58 releases covered live on the show — every model, product, paper and tool that mattered, with links and our analysis.
Tongyi Lab releases Z-Image generation model
Alibaba's Tongyi Lab released Z-Image, a new image generation model, with support landing in the open-source DiffSynth-Studio toolkit on GitHub. Covered in the AI Art segment alongside HunyuanImage 3.0.
Arcee AI ships Trinity Large: 400B MOE trained in 33 days for $20M
Arcee AI's Trinity Large is a 400B-parameter MOE with 13B active parameters, trained on 17T tokens across 2000 B300 GPUs in 33 days for $20M. It has 512K native context (twice Kimi K2.5), is free on OpenRouter until February 2026, and the panel called it the largest Western open-source lab model.
Lucy 2.0 real-time video generation model
Lucy 2.0, a real-time video generation model, was discussed in the AI Art segment. The episode covered its real-time video capabilities.
Google DeepMind launches Project Genie 3, real-time 24fps world model
Google DeepMind's Genie 3 generates interactive, controllable 3D worlds in real time at 24 frames per second, demoed live on the show with a spaceship exploration and paint persistence on walls. It ships alongside SIMA 2, a self-improving game-playing agent built on Genie 3, and is available to Gemini Ultra subscribers in the US with a one-minute session limit.
Jan AI releases Jan v3, a 4B model built for fast local inference
Jan v3 is a 4B-parameter open model optimized for local inference, hitting 132 tokens/sec with a 262K context window and a 40% improvement on coding. The Jan desktop app it powers has reached 5M downloads.
Moonshot AI releases Kimi K2.5, the new open-source king
Moonshot AI's Kimi K2.5 takes the open-source crown, becoming the most-used model on OpenRouter and topping open-source leaderboards. The panel highlighted its strong agentic coding performance and tool use.
NVIDIA releases PersonaPlex-7B voice model
NVIDIA released PersonaPlex-7B, an open voice/audio model published on Hugging Face with code on GitHub. Listed in the week's Voice & Audio releases.
Tencent launches HunyuanImage 3.0-Instruct image model
Tencent's Hunyuan team launched HunyuanImage 3.0-Instruct, an instruction-tuned version of its image generation model. Covered briefly in the AI Art segment alongside other new image models this week.
Qwen3-TTS: open-source TTS family with 97ms latency and voice cloning
Alibaba's Qwen team released Qwen3-TTS, a full open-source text-to-speech family under Apache 2 that dropped 30 minutes before the show. It spans 5 models from 0.6B to 1.7B parameters, with 97ms latency, voice cloning from just 3 seconds of audio, voice description prompting, and 10-language support.
FlashLabs Chroma 1.0: open-source real-time speech-to-speech under 150ms
FlashLabs released Chroma 1.0, billed as the world's first open-source end-to-end real-time speech-to-speech model with voice cloning under 150ms latency. The 4B parameter model is built on Qwen 2.5 Omni and released under Apache 2; its live demo with RAG and document upload impressed the whole panel.
Inworld TTS-1.5 claims #1 TTS ranking at half a cent per minute
Inworld AI launched TTS-1.5, a closed-source text-to-speech model claiming the #1 ranking with sub-250ms latency. Its headline is price: roughly $5 per million characters (about half a cent per minute) versus ElevenLabs' $120 per million characters.
Liquid AI's LFM2.5-1.2B-Thinking: on-device reasoning under 900MB
Liquid AI released LFM2.5-1.2B-Thinking, a 1.2B parameter reasoning model that runs entirely on-device with under 900MB of memory. Its hybrid architecture with gated convolutions delivers 239 tokens/sec on an AMD CPU and 82 tokens/sec on a mobile NPU, making it practical for edge devices, Raspberry Pi, and older iPhones.
Overworld's Waypoint-1: real-time AI world model at 60fps on consumer GPUs
Overworld released Waypoint-1, a real-time AI world model that runs at 60fps on consumer GPUs. It generates interactive environments live, bringing world-model tech out of research demos and onto hardware people actually own.
Runway 4.5 launches with image-to-video and audio
Runway launched version 4.5 of its video generation model, adding image-to-video and audio support. It was mentioned in the week's news rundown as part of a busy week for vision and video releases.
GLM-4.7-Flash: 30B MoE local coding agent with only 3B active params
Z.AI released GLM-4.7-Flash, a 30B parameter MoE model with only 3B active parameters, designed as the ultimate local coding and agent assistant. It hits 59% on SWE-Bench Verified (approaching Sonnet 4's 64%) and runs at 120 tokens/sec on a stock Mac Studio M3 Ultra, fast enough to run RALF autonomous coding loops even on CPU.
Black Forest Labs drops Flux 2 Klein, fast open-weights image model
Wolfram broke the news mid-show: Black Forest Labs released Flux 2 Klein, a fast 4B/9B image generation model with open weights under Apache 2.0. It is designed for near-real-time editing and style iteration, and Alex used it minutes later in his live Claude Cowork demo.
M3: 235B open-source medical LLM claims to beat GPT 5.2 on HealthBench
Byte released M3, a 235B parameter medical LLM fine-tuned from Qwen3 and licensed Apache 2.0. With only 22B active parameters, it is runnable at usable speeds on an M3 Ultra, and it claims to beat GPT 5.2 on HealthBench. Nisten suggested pairing it with smaller imaging models like MedGemma rather than treating them as substitutes.
Google releases MedGemma 1.5 for offline medical imaging
Google released MedGemma 1.5, a small (4B-class) open model for medical use cases, compact enough to run offline for medical imaging. The panel stressed it is a different model class from Byte's giant M3 medical LLM and that the two pair well together rather than replacing each other.
Meituan's LongCat Flash Thinking: 560B MoE with 27B active, MIT licensed
Meituan released LongCat Flash Thinking, an open-source reasoning MoE with 560B total parameters and only 27B active, under an MIT license. It continued the run of large sparse Chinese open-weights models offering frontier-style reasoning at low active-parameter cost.
Lightricks open-sources LTX-2 synchronized audio-video model
Lightricks open-sourced LTX-2, billed as the first truly open audio-video generation model with synchronized audio and video output, releasing full training code alongside the weights. A distilled version is available to try on Replicate.
Liquid AI LFM 2.5: 1B on-device family with end-to-end audio
Liquid AI released LFM 2.5, a family of ~1.2B parameter on-device models spanning text, vision, and audio, announced at CES alongside AMD's Lisa Su. The models hit 239 tokens/sec on AMD CPU and 100 tokens/sec on iPhone 16 Pro Max, and include a revolutionary end-to-end audio model that skips the traditional ASR-LLM-TTS pipeline entirely, running in as little as 8GB of RAM.
MiroThinker 1.5: 30B search agent beats trillion-param models
MiroMind AI released MiroThinker 1.5, a 30B parameter open source search agent that achieves 56.1% on BrowseComp and 66.8% on BrowseComp Chinese, outperforming trillion-parameter models. It introduces 'interactive scaling' as a third scaling dimension beyond parameters and context, and is a fine-tune of Qwen 3 Thinking with 147K open training samples.
NousCoder 14B: 7% LiveCodeBench jump in 4 days of RL training
Nous Research released NousCoder 14B, an open source competitive programming model that achieved a 7% jump on LiveCodeBench accuracy in just four days of RL training on 48 NVIDIA B200 GPUs. Training used 24,000 verifiable problems, and the release ships under a full Apache 2 license with training code and a benchmark harness.
NVIDIA Alpha Mayo: open source reasoning self-driving models
NVIDIA announced Alpha Mayo at CES, a family of open source reasoning-based self-driving AI models. The models perform end-to-end autonomous driving with explicit reasoning steps, like identifying jaywalkers and stopping accordingly, demoed in a Mercedes-Benz.
Nemotron Speech ASR: 600M streaming model with 24ms latency
NVIDIA released Nemotron Speech ASR, a 600M parameter open source streaming speech recognition model with 24ms median latency and support for 900 concurrent streams on a single H100. Kwindla Hultman Kramer of Daily.co demoed sub-500ms voice-to-voice latency using a three-model pipeline of Nemotron ASR, Nemotron Nano LLM, and Magpie TTS.
Qwen Edit 2512 optimized by PrunaAI: high-res images in under 7s
PrunaAI released an optimized version of Qwen Edit 2512 that generates high-resolution realistic images in under 7 seconds. The optimized model is available to run on Replicate.
Upstage Solar Open 100B: 102B MoE trained on 19.7T tokens
Upstage released Solar Open 100B, a 102B parameter MoE model with only 12B active parameters per token (129 experts, top-8 activation), trained on 19.7 trillion tokens including 4.5T synthetic via a 'data factory' approach. It outperforms GLM 4.5 Air on many benchmarks, features the SNAP PO reinforcement learning technique with a 50% training speedup, and delivers best-in-class Korean language performance.
Claude Cowork: Claude Code for non-developers, 100% written by Claude Code
Anthropic launched Claude Cowork, a research preview that brings Claude Code-style agentic workflows to non-technical users. It was built in a week-and-a-half sprint with 100% of the code written by Claude Code itself; it is Mac-only, requires a Max subscription, and includes a Chrome connector for browser automation. Alex demoed it live, adding Flux Klein support to an image extension project without looking at a single line of code.
Anthropic launches Claude for Healthcare with HIPAA compliance
Anthropic launched Claude for Healthcare, a HIPAA-ready offering as the major labs push into medical AI. The panel noted Claude's Opus 4.5 scoring 92% on Med Agent Bench as part of Anthropic's healthcare positioning.
Doctronic launches first US pilot for AI prescription renewals
Doctronic launched the first US pilot in Utah where AI can autonomously renew prescriptions without a physician in the loop. The service costs $4 per renewal and covers 190 routine medications, excluding controlled substances.
NVIDIA Vera Rubin platform: 5x Blackwell inference at CES 2026
Jensen Huang unveiled the Vera Rubin platform at CES 2026, NVIDIA's next-gen AI computer delivering 50 PFLOPS and 5x inference performance over Blackwell while adding only ~200W of power draw. It needs 75% fewer GPUs for 10 trillion parameter MoE training, packs 72 GPUs per rack with 20.7TB memory and 13 TB/s bandwidth, is 100% liquid cooled, and entered full production just four months after the B300.
OpenAI launches ChatGPT Health waitlist with health record sync
OpenAI launched a waitlist for ChatGPT Health, a privacy-first vertical for health conversations with connected health records and fitness apps including Apple Health, Function Health, MyFitnessPal, and Peloton. The panel noted LLMs are well-suited to medicine since there are only ~2,000 diseases and ~2,000 prescription drugs to master.
Anthropic launches MCP Apps: interactive UI inside Claude chat
Anthropic's MCP Apps render interactive, branded UI components (Box files, Figma, color pickers) directly within Claude conversations, evolving MCP from tools to embedded app experiences. It is protocol-based, so any app can integrate, letting brands reclaim identity from text-only LLM responses.
Google launches agentic Auto-Browse in Chrome with Gemini 3
Google unveiled Chrome Auto-Browse with Gemini 3 Nano integration, bringing agentic browsing to Pro and Ultra subscribers in the world's most-used browser with 4 billion daily users. Native browsing avoids Cloudflare bot detection, and Gemini's 2M context window suits long browsing sessions.
Google adds Agentic Vision to Gemini 3 Flash
Gemini 3 Flash gains agentic vision: a Think-Act-Observe loop that can zoom, crop, annotate, and plot images by generating and executing Python code in the backend. Available in the Gemini app, AI Studio, and Vertex AI.
Browser Use ships as an installable agent skill
Browser Use was released as an agent skill, installable via registries like Vercel's skills.sh. Wolfram flagged it as a signal of the broader shift away from MCP servers toward skills, since skills are easier to use with the CLI or API directly.
OpenAI begins testing ads in ChatGPT Free and Go tiers
OpenAI announced it is testing ads in the ChatGPT Free and Go tiers, framing the rollout around user trust and transparency. The company also announced age detection models for the upcoming adult mode, putting the memory and personalization data of 900M weekly active users in a new light.
Chorus adds agent skills support for every LLM via OpenRouter
Alex used a Ralph loop with Claude Code to add full agent skills support to Chorus, the open-source app that compares answers across multiple LLMs, in about 3.5 hours. The work added a settings panel, filesystem skill discovery, front-matter parsing, and cross-model skill injection, letting the same Claude-style skills run on GPT 5.2 Codex, Gemini, and any OpenRouter model.
Gemini personal intelligence reasons across Gmail, YouTube, Photos, Search
Google shipped personalized AI in Gemini, letting it reason across a user's Gmail, YouTube, Photos, and Search history with explicit opt-in for US Pro and Ultra subscribers. Alex tested it live: it inferred he drives a Tesla Model Y from emails and noticed his recent Honda Odyssey searches, highlighting Google's data moat over OpenAI and Anthropic.
Amazon brings Alexa+ to the browser as a $20/month web chat
Amazon made Alexa Plus, its upgraded smart assistant, available as a web chat interface for $20/month. It supports free-flowing conversations without repeating the wake word, integrates with smart home devices via natural language, and can continue conversations across devices; voice on the web is coming later.
Google brings Gemini 3 into Gmail for 3 billion users
Breaking during the show: Google integrated Gemini 3 into Gmail for 3 billion users, adding AI Overviews, smart replies, and natural language inbox search. It marks one of the largest consumer AI rollouts to date, bringing Gmail into the 'Gemini era.'
xAI launches Grok Imagine API with video generation
xAI released the Grok Imagine API, exposing its image and video generation capabilities to developers through the xAI console. The show subtitle notes Grok Imagine ranking #1 among generation models this week.
Moonshot AI releases Kimi Code coding agent
Alongside Kimi K2.5, Moonshot AI shipped Kimi Code, a coding tool that pairs with its new flagship model's strong agentic coding abilities. The code is available on GitHub with an announcement page at kimi.ai/code.
Claude Code VS Code extension hits general availability
Anthropic's Claude Code VS Code extension reached general availability, bringing full agentic coding directly into the IDE. The GA release makes Claude Code's agent workflows accessible from the VS Code Marketplace without the CLI.
Clawdbot: open-source self-improving personal AI assistant for macOS
Clawdbot, created by Peter Steinberger, is an open-source personal AI assistant that runs locally on your Mac and connects via WhatsApp, Telegram, or Discord. Its killer feature is self-improvement: ask it to learn something and it writes its own skill files, giving a single chat conversation control over multiple agents, persistent memory, voice messages, image generation, and browser automation on your actual computer.
Vercel launches skills.sh, an 'npm for AI agents'
Vercel launched skills.sh, a registry where you can browse and install agent skills from the command line for any agent, including Clawdbot. It hit 20K installs within hours, and releases like Browser Use shipping as a skill signal a broader shift from MCP servers toward skills.
Vercel releases official agent skill packs for Next.js and React
Vercel began releasing official agent skill packs for Next.js and React, packaging its framework expertise in the agent skills standard. Ryan Carson highlighted that you can point any skills-compatible coding agent at the pack and it installs the skills for you, an early sign of experts shipping domain knowledge as skills.
Catnip by W&B: open source iOS app to run Claude Code anywhere
Chris Van Pelt of Weights & Biases released Catnip, an open source iOS app that lets you run Claude Code from anywhere via GitHub Codespaces. It is available on the App Store with source on GitHub.
Sakana AI's RePo lets LLMs dynamically reorganize their context
Sakana AI introduced RePo, a research technique that lets language models dynamically reorganize their context for better attention. The paper proposes a new way to manage what a model focuses on, aimed at improving performance on long-context tasks.
KAIST's Avatar Forcing: real-time interactive talking heads
KAIST published Avatar Forcing, a framework for real-time interactive talking-head avatars with approximately 500ms latency. The paper targets responsive, live avatar interaction rather than offline video generation.
xAI raises $20B Series E at $230B valuation with NVIDIA backing
xAI raised a $20B Series E at a $230B valuation with NVIDIA and Cisco as strategic investors, even as Grok faced major backlash over its image model's lack of NSFW guardrails ('bikini-gate'). The company claimed 600M active users by counting all X users.
Zhipu AI becomes first major LLM company to IPO, raising $558M
Zhipu AI, the makers of the GLM model family, became the world's first major LLM company to go public, listing on the Hong Kong Stock Exchange and raising $558M. The IPO marks a milestone for Chinese AI labs monetizing frontier model development.
Klein team acqui-hired by OpenAI Codex
The Klein team was acqui-hired by OpenAI's Codex group following the viral 'imagine the smell' hackathon controversy. Discussed as part of the growing Codex ecosystem, which Peter Steinberger used to build Clawdbot entirely.
OpenAI acquires Torch Health to power GPT Health
OpenAI acquired Torch Health as part of its push into healthcare with GPT Health. The move came the same week Anthropic launched Claude for Healthcare, with both labs racing toward HIPAA-ready medical AI products.
NVIDIA acquires Groq team and licenses its tech for ~$20B
NVIDIA entered an exclusive licensing deal with Groq and acquired most of its team for approximately $20B. Groq's inference-optimized chips, created by former Google TPU lead Jonathan Ross, complement NVIDIA's training dominance as inference demand grows exponentially across AI use cases.
Anthropic publishes 90-page Claude Constitution values document
Anthropic published a roughly 90-page Constitution for Claude, a values document baked into the model at training and reinforcement learning time rather than a runtime system prompt. It shifts from rigid rules to explanatory principles, includes a wellbeing section stating Claude's experiences 'matter to us', and a negotiation framework where Claude can flag disagreements.
OpenAI inks $10B deal with Cerebras for 750MW of high-speed compute
OpenAI announced a $10 billion partnership with Cerebras for 750 megawatts of high-speed inference compute, with capacity starting in 2028. It extends OpenAI's pattern of locking in massive compute supply deals beyond its existing cloud partners.
Ralph Wiggum autonomous coding technique hits 1.2M views
Ryan Carson published a viral breakdown (1.2M views on X) of Ralph Wiggum, the autonomous coding technique created by Jeff Huntley: write a PRD, break it into atomic user stories with acceptance criteria in JSON, then run a bash loop that has a CLI agent pick the next story, code it, commit, and loop. The technique works with any CLI agent (Amp, Claude Code, Cursor CLI, Gemini CLI), compounds learning via agents.md, and won a YC hackathon running overnight on Sonnet 4.5.
Don't wait a month — get every week's releases recapped in your inbox each Thursday.
