
vLLM's DSpark Beats Every Fixed-Length Config on DeepSeek-V4 at Any Load
vLLM's new adaptive verification for DSpark automatically tunes speculative decoding depth per step, holding the Pareto frontier from concurrency 1 to 256 with a single config.
20 stories

vLLM's new adaptive verification for DSpark automatically tunes speculative decoding depth per step, holding the Pareto frontier from concurrency 1 to 256 with a single config.

models.dev is a community-built, open-source database of AI model specs, pricing, and capabilities — queryable via a free JSON API.

A new agent runtime called StateM pushes Terminal-Bench 2.1 to 95.3% without retraining, and reaches frontier scores on DeepSeek for about $15 in API calls.

METR raises $71M in six months to scale independent AI safety evaluations as rogue deployment risks grow more real

Zed v1.15 ships smarter Git diffs and JSX improvements while Delta, a new multiplayer agent coding environment, enters private beta

Pika launches four generative audio models — SFX, Speech, Music, and Soundtrack — priced up to 20x cheaper than ElevenLabs and other rivals

NVIDIA's open-source NeMo Switchyard routes each agent workflow step to the right model automatically, cutting costs by up to 74% without rebuilding your app
Turboderp's EXL3 quants of Qwen3.8-27B use trellis coding from QTIP to run a 27B dense model on a 12GB GPU with near-lossless quality.
A slot-aware LoRA for Lightricks' LTX-2.5 lets you condition video generation on up to five reference images and two voice tracks.

Alibaba drops open weights for Qwen3.8-27B and the massive 2.4T-A95B Max model, both free under Apache 2.0

Alibaba's new 27B vision-language model runs locally on a laptop at 17GB quantized, with 262k context and coding scores beating closed rivals.
Fastino's new multilingual checkpoint swaps span enumeration for boundary prediction, handling entities, relations, records, and classification in one 287M encoder.
SpaceX closes its $60B all-stock acquisition of Cursor, folding the AI coding giant into SpaceXAI to challenge Anthropic and OpenAI in the developer tools market.

MiniMax open-sourced a text-to-music model that generates full five-minute songs with vocals from lyrics and a structured caption.

RadixArk shipped an NVFP4 W4A4 quantization of Qwen3.8-27B that runs on a single RTX 5090 while keeping GSM8K accuracy above 97%.

Z.ai's GLM-5.3 pushes its 743B base model further with post-training focused on agentic coding and cybersecurity, with open weights coming after safety review

Stanford researchers mined 1.5 billion puzzle-solving sessions with offline reinforcement learning to identify chess puzzles that actually teach beginners something.
L&T and Together AI are deploying 10,000 NVIDIA B300 GPUs in Chennai, marking India's largest single-cluster AI factory in a deal worth up to $1.8B.

An open-source workspace turns Claude Code into a full video production studio, chaining voiceover, music, images, and clips into a rendered MP4 for cents.

NanoNets released Graft, an open-source context layer that turns your codebase into linked markdown nodes so coding agents skip the grep-from-cold problem.