Calathea

Agree, this was a huge week for open AI. The named releases here span 16 drops across 12 companies: 7 US-origin drops, 7 China-origin drops, 1 from Canada, and 1 from Czechia. US (7): NVIDIA = 3: Nemotron 3, NeMo ASR, Cosmos; Google = 2: Gemma 3n, Magenta RealTime; LiquidAI = 1: LFM2; Boson = 1: Higgs Audio v2 China (7): StepFun = 1: Step-3.1; RedNote = 1: dots.tts; Baidu PaddlePaddle = 2: PaddleOCR-VL-1.6, NAVA; JD.com = 1: JoyAI-Echo; ByteDance = 1: Bernini-R; VAST = 1: TripoS Canada (1): Ideogram = 1: Ideogram 4; Czechia (1): JetBrains = 1: Mellum 12B Not just language models. Image, audio, speech, docs, video, 3D, and world models all had credible open releases in the same week.

Before the week ends, let's acknowledge one of the most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality: 🧠 LLMs → NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models. → Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week. → StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0. → Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week. → JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0. 🎨 Image gen (the surprise of the week) → Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights. 🔊 Audio & Speech (a breakout week for open TTS, 4 labs shipped) → Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA. → RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0. → Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos. → NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B. 👁️ Vision & VLMs → PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0. → Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0. 🎬 Video, 3D & World Models → NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI. → JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3. → ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).

Before the week ends, let's acknowledge one of the most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality: 🧠 LLMs → NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models. → Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week. → StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0. → Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week. → JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0. 🎨 Image gen (the surprise of the week) → Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights. 🔊 Audio & Speech (a breakout week for open TTS, 4 labs shipped) → Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA. → RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0. → Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos. → NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B. 👁️ Vision & VLMs → PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0. → Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0. 🎬 Video, 3D & World Models → NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI. → JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3. → ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).

Thrilled to have worked closely with OpenAI to build @OdessiaTravel. In 5 months, we built a travel agent that can plan and book entire trips: flights, hotels, experiences, places to see – all personalized to you. Options come back in a few seconds with rich visuals. Grateful for @OpenAIDevs for the partnership.

Finally! the first eval ship from cog!!!!!!!!!! 👼🏼 To contextualize: @METR_Evals cap out at ~16 hours. Cog has private enterprise evals up to 100hrs, and is confident enough to put a financial guarantee on it 🤯 METR dataset: ML eng, GPU kernels, cybersecurity > "METR (2026) used a combination of GPT-4o and GPT-5 to estimate the human-equivalent times from compressed Claude Code transcripts. These transcripts were collected from 7 METR technical staff on 34 sessions labeled on human ground truth". rlog​ of 0.83 Cog dataset: real life java/typescript/python/c# feature dev, bugfixes, migrations > "We collected a ground-truth dataset by asking Devin users to review recent representative sessions, and estimate how long each completed session would have taken without Devin. Our dataset consists of 258 sessions from 126 users across a diverse set of enterprise customers." rlog​ of 0.74 on held out set this is pioneering real world evals work and part 1 of a broader frontier code evals drop that I'm really looking forward to writing up. huge kudos to @annarmitchell and @ryanbai1412 for leading the unglamorous last mile data collection!!

Thanks everyone for coming down last Friday night for Hermes Night! Met plenty of great people. What stood out from my previous experiences in SG was how many were already thinking past “AI demos” and into building real operator tools, multi-agent setups, sandboxing infra and are in the midst of pushing actual products. Once again, a big thanks to @NousResearch for supporting Singapore’s first Hermes Agent meetup, and to everyone who made the room feel alive! 🫶