Updated

AI models released in October 2026

58 model releases from 39 providers in October 2026, with announcement dates, official sources, and practical notes for builders.

Releases this month

Dates refer to the linked announcement or release milestone. Previews are labeled in the entries; historical availability and pricing may have changed.

Release details

  1. DiffuSpace

    Clover-35B-A3B-Preview (preview)

    DiffuSpace released Clover-35B-A3B-Preview, a 35B-parameter mixture-of-experts diffusion language model with about 3B active parameters and a thinking mode. DiffuSpace reports results on coding and agent benchmarks; those are the company's evaluations.

    For buildersIf evaluating the preview for agent work, compare end-to-end task success and total latency on your own fixed tasks, not just tokens per second.

  2. OpenAI

    GPT-6 Rosalind Research (limited access)

    OpenAI released GPT-6 Rosalind Research for multistep life sciences research, bringing GPT-6 reasoning capabilities with safeguards designed for biological work. Access is limited to approved Research organizations and users.

    For buildersFor approved research teams, assess it on a bounded literature or data-analysis workflow and keep scientific conclusions subject to expert review.

  3. TBC mischief 1

    The Biological Computing Co. launched mischief 1, an image-to-video model built on Wan 2.2. The company says its neuron-derived optimization generates video five times faster than the base model; it is available through fal and OpenRouter.

    For buildersTry it on a small image-to-video batch at the resolutions and clip lengths supported by the endpoint, then compare quality and total generation cost with your current workflow.

  4. InclusionAI

    Ling-3.1-flash

    InclusionAI released Ling-3.1-flash, a hybrid reasoning model with 560B total parameters, about 25B active per token, and a model context limit of 1M tokens. It builds on Ling-3.0-flash and targets real-world agent, search, office, and software-engineering tasks.

    For buildersFor local deployment, start with the model card's hardware and serving guidance, then benchmark representative long-context tasks before choosing a hosted or self-managed setup.

  5. Pika

    Pika H3 Flex

    Pika announced H3 Flex, a post-trained version of MiniMax H3 for iterative video creation. Pika says it supports up to 12 multimodal references, synchronized audio, and open weights for local use.

    For buildersFor a short video workflow, compare a small reference-guided generation with your current process, and account separately for local setup and Pika-hosted finishing features.

  6. Google

    Gemini 4 Carbon

    Google DeepMind first introduced Gemini 4 Carbon in limited access on October 28, then expanded availability on November 10 to paying developers and enterprises through its API and business products, with rollout to Google AI Ultra subscribers. The model targets complex, long-horizon agentic work and advanced cybersecurity tasks.

    For buildersFor teams admitted to the rollout, evaluate it on a bounded defensive code-review task and independently verify every suggested change before applying it.

  7. NVIDIA

    Nemotron 3.5 Transcribe

    NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech-augmented language model for instruction-guided transcription, translation, speaker attribution, and spoken-language identification across 40 languages. NVIDIA says it handles recordings up to one hour and supports word-level timestamps on request.

    For buildersEvaluate it on representative recordings with domain-specific names, code-switching, and overlapping speakers. Compare verbatim and cleaned-up output before choosing a transcription workflow.

  8. Particle

    Particle-1

    Particle launched Particle-1, a model post-trained from DeepSeek V4.1 Flash and initially focused on software engineering. Particle says its training approach reduces token use while keeping task success within the base model’s reported confidence intervals on evaluated tasks.

    For buildersCompare it with your current coding model on a small, repeatable software task, tracking both completion quality and token use rather than treating efficiency claims as a quality guarantee.

  9. Rime

    Prism

    Rime launched Prism, a speech-to-speech model designed for enterprise voice agents. Its generated text can be inspected before speech, while tool calls can run alongside spoken responses.

    For buildersEvaluate it on a narrow voice workflow where you need both responsive conversation and a reviewable text checkpoint before an action.

  10. Google

    Gemini Robotics ER 2 (public preview on Google Cloud)

    Google Cloud announced Gemini Robotics ER 2 is now available in public preview on the Gemini platform. The enterprise offering pairs the robotics reasoning model with Google Cloud security, orchestration, and lifecycle-management tools.

    For buildersFor a robotics prototype, test one low-level action interface as a model tool in simulation and verify task progress before connecting physical hardware.

  11. Google

    Genie 4 (limited access)

    Google DeepMind announced Genie 4, a world model for real-time interactive simulation with native audio. Users can steer scenes with natural language, and it is rolling out to a limited group of trusted testers.

    For buildersNot broadly usable yet; teams in simulation, game prototyping, or agent training should watch for Gemini API early access.

  12. Liquid AI

    LFM3

    Liquid AI released LFM3, a new generation of open-weight Liquid Foundation Models that puts text, vision, audio, omni, and decision models on one shared backbone. The first release covers six 2B-class models trained from LFM3-1.7B-Base.

    For buildersA reasonable starting point for teams that want a small model to fine-tune locally for one specialized task on device.

  13. Microsoft

    MAI-Image-2.7 (preview)

    Microsoft AI launched MAI-Image-2.7 in preview, an image editing model for restyling, relighting, text and symbol edits, and asset restoration. Microsoft AI says it improves on MAI-Image-2.6 across all 10 use cases and 7 editing actions evaluated by Artificial Analysis.

    For buildersWorth comparing against your current image editor on product-placement and text-in-image edits, where integration with the scene matters most.

  14. OpenAI

    GPT-6.1 Terra

    OpenAI released GPT-6.1 Terra, a fast, efficient model in the GPT-6 family that combines visual reasoning and computer use with clearer writing. OpenAI says its standard API prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol.

    For buildersWorth evaluating for high-volume, form-filling or document-entry agents where GPT-6.1 Sol is more model than the task needs.

  15. ByteDance

    Doubao-Seedance-1.0-pro-fast

    Volcengine released Doubao-Seedance-1.0-pro-fast, a video-generation model now available on Volcano Ark. The company says it builds on Seedance 1.0 Pro and generates up to about three times faster, with lower cost and multi-shot generation.

    For buildersCompare it with Seedance 1.0 Pro on the same short, multi-shot prompt, measuring generation time and reviewing whether the faster output still fits your quality needs.

  16. Block

    Dessa (research preview)

    Block introduced Dessa, a full-duplex speech model trained from scratch on human conversation, with speech input and output and no text or pretrained-model training. It is a research preview, with early access offered to a small group of AI safety researchers.

    For buildersIf you can access the preview, evaluate interruption handling, latency, and emotional nuance in a short spoken task, with human review before relying on responses.

  17. Moonshot AI

    Kimi K3.1 Pro and K3.1 Pro Highspeed

    Moonshot introduced Kimi K3.1 Pro, the first model in its K3.1 series, alongside K3.1 Pro Highspeed. The company describes a 3T-class open-weight model with a 1-million-token context window and improvements for coding, 3D creation, financial research, and agentic tasks.

    For buildersEvaluate both variants on the same long-context coding or research task, comparing output quality, response time, and review effort on your own inputs.

  18. NVIDIA

    Nemotron 3.5 Bolt E4B 4GB

    NVIDIA released Nemotron 3.5 Bolt E4B 4GB, an open small language model for conversational agents. NVIDIA describes multi-turn tool use, configurable reasoning, a 64K context window, and deployment on supported local NVIDIA hardware.

    For buildersTry it on a narrow local voice-assistant or tool-calling workflow, then measure response quality, latency, and memory use on your target hardware.

  19. Alibaba

    RynnUnify

    Alibaba DAMO Academy introduced RynnUnify, a 16B embodied model pairing an 8B reasoner with an 8B video generator to support visual planning through interleaved video-language reasoning. The project also describes agentic and world-action extensions.

    For buildersFor a robotics or simulation prototype, compare direct reasoning with generated visual evidence on a multi-step task, then check whether the predicted scene changes improve the next action.

  20. Canva

    Canva Craft

    Canva introduced Craft, its first image-generation foundation model built for design workflows. It is available in Canva and through a public API, producing images up to 2K resolution with composition planning and editable outputs.

    For buildersTry the same poster or product-graphic prompt in Craft and your current image model, then compare text placement, composition, and how easily you can revise one element.

  21. InclusionAI

    Ling-3.0-flash-Sante

    InclusionAI open-sourced Ling-3.0-flash-Sante, a healthcare-focused research model with 124B total parameters and 5.1B active per token. Its model card describes medical research, long-document analysis, tool-assisted workflows, and a 256K context; it is not a medical device and requires qualified human review for high-stakes use.

    For buildersUse it only in a research sandbox to summarize medical literature with cited source passages, and have a qualified reviewer verify every claim before any consequential use.

  22. Tencent

    WeVisOne-VL

    Tencent’s WeChat Vision introduced WeVisOne-VL, a vision-language model that combines object-level tasks such as detection, localization, counting and OCR with pixel-level segmentation and matting. The open model family includes 2B and 4B variants.

    For buildersTry the 2B and 4B variants on the same image set to compare structured object localization with fine-grained segmentation and matting outputs.

  23. Recursive

    Flint

    Recursive released Flint, an open mixture-of-experts language model designed for coding agents. Its model card describes repository exploration, bug fixes, feature implementation, tests, and multi-turn tool use; model weights are available from Recursive’s Hugging Face repository.

    For buildersEvaluate it on a small, representative repository task, such as diagnosing a failing test and proposing a fix, then verify every change with your usual tests and review process.

  24. Base Labs

    Mammoth-1 and Elephant-1

    Base Labs released mammoth-1 (27B) and elephant-1 (9B), open-source decision models that return probabilities and structured choices, yes/no answers, or scores instead of free-form text. They are available through Baseten Model APIs.

    For buildersCompare them on a small set of your own routing, classification, or rubric-scoring cases, and set a review threshold for uncertain outputs before automating decisions.

  25. Mistral

    Voxtral 3

    Mistral introduced Voxtral 3, a family of speech-to-text and text-to-speech models with a voice-agent SDK for real-time, multi-step conversations. Mistral says it is available through its EU-hosted API and for self-deployment.

    For buildersEvaluate it on a short, representative voice workflow that combines transcription, a tool call, and a spoken response, then review accuracy, latency, and handoff behavior on your own samples.

  26. Inception Labs

    Mercury 3

    Inception Labs released Mercury 3, a diffusion-based language model with a 1M-token context window. It is available through the Inception API Platform and listed distribution partners.

    For buildersCompare Mercury 3 with your current model on a representative time-sensitive task, measuring response quality, latency, and review effort on your own inputs.

  27. MiniMax

    MiniMax M3.1 Flash

    MiniMax released M3.1 Flash for agentic tasks, coding, computer use, and cybersecurity. It is available through the MiniMax API, M Plan, and MiniMax Code; model weights are expected later, according to MiniMax.

    For buildersTest it on a bounded workflow involving several tool steps or a coding task, then compare task completion, latency, and human review needs with your current model.

  28. Reflection

    Beam

    Reflection released Beam’s open weights on October 16 under the Apache 2.0 license, following its earlier early-access announcement. The 501B-parameter sparse MoE model has 23B active parameters and is built for coding, reasoning, and agentic workloads; it can be self-hosted or accessed through Reflection and partners.

    For buildersCompare the released weights or a hosted endpoint on a bounded coding task from your own workflow, tracking successful completion, tool-use errors, latency, and review effort.

  29. Perplexity

    pplx-decider-v1.2 and pplx-decider-flash-v1.2

    Perplexity released two open-weight multimodal decision models: pplx-decider-v1.2 and pplx-decider-flash-v1.2. They process multiple questions about the same input in one pass and return structured decisions with probabilities, rather than free-form text.

    For buildersTry both sizes on a small set of your own routing or classification cases, then review probability calibration and uncertain decisions with a human before automating actions.

  30. OpenAI

    GPT-Live-1.1 and GPT-Live-1.1 Quip

    OpenAI began rolling out GPT-Live-1.1 and the mini GPT-Live-1.1 Quip model in ChatGPT Voice, with more natural accents and additional language-specific voices. Availability varies by region and account; the announcement describes a ChatGPT Voice rollout, not general API access.

    For buildersIf the models are available to your account, compare them on short voice interactions across supported languages, checking transcription, turn-taking, and response quality.

  31. Jais 3

    The Institute of Foundation Models and Inception42 released Jais 3, a family of six open Arabic language models from 900 million to 375 billion parameters, alongside Jais Speech for Arabic speech recognition. The models support Arabic across regional dialects and are available on Hugging Face.

    For buildersEvaluate the model family on representative Modern Standard Arabic and regional-dialect prompts, plus speech samples if using Jais Speech; check transcription and task completion against a reviewed baseline.

  32. Liquid AI

    Liquid AI d1-ultra

    Liquid AI released d1-ultra, a multimodal decision model that accepts text and images and returns probabilities for yes/no, choice, or scoring questions without generating tokens. It is available through the Liquid AI API; performance comparisons in the announcement are provider-reported.

    For buildersTry it on a narrow classification or visual-inspection task with labeled examples, then check its decisions and probability calibration against human-reviewed outcomes.

  33. Meta

    Muse Spark Open

    Meta released the open-weight Muse Spark Open, a natively multimodal model with 816B total parameters, 42B active parameters, and a 1M-token context window. Meta says it is built for coding and agentic work and is available through hosting platforms and downloadable weights.

    For buildersCompare it with your current model on a bounded coding or multimodal agent task, tracking task completion and review effort on your own examples.

  34. NVIDIA

    Nemotron 3.5 Super VL and Bolt (Bolt E4B 4GB now available)

    NVIDIA released Nemotron 3.5 Super VL and Bolt 74B-A9B 24GB on October 15, then released Nemotron 3.5 Bolt E4B 4GB on October 23. The family targets multimodal agent workflows and local conversational agents; NVIDIA describes E4B as a small open model with multi-turn tool use and a 64K context window.

    For buildersEvaluate the available Super VL or Bolt model on a representative visual-document or local-agent task, and compare accuracy, latency, and hardware needs with your current setup.

  35. xAI

    Grok Voice 3.0

    xAI released Grok Voice 3.0, a speech-to-speech model available through its Voice Agent API. The API is unchanged, and the grok-voice-latest alias now points to version 3.0.

    For buildersCompare it with your current voice agent on representative calls, including interruptions and varied accents, and review task completion, latency, and handoff quality.

  36. Sqwish Labs

    kettei models

    Sqwish Labs released the kettei decision-model family: kettei fat and chub handle text and images, while core and dot are text-only; core and dot are open source. kettei fat, chub and core are available through OneHot, with fine-tuning for fat in early preview.

    For buildersTry an exact, bounded decision task such as routing or evidence checks with fixed choices, then compare its scores and review workload against the current classifier or LLM workflow.

  37. Google

    Gemini 4 Argon (broader access)

    Google announced broader availability for Gemini 4 Argon, a model designed for long-horizon reasoning across software engineering, enterprise knowledge work, and cybersecurity defense. Starting October 13, developers can use it through the paid Gemini API and Android Studio, with access also available in Google AI Studio and Antigravity for AI Ultra subscribers, and in Gemini app and Search AI Mode for eligible subscribers.

    For buildersIf you have access, compare Argon with your current model on a representative long-horizon coding or document-work task, tracking completion quality and review effort rather than relying on provider benchmark claims.

  38. Noiz AI

    Helix World (preview)

    Noiz AI and HKUST introduced Helix World, a real-time interactive world model that generates video and spatial stereo audio together from a starting image, conditioned on camera movement and keyboard input. A preview checkpoint is available on Hugging Face alongside a public demo, paper, and code.

    For buildersTry the preview on a small set of starting images and camera movements, then review whether generated motion and stereo sound remain coherent as the viewpoint changes.

  39. Meta

    Muse Wonder and Muse Wonder Full

    Meta introduced Muse Wonder and Muse Wonder Full, multimodal models for personal and professional agentic tasks that use Active Memory and Active Focus. Muse Wonder is available through the Meta Model API and Muse Code; Muse Wonder Full is available to Muse Code subscribers.

    For buildersCompare the models on a small set of multi-step workflows that require recalling earlier context, using tools, and recovering from a blocker. Track task completion and review consequential actions before relying on them.

  40. Cloudflare

    Cloudflare Clef-omni

    Cloudflare released Clef-omni, an open-weight decision model that accepts audio, video, image, and text inputs in one API call. Its release builds on the existing Clef family and makes its weights available on Hugging Face.

    For buildersTry Clef-omni on a workflow that currently chains transcription with image or video analysis, then compare decision quality and latency against that pipeline.

  41. Microsoft

    Microsoft-Decision-1

    Microsoft introduced Microsoft-Decision-1, a decision-scoring model for structured tasks such as routing, classification, prioritization, and verification. It returns calibrated scores for fixed choices and is available through Microsoft Foundry and OpenRouter.

    For buildersEvaluate it on a representative routing or classification task with a fixed set of choices, and measure decision quality and latency against the current approach.

  42. JetBrains

    Mellum2.1

    JetBrains released Mellum2.1, an open 12B mixture-of-experts model for coding agents. Its post-training uses reinforcement learning in real coding environments, enabling it to explore repositories, edit files, and check its changes; the model is available on Hugging Face under Apache 2.0.

    For buildersTry it on a representative codebase task that includes diagnosing a failing test, proposing a patch, and checking the result; compare quality and latency with your current coding model.

  43. Anthropic

    Claude Haiku 5.5

    Anthropic released Claude Haiku 5.5 for high-volume, cost-sensitive tasks including summaries, database queries, classification, and coding subagent work. It is available now across the Claude Platform and cloud providers, with an adjustable effort setting.

    For buildersCompare Haiku 5.5 with a larger model on representative classification, summarization, or coding-subagent tasks, tracking quality, latency, and cost before routing production work.

  44. Liquid AI

    Liquid AI d1-3B and d1-omni-600M

    Liquid AI released two open-weight decision models, d1-3B and d1-omni-600M, which produce answers in a single forward pass rather than generating tokens. d1-3B supports text and image inputs, while the experimental d1-omni-600M supports text+image or text+audio; both are available on Hugging Face.

    For buildersTest the d1 models on a bounded decision task using representative text, image, or audio inputs, then compare answer quality and latency with your current model before adopting them.

  45. Perplexity

    pplx-embed-v2-late (0.6B and 9B)

    Perplexity released two late-interaction embedding models, pplx-embed-v2-late-0.6b and pplx-embed-v2-late-9b, for text and image retrieval. They retain token-level vectors and share an embedding space, and both are publicly available on Hugging Face.

    For buildersCompare the 0.6B and 9B variants on a representative set of text queries and scanned or rendered documents; measure retrieval relevance alongside index size and query latency.

  46. Google

    EmbeddingGemma 2

    Google launched EmbeddingGemma 2, an open multimodal embedding model that maps text, code, images, video, and audio into a shared embedding space. Its 740M-parameter weights are available under Apache 2.0 for on-device search and retrieval.

    For buildersCompare local retrieval with your current embedding model using the same text, code, image, audio, and video queries; review relevance, latency, and device memory before adopting it.

  47. Mistral

    Mistral Large 4

    Mistral released improved Mistral Large 4 weights on October 29, following its public preview earlier in October. The natively multimodal model has a 1 trillion parameter scale and 52 billion active parameters, according to Mistral.

    For buildersTest the released weights on representative coding and defensive-security tasks, then compare output quality and token use against the public preview or another model before deployment.

  48. Google

    Nano Banana 2.1

    Google’s Nano Banana 2.1 is an image-generation and editing model in the Gemini 3 family. Its model card lists text and image inputs, image and text outputs, and distribution through Gemini, Google AI Studio, the Gemini API, Search AI Mode, Ads, Flow, and Stitch.

    For buildersCompare it with your current image model on a small set of prompts that test text rendering, image edits, and multi-reference consistency; review outputs for fidelity and instruction-following errors.

  49. Perplexity

    pplx-decider-v1.1-27b

    Perplexity released pplx-decider-v1.1-27b, an updated open-weights multimodal decision model. It is designed to return structured decisions such as yes/no, choice, or score outputs rather than generated text.

    For buildersEvaluate it on a small set of representative classification or routing examples, comparing decision probabilities with your current approach and checking uncertain cases with human review.

  50. Amazon

    Amazon Nova 2.5 Sonic

    Amazon announced general availability of Nova 2.5 Sonic, a real-time speech-to-speech model for voice agents. The release adds improved reasoning, instruction following, tool-calling accuracy, and lower latency, according to AWS.

    For buildersEvaluate it with representative voice-agent conversations that require turn-taking, multi-step instructions, and tool use, then review accuracy and handoff behavior.

  51. Reka

    Rho-1 (research preview)

    Reka released a research preview of Rho-1, a 19B omni-reasoning model trained from scratch. The company says it can understand and generate text, images, and video, reason across them, and take actions within one model.

    For buildersIf you can access the preview, test a task that combines visual understanding, generation, and follow-up editing, then review each output for consistency and errors.

  52. Aleph Alpha

    Kolibri

    Aleph Alpha released Kolibri, an English-German open-weight Mixture-of-Experts language model with 78B total parameters, 3B active parameters, and a context length of up to 1 million tokens. Full weights are available under Apache 2.0; the provider describes it as specialized for regulated, mission-critical work.

    For buildersEvaluate it locally on representative English and German documents, including long-context retrieval and abstention cases, before considering it for regulated workflows.

  53. Ai2

    AstaBrief 8B

    Ai2 released AstaBrief 8B, an open-weights model that turns a research question and retrieved literature into a cited scientific report in one pass. The model and training data are available for download; it is also offered as Fast mode in Ai2’s Asta platform.

    For buildersTry it on a small set of research questions with known source papers, then check citation support, omitted caveats, and whether the report is useful as a first draft.

  54. Cactus

    Cactus Whistle

    Cactus released Whistle, a 16.9 MB on-device speech-recognition model for CPU deployment. It transcribes up to 30 seconds of audio in seven languages and also returns word timestamps and speech embeddings without sending audio off-device.

    For buildersTest it on short, representative clips on the target device, measuring word errors, timestamps, latency, and performance on names before using transcripts to trigger actions.

  55. Decagon

    Chord

    Decagon introduced Chord, its first speech model, trained and post-trained for customer conversations and used by its Voice 3 system. The company says Chord shapes speech phrase by phrase and is trained on licensed data and consented voice talent; Voice 3 supports more than 70 languages.

    For buildersIf evaluating a customer-service voice stack, test real turn-taking, interruptions, confirmation details, pronunciation, and handoff behavior across representative calls.

  56. Cloudflare

    Cloudflare Clef and Clef-flash

    Cloudflare released Clef and Clef-flash, open-weight decision models hosted on Workers AI that return probabilities for typed choices instead of generating text. Clef adds image input and a 64K context window; the weights are available under Apache 2.0.

    For buildersEvaluate one bounded task such as support-ticket routing against your current classifier or model, measuring decision quality, confidence calibration, latency, and how often a human should review the result.

  57. Tavus

    Griffin-Lite (research preview)

    Tavus introduced Griffin, a real-time video-to-video model that combines perception, conversational behavior, and audiovisual generation. Only Griffin-Lite is currently available, as a research preview for select early testers; a wider release is still to come.

    For buildersFor a face-to-face agent prototype, evaluate whether live video and interruption handling improve conversation flow, and assess disclosure and consent requirements before user testing.

  58. Microsoft

    MAI-Transcribe-2-Streaming, MAI-Voice-2.1, and MAI-Voice-2.1-Flash

    Microsoft introduced MAI-Transcribe-2-Streaming for live speech transcription and two speech-generation models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. Transcribe supports 60 languages with automatic detection and partial transcripts in just over 100 ms; the voice models support multilingual speech generation, with Flash designed for low-latency workloads.

    For buildersPrototype a voice-agent loop using representative call audio, then compare transcription latency and short-utterance accuracy alongside response quality and voice consistency across the languages you need.