and Falcon-Emirati ‘will be available’ through Falcon Chat; Hy-Nomas-Diffusers is a format conversion; Kling 4.0 on fal and Kimi K3.1 Pro in Vercel’s AI Gateway are partner listings; Sakana’s Aquarium is an agent platform; the anonymous Kalmia on Arena has no named maker; Qwen4 names in third-party kernel code are not a release. Bria’s Fibo 2.0, Underdog’s Bark 3.0, JetBrains’ Mellum2.1 Thinking, Tencent’s Youtu-LLM-A7B-Base, Seedream 5.0 Flash and Doubao 2.1 Lite were seen only on aggregators, and Block’s Dessa only in a digest, and no page of their makers was checked. Checked on 27 Oct: GPT-6.1 Terra Pro on OpenRouter is GPT-6.1 Terra served with a pro reasoning mode, a setting, and GPT-6.1 Terra (Fast) in Vercel’s AI Gateway and Terra on Amazon Bedrock are partner listings; Baidu’s Unlimited OCR 2 (27 Oct) is a document-parsing model, like Parse 5; inclusionAI’s LBS-27B is a narrow agentic post-train of Qwen3.8-27B, and its two singprobe checkpoints are probes, not models; Microsoft’s CogACT-v2-RoboDojo is a robot policy, a kind the list does not track; Kling 4.0 on Runway, Runware and other partners is a set of partner listings of a model listed from 15 Oct, and a reported ‘full launch’ was not confirmed; MiniMax-M3.1-Flash post-training on Together and its GGUF quantisations are not new models; Runway’s App Studio is an app and fal’s FLUX 3 Image Klein LoRA trainer a feature. ByteDance Dreamina’s DreamScript, Anycook and DreamShot and Cadence’s chip-design models were seen only in news reports, and no page of their makers was checked. Checked on 28 Oct: Liquid’s d1-2B and d1-Omni-2B are decision models, like d1-ultra (16 Oct); OpenBMB’s MiniCPM-OCR 1.0 is a document-parsing model, like Parse 5 and Unlimited OCR 2, and MiniCPM-OCR Jev a model that checks OCR output; Microsoft’s MAI Cyber 1.1 Flash is available only inside MDASH, its vulnerability-scanning harness, to verified customers, and MAI-Image-2.6 and 2.6-Flash reaching general availability in Foundry is the general release of models already listed; ElevenLabs’ Cue (28 Oct) is the forced-alignment model already behind every Scribe transcription and its Forced Alignment API, offered since July 2025; Gemini Robotics ER 2 in public preview on Google Cloud’s Gemini platform is a new surface for a robotics model Google introduced earlier; NVIDIA’s Metatron-v1 is a humanoid robot policy and Lightricks’ LTX-2-Marking a set of watermarking weights for LTX-2, kinds the list does not track; BAAI’s AREX-2, a deep-research post-train of Qwen3.8-27B, has had public weights since 29 Sep; Tripo-ImageToSplat in Model Studio, DeepSeek V4.1 Flash and GLM-5.3-Flash on Mistral’s cloud, Kling 4.0 on Pika, VEED Dubbing on fal and Grok 4.7 reaching general availability on Vertex AI are partner listings, and qwen3.8-flash in the Frankfurt and Hong Kong regions is a regional expansion; Mistral’s Responses API, Claude Duet, Mirendil’s Lor, Runway Ads, Cartesia’s Managed Agents and xAI’s Grok Bot are products and features, and Artificial Analysis’s Intelligence Index v5 is a benchmark; Mirendil-Base, the pretrained model of 21 Oct that Fern is built on, has no public access. Particle AI’s Particle-1, a post-train of DeepSeek V4.1 Flash, is new on OpenRouter and Vercel’s AI Gateway, but no page of its maker was checked; Cadence’s own post of 26 Oct, before this pass, names a Maestro AI Model Portfolio of engineering-design models, but cadence.com still refused reads; and Dreamina’s own feed says nothing of DreamScript, Anycook or DreamShot. These remain unverified. Checked on 29 Oct: Particle AI’s Particle-1 is now listed, from its maker’s blog and docs. Cadence’s Maestro AI Model Portfolio (26 Oct), read through the Wayback Machine because cadence.com still refused reads, says ‘The first Maestro models are available to select early access customers in the fourth quarter of 2026’, so it is not a release. ByteDance’s Jimeng short-drama studio, which IT之家 reported on 26 Oct with in-house DreamScript, Anycook and DreamShot models, is live on Jimeng’s site, but no Jimeng, Dreamina or ByteDance page names the three models and Dreamina’s own feed is silent; they remain unverified. GitHub’s Copilot Fusion (28 Oct) is ‘an intelligent runtime model router and orchestrator’ over Anthropic’s and OpenAI’s models, like MAI-DxO; ‘DeepSeek V4.1 Flash Cyber’ on LLM Gateway is an abliterated DeepSeek V4.1 Flash 1017 served by inference.net on an enterprise plan, not a DeepSeek model; SenseTime’s SenseNova 6.8 Code2Video (25 Oct) is a narrow code-to-video fine-tune of Qwen3.8-27B; Hugging Face’s Carbon-X (22 Oct) is a DNA model, a kind the list does not track; Moss-Image-Edit is an individual’s LoRA; IBM’s granite-speech-5.0-2b-ONNX is a format conversion, and the Nemotron Labs Dory 1.0 4B repositories changed only their generation config; OpenAI’s texting your dot and API credits for ChatGPT Pro and Business, Google’s Connected Apps and XREAL glasses, and xAI’s moderation guidelines in its API are features, products and hardware. A 29 Oct date for Qwen 4 from an account that is not Qwen’s, and talk of a Gemini model beyond Carbon, are rumours. IBM’s Granite Speech 5.0 2B (20 Oct) and KRAFTON’s Raon-1-mini-preview (21 Oct), seen in this pass, fall in the closed week of 15–21 October and were not checked. Checked in the second pass of 29 Oct: Tencent said at its Global Digital Ecosystem Summit on 29 Oct that Hy4, with visual understanding for the first time, will be released soon and Hy4 flash in the near future, but only press reports carry this and no Tencent page or post does yet, so it is not yet listed as an announcement. The China Meteorological Administration’s first Fengyun satellite AI models, for typhoon cloud-image interpretation and severe-convection warnings (29 Oct), are weather models, a kind the list does not track. inclusionAI’s Ling-3.0-flash-UBTree, announced with Ling-3.0-tiny-UBTree on 29 Oct, is also a speculative-decoding drafter; OpenBMB’s GPTQ, AWQ and DSpark repositories of MiniCPM-OCR 1.0 are quantisations and a drafter, and FlagOS bringing MiniCPM-OCR to ten more chip platforms is a port; Efficient-Large-Model’s Open-H3-Regenerate-2K (26 Oct) is a LoRA refinement pass for MiniMax-H3; and Microsoft’s rho-yam-box and rho-ur-ai-trainer repositories of 26 Aug only changed format. Kling 4.0 in Cloudflare’s AI Gateway is a partner listing; fal’s Motion to Code, which has MiniMax H3 Max animate a prompt or image and rebuilds the clip as GSAP or GLSL code, is a pipeline over existing models; precise editing in Ideogram 4.5 is a feature; and Apple Intelligence with Alibaba’s Qwen in China, reported by Leiphone, is a product. Fish Audio’s Drama 3 (13 Oct), an open-weight text-to-speech model whose weights were uploaded again on 29 Oct, and Articul8’s Bivia (28 Oct), a decision model, seen in this pass, fall in the closed weeks of 8–14 and 22–28 October and were not added. Checked on 3 Nov, for LongCat-2.5 only: it is the general release of LongCat-2.5-Preview (25 Sep) and, against the rule above, has its own row, because Meituan launched it as a new model with its own technical post and benchmarks. Nothing else from 30 October to 3 November was checked. Checked on 4 Nov, for Hy4 only: it is the general release of Hy4 preview (28 Aug) and, like LongCat-2.5 and against the rule above, has its own row, because Tencent launched it as a new model with vision and computer use added. Nothing else from 30 October to 4 November was checked. Checked from 29 Oct to 5 Nov, to 11:30 UTC: GPT-6 Rosalind (OpenAI, 30 Oct) is trusted access for life-science researchers to the same weights as GPT-6 Astra, per its system card, so it is not a new model; Kling 4.0 Flash reaching all users and Kling’s API on 4 Nov is the wider release of the model listed here from 28 Sep; Pika H3 Flex (29 Oct) is a LoRA adapter on MiniMax H3, per its model card; Deep Principle’s Hydrogen-1.0 is a scientific post-train of GLM-5.3, inclusionAI’s ZwZ2 an RL post-train of existing vision-language models, and Microsoft’s Mage-VL-Cowork a computer-use fine-tune of Mage-VL; Sarvam Prism (3 Nov) is a decision model built on Sarvam 105B, and ElevenLabs’ SFX v3 (29 Oct) makes sound effects, kinds this list does not track; Tencent Hunyuan’s HyInv and HyPLM-Flash (2 Nov) are protein models, and the HyCell and HyPDE models its Hy4 post calls open-sourced could not be found; ByteDance’s doubao-seed-evolving update on Volcengine Ark (5 Nov) keeps an existing model ID; Baidu’s Apollo World (shown 3 Nov) is a driving-simulation world model with no weights or API, and Mistral’s unnamed cyber variant of Large 4 (29 Oct) is a preview for chosen partners only; reports of Mythos 5.5 in Anthropic’s site code (1 Nov), GLM 5.5 in post-training (31 Oct, denied without a source from Z.ai) and a public Grok 5 API tier (31 Oct) have no source from the maker. NVIDIA’s NVFP4 builds of models already listed are new formats of released weights. Checked from 5 to 7 Nov, to 09:23 UTC: Inception’s Mercury 3 is now listed after all, from Inception’s post of 19 Oct; OpenAI’s GPT‑6 Cyber (5 Nov) is a cyber post-train of GPT‑6 Astra available only to approved Daybreak Red users, like Gemini 4 Argon Cyber Permissive and MAI Cyber 1.1 Flash; the specialised Muse models Meta described on 6 Nov have no names; ByteDance Seed’s Cola2 (Hugging Face, 6 Nov) is a research base model with no licence and no post from its maker; Nous’s Arete-27B is a fine-tune of Qwen3.8-27B and Sage-SciPerception a research projector; Index-Translate-omni (2B, 9B and 35B) is a translation fine-tune posted on 4 Nov; NVIDIA’s Nemotron 3.5 Parse DFlash2 is a speculative-decoding drafter dated 4 Nov, and its Cosmos3-Nano checkpoints are robotics and driving checkpoints of an existing family; Runway’s ‘Real-Time Video at Batch Size One’ (5 Nov) is an engineering post on serving Characters, and Microsoft’s Proof-Driven Development Kit is a toolkit. Reports of GPT-6.2 (including ‘gpt-6.2-betelgeuse’), Fable 5.5, Mythos 5.5, GLM-5.5, Doubao 3, Step 6 and a 10 Nov date for Qwen4 have no source from the maker. Checked from 7 to 8 Nov, to 11:02 UTC: Qoduck (7 Nov) is now listed after all, as a stealth model whose maker is not disclosed; testers’ claim that it is Qwen4 rests on a tokenizer match. Tencent’s Qwen3.8-27B-DSpine-b16 (Hugging Face, 7 Nov) is a speculative-decoding drafter; HySparse2 is a paper of 22 Sep; Liquid’s NVFP4 builds of LFM3 were posted on 24 Oct; the heading ‘Hy4 Flash Recently Released’ on Tencent’s Hy4 post sits over text that still calls it ‘scheduled for release soon’. Reports of a Gemini 4 ‘Hydrogen’, of claude-donut-eap, Silicon and Obsidian 3, of GPT-6.2 Astra, Sol and Bell and a delay to them, of GLM-5.5 sizes (called fake by a Z.ai researcher), of Kardashev-0.7 having shipped and of a 10 Nov date for Qwen4 have no source from the maker. Checked from 8 to 9 Nov, to 10:08 UTC: posts that GPT-6.2 or GPT-6.2 Sol is out or being served still have no OpenAI source, and OpenAI’s teaser of 8 Nov was for a ModRetro Chromatic console; Gemini for Google, seen by testers on 8 Nov, is the name of an enterprise agent, not a model; Cursor’s Composer 3 is still not in Cursor’s changelog or model docs; Efficient-Large-Model’s Open-H3-Regenerate-2K-IC (9 Nov) is another LoRA for MiniMax-H3; Qoder still calls Qoduck a stealth model of undisclosed origin, so posts calling it Qwen 4 remain guesses.