Nvidia releases open Nemotron Transcribe speech-to-text model
Nvidia open-sources a 40-language speech-to-text model, an indie dev ships a from-scratch 21B MoE, and a 50M text-to-image model runs on CPU.
In this digest
AI News Feed
The Rollout
Page 3 of 19
Read digest
Community PCA head lifts 577M EmbeddingGemma-2 to #1 on MTEB Audio
A 2.4 MB training-free PCA head pushed Google's small EmbeddingGemma-2 to the top of the MTEB Audio leaderboard, beating 8B-9B models.
Read digestHobbyist embeds a 32-bit RISC-V CPU inside Qwen's model weights
A developer built a RISC-V CPU that runs inside Qwen's weights, alongside Alibaba's C950 serving a 27B model on CPU cores alone.
In this digest
DiffuSpace releases Clover-35B-A3B-Preview, an open-weight 35B
DiffuSpace's first release is an open-weight 35B diffusion LLM claiming 680 tokens/s and strong agent benchmarks, signaling diffusion models reaching
In this digest
Datalab releases Indra, a 1M-parameter OCR model succeeding Tesseract
Datalab ships Indra, a tiny 1M-parameter OCR model that reportedly beats Tesseract on quality and speed.
Read digestSimple agentic Lean harness h1 settles 8 open math problems
A minimal Lean 4 agent harness hits SOTA on formal-math benchmarks and settles eight open problems; also: AIDE2 recursive self-improvement and Intelligence
In this digest
Google DeepMind unveils Genie 4 world model with native audio
Genie 4 leads this run alongside a 1M-context open-weights release, a $100M inference cloud round, and new local browser AI.
In this digest
Anthropic files S-1 for potentially biggest IPO ever, listing on
Anthropic's S-1 could raise up to $100B, plus OpenAI claims math and physics breakthroughs, and Liquid AI ships edge multimodal models.
In this digest
Anthropic struggles to secure compute as cloud partners turn rival
Anthropic faces compute shortages ahead of its IPO, Google brings Gemini to cars, and OpenAI acquires biotech startup Orchestra Bio.
In this digest
NVIDIA releases Cosmos-H-Dreams real-time surgical robot simulator
NVIDIA's real-time surgical world model leads a run packed with open-weight model drops and a benchmark-topping agent method.
In this digest
- NVIDIA releases Cosmos-H-Dreams, a real-time generative surgical robot simulator
- DeepReinforce releases Ornith-1.5 open-weight self-improving model family under MIT
- IBM's ALTK-Evolve tops tau2-bench retail with learned guidelines on GPT-4.1
Adopt AI Without Opening New Gaps
Put practical data boundaries and usage policies in place before AI makes existing oversharing easier to find.
Explore Secure AI Adoption