June 2026 — Research-backed analysis of new model architectures and inference breakthroughs.
A 3B model matches Claude Opus and DeepSeek V3.2 on math and code. Here's how WeiboAI's Spectrum-to-Signal pipeline makes it possible.
A family of coding models that learn to write their own agent harness during training. The 397B variant beats Claude Opus 4.7 on Terminal-Bench.
A startup claims to have escaped the quadratic attention bottleneck. 12M tokens, 64x less compute, frontier-level retrieval accuracy.
A 9B model fine-tuned on 500M+ tokens of Claude Mythos traces gains +34 MMLU and native tool calling. Apache 2.0 licensed.
Speculative decoding done right: draft model guesses, target model verifies, confidence scheduling saves GPU. Already in production.