Flint is a mixture-of-experts model: it routes each token to 6 of 256 expert sub-networks plus a shared one, so only about 2.5 billion of its 39.8 billion parameters do work per token. In Recursive's release evaluations it scores 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1. The BF16 weights are on Hugging Face under Apache 2.0, with a 262,144-token context, and run on one B300 or H200 GPU per replica.
Cohere's Rerank 5 adds Context Focus, which returns the most relevant passages within each document, so a search agent reads less. On six deep-search benchmarks the agent used 41% fewer tokens and its accuracy rose from 39% to 52%. A reranker re-orders search results so the model sees the most relevant text first. Pricing is $0.05 per 1M tokens for Pro and $0.02 for Fast, on the Cohere API, Microsoft Foundry and Amazon SageMaker.
Humming is an open-source library of GPU kernels, the small programs that do a model's math, for quantized models that store numbers in fewer bits to save memory. One library covers many integer and floating-point formats on NVIDIA GPUs from Turing onward. In a vLLM test, switching a MoE model from Marlin W4A16 to Humming W4A8 raised end-to-end throughput by about 14%, 43% and cut mean time to first token by about 25%, 32%.
Ai2 released OlmoEarth v1.3, a satellite embedding model, with precomputed annual embeddings for the whole planet. The model reads a year of radar and multispectral satellite images and outputs one 128-dimensional vector, a compact numeric summary, for every 10-meter pixel; the first 64 dimensions also work on their own. Until now, using OlmoEarth meant training a model for each task, which required GPU access and stored imagery. A new browser tool, the OlmoEarth Embeddings Explorer, needs no sign-in.