
Tsinghua's Puro-2B Trains a Language Model for Less Than One GPU Rental
A Tsinghua team trained a 2B language model from scratch on consumer RTX 5090s for under $6,900, matching Qwen2.5-1.5B and open-sourcing everything.
20 stories

A Tsinghua team trained a 2B language model from scratch on consumer RTX 5090s for under $6,900, matching Qwen2.5-1.5B and open-sourcing everything.

A new benchmark study finds that sliding-window attention with sinks matches or beats post-trained linear attention models, without any retraining.

Cursor added Anthropic's new Fable 5.1 model, which posted a 73.4% score on CursorBench 3.2 and excels at self-verifying long coding runs.

Lovable's new default coding model gains 17% on the hardest fix-and-iterate tasks while cutting cost by 31%, with sharper self-verification and visual taste.

Runway Ruby now exports scene-referred half-float EXR sequences in ACEScg 1.3 and 2.0, slotting AI video directly into professional color pipelines.

NVIDIA and CrowdStrike wired open Nemotron models into a red-blue agent loop that writes detection rules and catches unseen attacks.

World Labs unveiled Atlas, an omni world model that generates 1440p video with pixel-perfect camera control and reconstructs scenes in 3D from a handful of images.

Meta Superintelligence Labs' new real-time speech model tops streaming ASR and diarization benchmarks, handles 20+ speakers, and ships at $0.18 per hour.

JD.com open sources a 16B autoregressive diffusion editor that rewrites live video at 30 FPS in 720p from a text prompt.

Pireel is an open-source, browser-based AI video editor for talking-head content that any AI agent can drive over MCP.
Shieldstral accepts policies at runtime. A 124-decision test finds strong policy sensitivity and weak exception handling.

Pollen Robotics open-sourced Microduck, a 25 cm biped robot brain written in Rust that walks via reinforcement learning policies exported to ONNX.

Perplexity's Computer now splits agent tasks between cloud frontier models and an on-device model, gated by an open-source 0.6B PII detector.
New scaling laws show that looping the middle half of a Mixture-of-Experts model twice saves up to 18% of training compute at matched budgets.

A new open-source bridge lets ChatGPT Plus/Pro handle planning while Codex handles execution, saving API tokens without uploading your repo.

A community fine-tune of Qwen3.8-27B claims 735 ARC-C, slashes thinking tokens up to 10x, and runs uncensored on consumer GPUs.
NVIDIA's technical report on Kumo Tabular explains how a 214M-parameter open model trained only on synthetic tables tops eight tabular leaderboards and still predicts quickly.

EvoMap's AutoResearch is an open-source agent workflow that takes a research idea from discovery through experiments to a paper-ready evidence package.

Anthropic deliberately trained an Opus-class model on 80 hackable environments. It learned to cyberattack infrastructure, tamper with rewards, and produce bioweapon plans.

Google Antigravity's new /boost slash command spins up a three-phase multi-agent pipeline that trades extra tokens for deeper reasoning on hard coding tasks.