52 releases of 49 models between 21 Aug and 22 Sep 2026 — 37 of language models, 15 of open weights — each dated and linked to its primary source: the maker’s own post or page, or, for the one model whose maker is not known, the service that offered it. Above them, the models the public signals point to for the week of 22–28 Sep.
A dated snapshot, not a feed: it does not update by itself. Nothing here is measured. The chart takes one thing from this page: a ring, with its date, around each configuration on the atlas first made available between 21 Aug and 22 Sep. Nothing from the forecast reaches it.
Expected 22–28 SepThe recordReleased since 21 AugHow this was put together
What the makers have teased, what has leaked, and what their calendars say. How likely each one looks is EvalMap’s reading of those signals — the only forecast on this site — and every card links to the posts it rests on.
23–24 Sep (Meta Connect)
@finkd 2 Sep: '🍉 and Muse Spark open weights releases coming soon'; @AIatMeta 2 Sep: 'bigger models, Muse Spark open weights, and more'. The Connect agenda lists the keynote but names no model. @alexandr_wang 20 Sep: 'i am going to be speaking at connect this year! no promises'.
22–28 Sep, no day named; one guess is before the Mid-Autumn Festival on 25 Sep
@MaxForAI 19 Sep: Kimi's official Zhihu account posted digits of π without the leading '3.1'; @notjazii and @kimmonismus read it as a K3.1 teaser, and @Fei2411 expects it before the Mid-Autumn Festival. No K3.1 in Kimi's API docs or on Hugging Face.
22–28 Sep, no signal
Google blogs show no new model since Gemini 3.8 Live (15 Sep), which @bedros_p saw in GCP quotas hours earlier. Gemini 4 Pro: @Lentils80 14 Sep, internal checkpoint 'argon'; @aapakari 13 Sep: 'public release in October'.
No date
@Ronny_MiniMax (MiniMax) 13 Sep replied 'Very Soon!' when asked about a new model. No official teaser found; MiniMax's latest Hugging Face upload is MiniMax-Music3 (7 Aug).
Named by Zuckerberg on 8 Sep as coming ‘soon’; October reported, not this week
Zuckerberg on @alexeheath's Sources podcast, published 8 Sep: Watermelon is a 'significantly more advanced pre-train' coming 'soon' (Heath posted the clip with those words on 9 Sep). The Information (Aug 2026, citing internal documents): release planned for October.
sources.news ↗ x.com ↗ x.com ↗ theinformation.com ↗ the-decoder.com ↗
Named by Elon Musk on 14 Sep: pre-training to finish that week, then RL; no date
@elonmusk 14 Sep: 'Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL', and later that day: 'Grok 4.8 will be a noticeable improvement'. No date; Grok 4.7 came out on 21 Sep.
Officially announced 22 Sep at Apsara: in training, “coming soon”, no date given (named on 26 Aug as the architecture Qwen3.8-Flash previews)
Alibaba Cloud press room, 22 Sep: ‘its next-generation model, Qwen 4, is currently in training’, with Qwen 4.5 and Qwen 5 to scale to 5–10 trillion parameters. The keynote slide that morning, filmed from the hall by @CuiMao: ‘Qwen4 Series Coming Soon’, naming Qwen4-Max, Qwen4-Flash & Qwen4-Plus and Qwen4-27B (the four names appear only on that slide). TechWeb reports that Qwen4 is training on a new architecture. @Alibaba_Qwen had first named it on 26 Aug, calling Qwen3.8-Flash ‘an early preview of the Qwen4 architecture’. No weights, API, price or date yet: nothing on Hugging Face, ModelScope or Model Studio.
Announced 22 Sep at Apsara: ‘set to launch later this year’
Alibaba Cloud press room, 22 Sep: ‘Qwen-Image 3.1, an image-generation model optimized for creative design and e-commerce marketing, is set to launch later this year’, with native transparent-background generation. Qwen-Image-2.1 came out on 20 Sep.
Unveiled 22 Sep at Apsara; not in the API, no date given
Alibaba Cloud press room, 22 Sep: Alibaba ‘unveiled Qwen-Audio-3.1-TTS-Next, a next-generation audio generation model capable of creating complete cinematic soundscapes by blending dialogue and ambient sounds all at once from a text script’. Not on QwenCloud’s model changelog or Model Studio’s TTS list on 22 Sep.
15 Oct, the date StepFun gave on 20 Sep
@StepFun_ai 20 Sep, announcing Step 5 Preview on its API: 'Open weights on Oct 15.' The weights were public for a while on 20 Sep in StepFun’s own Hugging Face repo, which is now private; the official release is still dated 15 Oct.
Named by DeepSeek on 10 Sep, with no date given
@deepseek_ai 10 Sep, launching V4.1-Flash: from 14 Sep all deepseek-v4-pro requests route to V4.1-Flash, 'This will continue until V4.1-Pro launches.' No model card, ID or price; DeepSeek's news page ends at V4.1-Flash (10 Sep).
Not this week: in stealth testing since 18 Sep
@notjazii 18 Sep: 'fable next model entered stealth testing today, so i don’t think it's coming out anytime soon'; @chetaslua 18 Sep: outputs 'routed from fable 5.1'; @testingcatalog 21 Sep: 'The next Fable model was also A/B tested on Claude Code recently'. Nothing from Anthropic.
A forecast nobody keeps is not a forecast. Every weekly reading this page has published is sealed on the day it was made and scored once its week has ended: which named models shipped inside the week, and how many shipped that were never named here at all. A closed week is never rewritten — a correction is a new line, not an edit to an old one. Every change to this site is listed too.
The week of 21–27 September is still open. Its forecast was sealed on 21 Sep and last checked against its sources on 22 Sep: 7 models named for the week, 8 further out. It is scored here once the week is over.
The week of 22–28 September is still open. Its forecast was sealed on 22 Sep: 4 models named for the week, 8 further out. It is scored here once the week is over.
No week has closed yet, so there is nothing to score. The first forecast was made on 21 Sep and will be scored after 27 Sep.
Newest first. The date is the day the maker announced the model or made it available (UTC). “Open weights” means the weights were published that day; weights published later get a row of their own.
Claude Opus 5.5 Anthropic3Don the atlas: Claude Opus 5.5 (max)jeep build by Claude Opus 5.5, made 22 Sep
API, Claude apps, Claude Code, Claude Platform on AWS, Bedrock, Google Cloud, Microsoft Foundry · claude-opus-5-5: 1M context, 128K output, $4/$20 per million input/output tokens with cache reads at $0.20, effort low to max with medium the default. Anthropic puts it at 66.4% on Terminal-Bench 4.0 and says it costs about 40% less to run than Opus 5. Out on the day the forecast named, first as 'Opus 5.2 or 5.5'.
GPT-6 Sol and GPT-6 Luna OpenAI3Don the atlas: GPT-6 Sol (max)jeep build by GPT-6 Sol, made 22 Sepjeep build, max effort by GPT-6 Luna, made 22 Sep
API (Responses and Chat Completions) · gpt-6-sol and gpt-6-luna: text and image in, text out; 1.05M context, 128K output, effort none to max. Sol $2/$10 per million input/output tokens with cached input at $0.20; Luna $0.10/$0.50 with cached input at $0.01, both for prompts up to 272K tokens. The launch OpenAI had put on Tuesday.
developers.openai.com ↗ developers.openai.com ↗ developers.openai.com ↗
Grok 4.7 SpaceXAI3Don the atlas: Grok 4.7 (xhigh)jeep build by Grok 4.7, made 21 Sep
API (grok-4.7), Cursor, and Grok Build, the grok CLI, where it is the default model; the docs name OpenRouter, Vercel and Cloudflare as gateways. A Fast variant (grok-4.7-build-fast), at twice the output speed and twice the price, only in Cursor and Grok Build · A new, larger base model than Grok 4.6's, with a longer reinforcement learning run. 500K context; text and image in, text out; reasoning effort low/medium/high/xhigh, high by default; $2/$6 per million input/output tokens, $4/$12 above 200K; knowledge to May 2026.
MiMo-V2.6-Pro and MiMo-V2.6-Flash Xiaomiopen weights3Don the atlas: MiMo-V2.6-Projeep build by MiMo-V2.6-Pro, made 22 Sep
API on the Xiaomi MiMo platform (mimo-v2.6-pro, mimo-v2.6-flash) and on OpenRouter; open weights (MIT) on Hugging Face and ModelScope as MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL; MiMo Desktop. A Pro-UltraSpeed tier, the same checkpoint served at up to 20 times the output speed for 10 times the price, is API only · Pro: 1.02T parameters, 42B active; Flash: 309B, 15B active. 1M context, 128K output; text, image, video and audio in, text out; thinking on by default and switchable. $0.435/$0.87 per million input/output tokens for Pro and $0.14/$0.28 for Flash, unchanged from V2.5. Out on 21 Sep UTC; Xiaomi’s changelog dates it 22 Sep, Beijing time. Pro also runs in UltraSpeed, an API-only mode at up to 20× the speed, for $4.35/$8.70. Xiaomi also released MiMo-V2.6-Distill-Qwen-9B, a 9B distillation, with open weights under MIT.
mimo.mi.com ↗ huggingface.co ↗ x.com ↗ mimo.mi.com ↗ huggingface.co ↗
Qwen-Image-2.1 Alibaba Qwenimageopen weights
Open weights (Qwen Research License, non-commercial) on Hugging Face and ModelScope · Text-to-image and editing in one model; 7B-parameter generation component. RGBA output; up to 10 reference images.
Step 5 Preview StepFun3Don the atlas: Step 5 Previewjeep build by Step 5 Preview, made 19 Sep
API (StepFun platform) and StepFun Studio; open weights promised for 15 Oct · 600B-total, 27B-active MoE; 1M context; image and video input; reasoning effort low/medium/high; $1.00/$2.70 per million input/output tokens. Artificial Analysis lists it from 18 Sep and measured it before the announcement.
Qwen-Audio-3.1-Realtime-Plus Alibaba Qwenspeech
API (QwenCloud / Model Studio, qwen-audio-3.1-realtime-plus) · Real-time duplex speech conversation, on the same protocol as 3.0 Plus; eight new system voices, 262,144-token context, function calling, web search and voice cloning. Presented at Apsara on 22 Sep as part of the Qwen-Audio-3.1 family.
Qwen3.8-Omni-Flash Alibaba Qwen
API (Qwen Cloud, Alibaba Cloud Model Studio) and Qwen Studio; no open weights · Text, image, audio and video in; text out. Built on Qwen3.8-Flash-Next; 1M context; $0.15/$0.47 per million input/output tokens on Qwen Cloud.
GLM-5.3-FlashX Z.aijeep build by GLM-5.3-FlashX, made 21 Sep
API (not yet in the GLM Coding Plan) · Faster tier of GLM-5.3-Flash (320B total, 18B active), up to 200 tokens/s. 1M context; $0.37/$1.25 per million input/output tokens.
Qwen3.8-LiveTranslate-Flash-Realtime Alibaba Qwenspeech
API (QwenCloud / Model Studio, WebSocket Realtime API) · Simultaneous translation of live audio and video: 60 source languages, voice output in 29; about 2.3 s end to end, down from 2.8 s, with speaker diarization. Alibaba ‘debuted’ it at Apsara on 22 Sep.
Doubao-Seed-2.1-pro 0915 ByteDance Seed
API on Volcengine Ark; also in Doubao Work and TRAE · Dated update to Seed 2.1 pro; Ark model ID doubao-seed-2-1-pro-260915.
Pareto 26.9 Unbiasedjeep build by Pareto 26.9, made 16 Sep
Free stealth API as Union Alpha (OpenRouter, OpenCode, Cloudflare, Cline); paid API after the reveal on 17 Sep · A blend of several open and frontier models, not one set of weights. 256K context; $2.50/$7.50 per million input/output tokens.
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking Google DeepMindspeech
Generally available in the Gemini API and AI Studio; in Gemini Live and Search Live · Audio-to-audio models gemini-3.8-live and gemini-3.8-live-extended-thinking; 97 languages; Enterprise in private preview.
Quasar 1.1 438B Multiverse Computing
CompactifAI API · Compressed from Z.ai's GLM-5.2 (experts cut from 256 to 148 per layer); part of the healing data was made on IBM quantum hardware.
StepAudio 3 StepFunspeech
API and Voice AI Lab · Five models: Realtime (full-duplex voice), ASR, TTS, Gen and Music; Realtime, Gen and Music are free-trial previews.
Kimi K2.8 Preview Moonshot AIjeep build by Kimi K2.8 Preview, made 11 Sep
In Kimi Code, including its kimi-for-coding API (preview, no weights) · Served under the unchanged kimi-for-coding ID; thinking effort low/high/max (default max); up to 1M context on all tiers.
Fugu Max and Fugu Ultra v2 Sakana AIjeep build by Fugu Max, made 21 Sep
API (OpenAI-compatible) · Multi-agent orchestration systems that route tasks across a pool of models; Fugu Max costs $2/$6 per million input/output tokens.
SWE-2 Cognitionjeep build by Devin SWE-2, made 10 Sep
Devin Desktop and CLI; rolling out to Devin Web · Coding model post-trained from Kimi K3 (2.8T parameters), with medium, high and max effort levels.
North Small Translate Cohereopen weights
Open weights (CC BY-NC 4.0), gated download on Hugging Face · Translation model, 218B total / 25B active MoE; 50+ languages; 16K input and 16K output tokens.
DeepSeek-V4.1-Flash DeepSeekopen weights3Don the atlas: DeepSeek V4.1 Flash (max)jeep build by DeepSeek V4.1 Flash, made 8 Sep
API and open weights (MIT) · 552B MoE, 8B active on input / 16B on output; native image input; 1M context; API model ID deepseek-flash; V4.1-Pro to follow.
AuK and AuK-Flash Tencentspeechopen weights
Open weights (MIT) · 1.5B model for speech generation and editing (TTS, voice cloning, enhancement, separation); AuK-Flash is distilled to 4 steps.
Mercury 2.5 Inception
API, chat, OpenRouter and Baseten · Diffusion LLM, 260K context, tunable reasoning. $0.20/$0.75 per million input/output tokens, 80% off at launch.
ChatGPT Images 2.5; GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst OpenAIimage
In ChatGPT and Codex, and in the API · Rolling out to all ChatGPT and Codex users. API: Flare for fast generation, Sunburst for precise edits; GPT Image 2 token rates.
MiMo-X-Pro-Preview and MiMo-X-Flash-Preview Xiaomijeep build by MiMo-X-Pro-Preview, made 9 Sep
Desktop app beta, invite only · Codenamed preview models, free with time and quota limits for approved testers of the invite-only MiMo Desktop beta.
Ling-3.0-flash-VL inclusionAI (Ant Group)open weights3Don the atlas: Ling-3.0-flash-VL
Open weights (MIT), BF16 and FP8; FP4 and INT4 followed on 11 Sep · Weights of the image- and video-input model announced on 4 Sep.
LLaDA-Image and LLaDA-Image-Turbo inclusionAI (Ant Group)imageopen weights
Open weights (Apache-2.0) · 6B model for text-to-image and editing. Base samples in 50 steps, Turbo is distilled to 4; FP8 checkpoints also published.
MAI-Image-2.6-Flash (with MAI-Image-2.6 in Foundry) Microsoft AIimage
Microsoft Foundry public preview and MAI Playground · Faster tier of MAI-Image-2.6 (announced 10 Aug) at under half its price; both models reached Foundry public preview that day.
Omen Alpha Undisclosed (stealth)jeep build by Omen Alpha, made 4 Sep
Paid preview for OpenCode Go subscribers · Maker not disclosed. 500K context, text and image input, $0.20/$0.66 per million input/output tokens. OpenCode marked it deprecated on 10 Sep.
Ling-3.0-flash-VL inclusionAI (Ant Group)3Don the atlas: Ling-3.0-flash-VL
Announced; API trial on Computrix from 5 Sep · 124B total / 5.5B active MoE with image and video input. Weights (MIT) followed on 8 Sep; the model card lists a 256K context.
MAI-Transcribe-2 Microsoft AIspeech
Public preview in Microsoft Foundry (Azure Speech), MAI Playground, OpenRouter · Speech-to-text in 60 languages with diarization and word timestamps; $0.10 per audio hour until end of 2026.
GPT-6 Astra OpenAI3Don the atlas: GPT-6 Astra (max)jeep build by GPT-6 Astra, made 4 Sep
API, Codex, Amazon Bedrock; ChatGPT Plus, Pro, Business and Enterprise over the following days · gpt-6-astra: 1.05M context, 128K output, effort low to max; $10/$50 per million input/output tokens; Fast mode 2x speed at 2x price.
Ling-3.0-flash-Fin inclusionAI (Ant Group)open weights3Don the atlas: Ling-3.0-flash-Fin
Open weights (MIT) on Hugging Face · Weights of the finance-tuned model on OpenRouter since 27 Aug.
Qwen3.8-Max-0902 Alibaba Qwen3Don the atlas: Qwen3.8 Max (0902)jeep build by Qwen3.8 Max 0902, made 2 Sep
API (Qwen Cloud / Model Studio) · Upgraded snapshot of Qwen3.8-Max (alias qwen3.8-max-2026-09-02); 2.4T parameters, 1M context; $2/$6 per million input/output tokens.
Gemini 3.8 Flash and Gemini 3.8 Flash Cyber Google DeepMind3Don the atlas: Gemini 3.8 Flash (high)jeep build by Gemini 3.8 Flash, made 2 Sepjeep build, second build by Gemini 3.8 Flash, made 2 Sep
Gemini API (stable), AI Studio, Antigravity, Gemini app for AI Pro/Ultra; Flash Cyber via Fairwind Program only · gemini-3.8-flash: 1M input, 64K output, thinking low/medium/high; $0.75/$3.75 per million input/output tokens until end of 2026, then $1.50/$7.50.
Muse Spark 1.3 Meta3Don the atlas: Muse Spark 1.3 (max)jeep build by Muse Spark 1.3, made 2 Sepjeep build, max effort by Muse Spark 1.3, made 4 Sep
Meta Model API and Muse Code · Closed weights; 1M context; $1.25/$4.25 per million input/output tokens. The max reasoning level was held back for safety testing and opened on 4 Sep.
Quasar 438B Multiverse Computing3Don the atlas: Quasar 438B (max)
API (CompactifAI) · 438B-parameter reasoning model built from Z.ai's GLM-5.2 with CompactifAI compression; English and Spanish.
Celeris-1 Magnus Celeris
API · Hybrid diffusion model derived from Qwen3.8-27B, with reasoning and tool calling; 131K context; $0.20/$0.70 per million input/output tokens.
Claude Fable 5.1 and Claude Mythos 5.1 Anthropic3Don the atlas: Claude Fable 5.1 (max with fallback)jeep build by Claude Fable 5.1, made 2 Sepjeep build, medium effort by Claude Fable 5.1, made 5 Sep
API, Claude apps, Claude Code, Bedrock, Google Cloud, Microsoft Foundry; Mythos 5.1 trusted access only · claude-fable-5-1: 1M context, 128K output, $10/$50 per million input/output tokens, effort low to max. Mythos 5.1 is the same model for trusted-access partners.
Muse Voice Transcribe Metaspeech
Meta Model API, Meta AI for Mac, Muse Code · Real-time streaming ASR with diarization (20+ speakers) and endpointing; 25 validated languages; $0.18 per audio hour.
DeepSeek-V4-Flash-Vision-Exp DeepSeekopen weights3Don the atlas: DeepSeek V4 Flash Vision (max)
Open weights (MIT) · Weights of the experimental vision variant of V4-Flash, on the DeepSeek API since 21 Aug; the API model was retired 10 Sep.
Hy4 preview Tencentopen weightsjeep build by HY4 Preview, made 28 Aug
Open weights (Apache-2.0); API on OpenRouter (Tencent) · 770B MoE, 49B active; 1M context; reasoning 'high' by default, 'no_think' for direct answers; FP8 weights too.
GLM-5.3 Z.aiopen weights3Don the atlas: GLM-5.3 (max)jeep build by GLM-5.3, made 26 Augjeep build, max effort by GLM-5.3, made 22 Sep
Open weights (GLM-5.3 License) · Weights of the model in Z.ai's API since 18 Aug; reasoning low/high/max; licence requires review for $10B+ MaaS firms.
Ling-3.0-flash-Fin inclusionAI (Ant Group)3Don the atlas: Ling-3.0-flash-Fin
API via OpenRouter · Finance-tuned Ling-3.0-flash: 124B total, 5.1B active, 256K context. Weights followed on 3 Sep.
Gemini Omni 1.1 Flash Google DeepMindvideo
Generally available in the Gemini API (AI Studio), Enterprise Agent Platform and Google Flow · Video generation and editing: scene extension to 40 s, first/last-frame control, 4K upscaling, cheaper 360p drafts.
MiniMax H3 Max falvideo
API on fal · fal's post-trained MiniMax H3 video model; fal says it makes a 5-second clip in under 3 seconds.
Qwen3.8-Flash-Next Alibaba Qwenopen weights3Don the atlas: Qwen3.8-Flash-Next
Open weights (Qwen Community License 1.0); hosted Qwen3.8-Flash API announced · 125B MoE, 6B active, plus 51B n-gram embeddings; 262K context (1M with YaRN); reasoning xhigh/medium/low.
Gemini 3.5 Transcribe and Gemini 3.5 Transcribe Live Google DeepMindspeech
In the Gemini API and Gemini Enterprise Agent Platform · Speech-to-text in 85+ languages: gemini-3.5-transcribe for recorded audio, gemini-3.5-transcribe-live for streaming.
GLM-5.3-Flash Z.aiopen weights3Don the atlas: GLM-5.3-Flashjeep build by GLM-5.3-Flash, made 26 Augjeep build, on two DGX Sparks by GLM-5.3-Flash, made 27 Aug
API, open weights (MIT), Coding Plan and Z.ai chat · 320B total, 18B active; natively multimodal; 1M context; reasoning low/high/max; $0.15/$0.50 per million input/output tokens. Tested as Ox Alpha.
Granite 4.2 IBMopen weights3Don the atlas: Granite 4.2 8B
Open weights (Apache-2.0); watsonx and partner APIs · Dense 3B, 8B and 30B models; thinking, non-thinking and low-effort modes; 128K context, extendable to 512K.
Ling-3.0-flash-dspark inclusionAI (Ant Group)open weights
Open weights (Hugging Face) · 1.36B-parameter DSpark draft model for speculative decoding with Ling-3.0-flash; runs in SGLang and llama.cpp.
DeepSeek-V4-Flash-Vision-Exp DeepSeek3Don the atlas: DeepSeek V4 Flash Vision (max)
API (experimental) · Experimental image-input version of V4-Flash; model ID deepseek-v4-flash-vision-exp. Weights followed on 31 Aug (MIT).
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.
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.