Coming models · checked 23 October 2026
Coming models
Models a lab has announced, teased or shown in a closed preview that nobody can download or call yet, with the day each was announced and when it's expected. We list a model when it comes out, so this is what we're watching. 13 waiting.
Not out yet
| Model | Announced | Expected | Status | What's known |
|---|---|---|---|---|
| Kimi K3.1Moonshot AI · language model · unknown | teaser video on Moonshot's @Kimi_Moonshot X account (cited via pages that describe it) | 27 October 2026a tester on X says before 27 October; Moonshot gave no date | Teased | Successor to Kimi K3; the 24-second teaser shows the Kimi Code terminal running 'Model: Kimi K3.1' and generating Three.js 3D scenes. Believed to be the anonymous 'Kibitz' model on Arena; size, licence and date not stated (a leaker, not Moonshot, says it ships before 27 October). 1 2 |
| Rho-1Reka · multimodal model · unknown | Reka blog post 'Rho-1: Collapsing the multimodal stack' | No date given | Closed preview | 19B omni-reasoning model trained from scratch that reads and generates text, images and video and outputs robot actions in one network. Research preview only: no public API, download or price; collaborators contact Reka. 1 2 |
| GriffinTavus · video model · unknown | Tavus research post introducing Griffin | No date given | Closed preview | Real-time, full-duplex video-to-video 'Human Interaction Model' for face-to-face AI conversation. A smaller Griffin-Lite is a research preview for select trusted testers (request form); Tavus says the full Griffin follows once safety work is done. 1 2 |
| Qwen 4 (Max, Plus, Flash)Alibaba (Qwen) · language model · unknown | Apsara Conference keynote and Alibaba Cloud press release | No date given | Announced | Next Qwen generation on a new architecture, said to be in training; previewed as Max (flagship), Plus and Flash tiers. No model card, API name, price or benchmarks yet; the Qwen LLM lead said only that it will come 'soon'. 1 2 3 |
| Qwen 4 27BAlibaba (Qwen) · language model · open weights | shown in the Qwen 4 lineup at the Apsara Conference | No date given | Announced | 27B model presented as the open, locally deployable tier of the Qwen 4 family. The open-weight plan comes from coverage of the keynote; Alibaba's press release does not spell it out, and no weights are published. 1 2 |
| Qwen-Image 3.1Alibaba (Qwen) · image model · unknown | Alibaba Cloud press release at the Apsara Conference | Later in 2026, Alibaba says | Announced | Image generation model following Qwen-Image 3.0. Alibaba said it will launch 'later this year' (2026) but gave no date; no specs published. 1 2 |
| Grok 4.8xAI · language model · unknown | Elon Musk posts on X (quoted by third-party pages) | No date given | Teased | Musk described a 2.5-trillion-parameter model on a new C++ training stack that would 'finish training this week and start RL', and later called it a noticeable improvement over Grok 4.7. No model ID, price or date from xAI. 1 2 3 |
| DarwinOpenEvidence · language model · API | press release introducing the OpenEvidence model family | No date given | Closed preview | OpenEvidence's most capable medical model, in an application-based research preview for institutional partners, research collaborators and accredited academic AI researchers (partner endpoint). Size not disclosed; capabilities are to move into its production models later. 1 2 |
| GLM-6.0Zhipu AI (Z.ai) · language model · unknown | named by founder Tang Jie on the half-year results call | No date given | Teased | Next Z.ai base model, framed around 'self-evolution' (the model judging when to stop and correct its own training). Tang Jie said the base model will grow while active parameters are kept in check; no date or licence given. 1 2 |
| Sarvam trillion-parameter model (unnamed)Sarvam AI · language model · unknown | roadmap announced at Sarvam's Epoch 2026 conference in Bengaluru | January 2027within six months, Sarvam says | Announced | Trillion-plus-parameter foundation model trained from scratch in India, aimed at coding, cybersecurity, simulation and science. Sarvam said it expects the model within six months (by about January 2027); no name or licence yet. 1 2 |
| FLUX 3 DevBlack Forest Labs · multimodal model · open weights | Black Forest Labs FLUX 3 launch post (launch plan section) | No date given | Announced | Planned open-weight release of the FLUX 3 multimodal backbone for video, audio and image generation and action prediction. The other FLUX 3 parts (Video, Image, Image Klein, Action, Max) have shipped; no date, size or licence for Dev. 1 2 |
| Nemotron 4NVIDIA · language model · open weights | NVIDIA press release launching the Nemotron Coalition (GTC) | As early as late autumn 2026, The Information reported | Announced | Family of open models to be built on the Nemotron Coalition's first model, co-developed with Mistral AI. The Information reported the largest model exceeds 1 trillion parameters and could be ready as early as late fall 2026; NVIDIA has not confirmed a date. Announced slightly more than six months ago. 1 2 |
| Grok 5xAI · language model · unknown | xAI Series E funding post ('Grok 5 is currently in training') | No date given | Announced | xAI's next flagship, in training on Colossus 2. Musk has given shifting targets (most recently that it comes after Grok 4.8 and 4.9); no xAI date. Announced more than six months ago but still unreleased. 1 2 |
An announcement date is the lab's own post, release notes or official account. An expected date is the lab's when it gave one, otherwise the report named beside it, and none is guessed. A model leaves this list when anyone can download it or call it, and gets graded then.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.