Cohere's North Small Translate Beats DeepL Across 50 Languages With Downloadable Weights
Cohere's open-weights 218B mixture-of-experts translation model tops WMT26 under 1T parameters, with 25B active params and no reasoning traces.
20 stories
Cohere's open-weights 218B mixture-of-experts translation model tops WMT26 under 1T parameters, with 25B active params and no reasoning traces.
OpenBMB released a 324M-parameter draft model that accelerates MiniCPM5-2B inference using semi-autoregressive speculative decoding, hitting 5.5x accepted length on average.
A free MIT-licensed bundle of eleven Claude skills chains recon, rebuild, test, and launch steps to clone any app end to end.
A compact 44M-parameter encoder from Microsoft Research keeps climbing the trending charts, offering ELECTRA-style efficiency and strong NLU accuracy at tiny inference cost.
A from-scratch Vietnamese TTS model ships 48 kHz speech, 23 preset voices, instant cloning, and 16 concurrent real-time streams on a single RTX 3060.
A new open-source benchmark called livenerf runs a frozen test panel every day for 30 days to detect whether a frontier model quietly degrades after launch.
Ideogram's new precision edit model debuts at #23 on the image editing leaderboard, leading on text edits but trailing the frontier on raw generation.
A new survey by Abdelkader Baggag and Yousef Saad bridges the gap between classical numerical linear algebra and the matrix-heavy machinery powering modern LLMs.
A new paper argues the right dropout placement, not just positional encoding, lets transformers extrapolate up to 64x past their training context length.
Amazon AGI's new diffusion language model spends more compute on hard tokens and less on easy ones, beating same-size autoregressive baselines across 11 benchmarks.
Prime Intellect launches its inference platform serving 600B tokens daily, with NVFP4 KV compression, prefill/decode disaggregation, and sub-second TTFT on GB200 NVL72.
PyTorch's Helion kernel DSL now powers a vLLM linear backend that beats CUTLASS and DeepGEMM on Hopper, with 10%+ end-to-end throughput gains.
Meta's Muse Spark worked alongside mathematicians through the standard chat interface to co-author six research papers, five of which answer previously open questions.
Princeton's Queen couples a Leela chess network to SmolLM3 via cross-attention, then self-improves through an AlphaZero-style natural language distillation loop.
A new loss function reframes training as a planning problem, beating cross-entropy on ImageNet with a one-line change that scales with label noise.
Wayfarer learns reusable skills from pixels alone, cracking Montezuma's Revenge and Private Eye while beating DreamerV3, Rainbow, and IQN.
Google's Project Suncatcher just put four Trillium TPUs into low Earth orbit on a Planet-built satellite, testing whether AI compute can survive and run in space.
Ai2 open-sources an 8B Qwen3-based model that generates cited scientific reports 3.5x faster than Claude-powered pipelines, with full training data.
Google rebuilt Federated Learning around Trusted Execution Environments, moving gradient computation server-side with externally verifiable differential privacy guarantees.
Arena's new recipe combines a 5M-vote preference model with rubric-based rewards, pushing FLUX.2-dev and Ideogram 4 to new leaderboard highs.