Inception Labs launches Mercury 2, the first diffusion-based reasoning language model, generating over 1,000 tokens per second on Blackwell GPUs at a fraction of the cost of conventional autoregressive models.
Today's arXiv picks: a state-machine framework that makes GUI agents 12x cheaper, a training method that forces chain-of-thought to be honest, and a KV cache system that matches full quality at 1% the memory.
DeepSeek V3.2 is a 671B-parameter MoE model activating 37B per token that delivers frontier-class reasoning and coding at the lowest API price in the industry - $0.14/$0.28 input, $0.42 output per million tokens.
Microsoft's 14B dense transformer that consistently beats models 5x its size on MATH and GPQA, available under the MIT license for unrestricted commercial use.
NVIDIA's hybrid Mamba2+MoE model packs 31.6B total parameters but activates only 3.2B per token, delivering frontier-class reasoning with 3.3x the throughput of comparable models on a single H200 GPU.
Anthropic's flagship model leads on agentic coding, enterprise knowledge work, and long-context retrieval with a 1M-token window, 128K output, and agent teams at $5/$25 per million tokens.
