Home / New releases / Week of Oct 19–25, 2026
44 verified AI tool, model and product launches from the week of Oct 19–25, 2026, newest first. Each entry links to the official announcement or credible press; dates come from the vendor's post. Last updated 2026-10-25.
Nanbeige LLM Lab (BOSS Zhipin) open-sourced Nanbeige4.3-33B under Apache 2.0: a 33B dense model with a 512K-token context for agentic coding, tool use and office 'cowork' tasks, plus a Base checkpoint. Benchmarks are vendor-reported; a distilled Nanbeige4.3-3B is planned.
Source ↗YC Fall 2026 startup Pomerado launched an open-source service that turns any website, including legacy portals and internal tools, into a hosted API or MCP server for AI agents, built from a single prompt, with self-healing integrations as an alternative to browser agents. Free to try.
Source ↗Moonshot AI released Kimi K3.1 Pro, a 2.8T-parameter (104B active) multimodal MoE model with a 1M-token context, aimed at long-horizon coding, 3D modeling in Blender/CAD, financial research and agents, plus a 2–3× faster Highspeed variant. Live in Kimi apps, Kimi Work, Kimi Code and the Kimi API ($2/$10 per 1M input/output tokens); full open weights are promised by Nov 4.
Source ↗NVIDIA shipped Nemotron 3.5 Bolt E4B 4GB, the small open model announced Oct 15: post-trained from Gemma 4 E4B for local assistants, interactive characters and robots, fitting in about 4GB with multi-turn tool calling. Weights (4GB GGUF, BF16 and a DFlash draft model) are on Hugging Face.
Source ↗MiniMax published the weights for M3.1 Flash (428B total, 23B active per token, 1M context, text/image/video input) on Hugging Face, four days after the Oct 19 API launch. Day-0 serving is on vLLM and Together AI. It is released under the MiniMax community license (attribution required; prior authorization above $20M annual revenue), not MIT/Apache.
Source ↗Hi3D open-sourced Hi3D V1.0 (formerly Sparc3D), an image-to-3D model trained on about 2M curated 3D assets with sparse-grid geometry reconstruction up to 1536³. Code is on GitHub and weights are on Hugging Face, with a free online demo. Hi3D also cut credit use on its hosted platform by over 50%.
Source ↗Canva released Craft, its first in-house image generation model from Canva Research. A two-stage design plans the composition and text before rendering images at up to 2K in 30+ aspect ratios, with text in 55 languages. Available to all Canva users in Dream Lab and via the Canva public API; Canva MCP support is coming.
Source ↗Ideogram launched its first vector model: it generates clean, editable SVGs with smooth gradients and sharp typography from a text prompt, or converts an uploaded image into a scalable vector. Live in the Ideogram app, the API and partner platforms such as fal.
Source ↗OpenAI made the Decisions API generally available on gpt-6-luna. POST /v1/decisions returns a probability, a choice or a score for bounded questions over text, images or JSON, up to 10x faster than the Responses API. GA adds prompt caching, native GPT-Live integration, US/EU inference residency and up to 400 questions per request.
Source ↗Framer launched an official remote MCP server (mcp.framer.com/mcp, OAuth) so AI tools such as Claude, ChatGPT/Codex and Cursor can create pages, edit layouts and content, manage CMS collections and publish Framer sites. It is now the recommended way to connect external agents to Framer.
Source ↗Sume Labs launched the Sume API, an image and video generation API (Seedance, Kling, Veo, Wan, Nano Banana and more) billed at the underlying provider's price with 0% markup. It is available via REST, a hosted MCP/CLI and a web studio, and the pricing page lists side-by-side rates against fal and Higgsfield.
Source ↗Ant Group's Ling team open-sourced Ling-3.0-flash-Sante, a healthcare-focused MoE model (124B total, 5.1B active per token) for medical reasoning, deep research and multi-agent workflows. It reports 54.29 on HealthBench Professional. Weights are on Hugging Face.
Source ↗Hugging Face released Carbon-X, an open (Apache 2.0) 14B model mid-trained on 300B tokens of interleaved DNA and text, so you can ask about a sequence or mutation in plain English or generate DNA. It is for research, not clinical use, and runs locally.
Source ↗ThoughtSpot launched AgentSpot, an agentic workforce platform where business users build, share and run AI agents and workflows grounded in governed company data ('multiplayer AI'), with outcome-based rather than token-based pricing.
Source ↗Anthropic opened a waitlisted research preview of Claude Duet, a proactive agent app (iOS, Android, claude.ai) powered by Claude Opus 5.5 that sends a morning brief, monitors connected apps, works on its own cloud computer and asks before sending, paying or deleting. Included in US Pro and Max plans.
Source ↗Recursive released Flint, an open-weight (Apache 2.0) 40B-parameter mixture-of-experts coding model with about 2.5B active parameters, built for fast local agentic coding (125+ tokens/s on an M4 Pro MacBook Pro) with day-0 support in vLLM, Ollama, llama.cpp and LM Studio.
Source ↗Base Labs released open-weight decision models mammoth-1 (27B) and elephant-1 (9B), post-trained from Qwen3.5, which return calibrated probabilities for routing, classification, yes/no and rubric questions in one pass. Mammoth-1 tops the Jev Decision Index; both are served on Baseten.
Source ↗Google Cloud added OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna to Vertex AI Model Garden in an invite-only private preview, usable via Google Cloud APIs and Codex workflows and payable with existing Google Cloud commitments. General availability is planned in the coming weeks.
Source ↗Breken launched a 'production before production' service that rebuilds your app in isolated worlds with synthetic users, aged data and failing dependencies, then reports reproducible bugs with traces and draft fixes. Its ESCAPES benchmark claims it caught 76 of 235 shipped bugs.
Source ↗Kakao open-sourced two Kanana-2 vision encoders (426M 'medium' and 98M 'tiny') under the commercially usable Kanana Open License, claiming top results across 42 benchmarks, especially Korean OCR and document understanding, with 60%+ lower latency.
Source ↗NVIDIA released Nemotron 3.5 Embed, Rerank and Parse at GTC Berlin: open multimodal retrieval models (embedding and reranking over text and page images, with NVFP4 variants) plus a document-parsing model, aimed at grounding enterprise agents in company data.
Source ↗Perplexity's Computer agent can now run fully on-device in the Mac app, with the model (PPLX Pro or Qwen) and the agent working with local files and native apps like Notes, Mail and Numbers without using Computer credits. For Pro, Max and Enterprise on Apple silicon; 36GB+ memory recommended.
Source ↗EAGLE-4 is an open-source speculative-decoding system with one trained drafter per target model that adapts draft-tree depth, width and verification budget to hardware and batch size. It reports about 4x faster decoding on Qwen3.8-27B, with SGLang and vLLM integrations for NVIDIA B300 and AMD MI355X.
Source ↗Mistral launched Voxtral 3: new speech-to-text models (Voxtral Small/Mini Transcribe 3, incl. realtime), Voxtral Mini TTS 3 with emotion and pace control and voice cloning ($13 per 1M characters), and a voice agent SDK, via its EU-hosted API or self-deployment.
Source ↗ElevenLabs introduced Eleven Realtime, end-to-end speech-to-speech models for voice agents: v1 Mini for natural conversation and v1 Max for complex tool-using tasks, with 90+ languages, 17,000+ voices and an OpenAI-compatible Realtime API.
Source ↗Bria released Fibo 2.0, an 8B open-weights image model trained only on licensed data that unifies generation and instruction-based editing (up to four reference images, masks, JSON controls); Dev and Edit Turbo weights are on Hugging Face and it ranks #1 among open-weights models on Artificial Analysis image editing.
Source ↗Ornith released Ornith-1.5-122B, an MIT-licensed open-weight MoE model for coding and agentic work that runs on a single NVIDIA DGX card, with FP8, NVFP4, GGUF and MLX variants and day-one support in Ollama, SGLang and Hermes Agent.
Source ↗humans& released its first model, Ampersand v0.1, a collaborator model trained in long multi-party simulations to help groups reach shared goals (and to stay quiet when it should); built from GLM 5.3 and available now via an OpenAI-compatible Agents API.
Source ↗Gradium open-sourced Phonon, tiny on-device text-to-speech models (31M and 40M parameters, the smallest about 64 MB) that run offline on phones, laptops, browsers and embedded hardware; the English 31M model is Apache 2.0.
Source ↗UCSD's Hao AI Lab open-sourced FastH3 Omni, a distilled, accelerated version of MiniMax H3 reference-to-video that takes text plus image, video and audio references and generates synced video with audio in eight steps, with NVFP4 checkpoints for local RTX/Spark GPUs.
Source ↗Asymptote Labs launched Beacon, an MIT-licensed, local-first memory and telemetry layer that captures sessions from 20+ coding agents (Claude Code, Cursor, Codex and more) and turns them into reusable skills, alongside a $5.5M seed round.
Source ↗Google added a small-business landing page to the Gemini app with on-brand visuals (upload a brand kit, paired with Nano Banana 2.1), financial insights from messy spreadsheets and receipts, and connections to Google Business Profile, Ads, Workspace and apps such as Canva, Shopify and Xero.
Source ↗NC AI officially released VARCO Sound, a multitrack generative AI sound platform that creates and edits scene-appropriate audio from text, images and video, with a plugin for existing production workflows.
Source ↗Mintlify launched Autopilot in beta on Pro and Enterprise plans: an always-on docs agent that watches your repos plus Slack, project trackers, support tickets and feature flags, drafts doc updates when something changes and opens a pull request for review.
Source ↗The CNCF-governed llm-d project (founded by Red Hat, IBM, Google Cloud, CoreWeave and NVIDIA) released llm-d 1.0, a hardened open-source core for distributed LLM inference aimed at production and context-heavy agentic workloads.
Source ↗MiniMax released M3.1 Flash (revealed as the anonymous 'Space Bunny' model): a 428B MoE with 1M-token context for agents, coding, computer use and security, ~3x faster than M3 at $0.15/$0.60 per 1M input/output tokens; on the MiniMax API, M Plan and MiniMax Code, weights due later.
Source ↗Inception launched Mercury 3, its flagship diffusion LLM: 1,068 tokens/sec on NVIDIA GPUs, 1M-token context, tool calling and structured outputs; $0.20/$0.70 per 1M tokens (50% off at launch) on its API, OpenRouter, Baseten and Vercel.
Source ↗Cloudflare previewed Clairic on Workers AI (@cf/cloudflare/clairic), an extraction model that returns typed values (text spans, numbers, dates, estimates, forecasts) with confidence scores instead of generated text; $0.10 per 1M input tokens, open weights planned.
Source ↗Viggle launched JST-2, a physics-native 3D world model that generates explorable scenes, animation-ready rigged characters from a single image and full-body motion from video in one shared 3D world; free to try and available via the Viggle API.
Source ↗fal launched FLUX 3 Max, a text/image/keyframe-to-video model post-trained with Black Forest Labs on FLUX 3; fal says it renders a 10-second clip in about 11 seconds, 22x faster than Seedance 2.5, at 720p and 1080p on fal's API.
Source ↗Exa released Exa 3.0, a major upgrade to its web search API for AI agents (larger embedding model, Dynamic Highlights GA using 95% fewer tokens, expanded index), plus a new Exa Lite search type at $4 per 1,000 requests.
Source ↗Together AI launched Together Post-Training, a managed service for custom SFT and reinforcement-learning loops on open-weight models (LoRA or full-weight, custom losses) on dedicated B200 GPUs, with a $9/GPU-hour launch promo.
Source ↗doubleAI unveiled Maxwell, billed as the first 'code execution model': it runs inside live Python/PyTorch programs, swaps in faster hardware-optimized code and proves each replacement is semantically equivalent; early access now.
Source ↗Replika founder Eugenia Kuyda's Wabi opened Wabi 2.0 to everyone on iOS, Android and Mac: a multiplayer AI messenger where you add friends or family to threads and the agent plans, books and builds mini-apps on demand.
Source ↗Nanbeige4.3-33B (Nanbeige LLM Lab), Pomerado, Kimi K3.1 Pro + Highspeed (Moonshot AI), Nemotron 3.5 Bolt E4B 4GB (NVIDIA), MiniMax M3.1 Flash open weights, Hi3D V1.0 open source (formerly Sparc3D), Canva Craft (Canva), Ideogram Vector 1.0, Decisions API GA (OpenAI), Framer MCP, Sume API (Sume Labs), Ling-3.0-flash-Sante open weights (InclusionAI / Ant Group), Carbon-X (Hugging Face), AgentSpot (ThoughtSpot), Claude Duet (Anthropic), Flint (Recursive), Mammoth-1 + Elephant-1 decision models (Base Labs / Baseten), OpenAI models on Google Cloud, Breken, Kanana-2 vision encoders (Kakao), Nemotron 3.5 Embed, Rerank + Parse (NVIDIA), Perplexity Computer, fully local on Mac, EAGLE-4 speculative decoding, Voxtral 3 (Mistral), Eleven Realtime (ElevenLabs), Fibo 2.0 (Bria), Ornith-1.5-122B, Ampersand v0.1 (humans&), Phonon on-device TTS (Gradium), FastH3 Omni (Hao AI Lab FastVideo), Beacon (Asymptote Labs), Gemini tools for small businesses (Google), VARCO Sound (NC AI), Mintlify Autopilot, llm-d 1.0, MiniMax M3.1 Flash, Mercury 3 (Inception), Clairic (Cloudflare Workers AI), Viggle JST-2, FLUX 3 Max (fal), Exa 3.0 + Exa Lite, Together Post-Training, Maxwell (doubleAI), Wabi 2.0.
Each launch is checked against the vendor's own post, its official X account or credible press before it is added, and every source link is confirmed to load. Rumored, unreleased or ambiguously dated items are left out.
Older: Oct 12–18, 2026 → · All releases · Atom feed
Browse by week: Oct 19–25, 2026 (44) · Oct 12–18, 2026 (33)