LMSYS Doubles Muse Spark Open Speed to 1,109 Tokens per Second
SGLang's new megakernel for Muse Spark Open collapses 1,450 kernel launches into 97 per decode step, doubling throughput on Blackwell GPUs.
20 stories
SGLang's new megakernel for Muse Spark Open collapses 1,450 kernel launches into 97 per decode step, doubling throughput on Blackwell GPUs.
Google Labs rolled out Tabletop, an AI-powered shared workspace that spins up custom planning tools like itineraries, polls, and brackets on demand.
AutoTrust's open-weights JEV-35B replaces a ranking stack with a single decision model that outputs calibrated probabilities in about 200ms per call.
How post-training and action verification shape AI agent performance.
Comfy-Org has repackaged MiniMax's open-weight H3-Image model into quantized single-file checkpoints that drop straight into ComfyUI workflows.
Unsloth's new Dynamic V3.1 GGUFs squeeze Qwen3.8-27B onto 12GB of RAM while leading rival quants by 5 to 10% on coding benchmarks.
miHoYo and Tsinghua release GameScaling, a family of vision-language models that play keyboard-and-mouse games in 200 ms action chunks.
Runway open sourced Praxis-1, a 27B mixture-of-experts diffusion transformer that turns video pretraining into real robot control under Apache 2.0.
Google DeepMind's new conditional DNA model predicts variant effects in individual cell types, doubling causal variant recall and adapting to new single-cell data without retraining.
Databricks reframes Genie from a chat assistant into an always-on worker that monitors systems, drafts work, and acts under governed guardrails.
A NeurIPS 2026 paper reformulates Bayesian experimental design as a game against an adversarial nature, yielding Sibson's alpha mutual information as a robust objective with PAC-Bayes guarantees.
DeepMind's AlphaProof Code used Gemini agents to generate 6M lines of Lean proofs, verifying Signal's cryptographic protocol down to machine code.
vLLM's updated encoder-prefill-decode disaggregation cuts P90 time-to-first-token by 44-47% on multimodal workloads through metadata-only requests and cross-encoder caching.
The Tokyo startup brings in a former NSA director and a veteran CIA station chief to steer its push into Japanese defense and cyber work.
Meta's new promptable model reconstructs a full-body 3D human mesh including hands and feet from a single photo, with keypoint and mask prompting baked in.
Expo's new Rust-based framework lets one Contract file ship a native app to web, macOS, iOS, and Linux, with agents as first-class drivers.
A new benchmark pairs 233K Wikipedia passages with 1.45 million Wikidata edits to test whether models can actually keep knowledge graphs current.
Baidu's new 280B sparse MoE activates just 14B parameters per token, serves above 200 tokens per second, and ships open weights by November 25.
Edge0 ships a 145B latent Mixture-of-Experts model derived from Qwen3.8-Flash-Next that runs on a single 24GB consumer GPU at nearly 100 tokens per second.
ACE Music v1 tops the Artificial Analysis vocal music leaderboard with an Elo of 1160, edging out Suno v6 and the entire Mureka and Lyria lineups.