More posts
Near-Fable-5 level intelligence at half the price. 43.3% on Frontier-Bench — more than 2x Opus 4.8, and cheaper. On CursorBench (high), it's within 0.5% of Fable 5. And it's getting more AGI by the day — takes initiative like a real person, reflects and verifies its own work. Other models just patch bugs. This one digs out the root cause. For long tasks it takes over your dev environment and runs monitoring itself — only pings you when it genuinely can't decide.

. 1M context, 65K output, takes image/audio/video. Google says it uses fewer tokens than 3.5 Flash, writes better code, and reasons better across modalities. 3.5 Pro is still MIA. My bet: they skip 3.x and drop 4.0 Pro to chase the top models.

. 1.6T params, ~48B active, native multimodal, aimed at long-horizon agents, trained start to finish on Chinese hardware. Images get quantized into multi-level discrete tokens, and that's where Visual Coding and 3D pick up. On the RL side, failed trajectories get turned back into verifiable training tasks. W4A4 inference, 43% faster prefill. Meituan's own evals put it between Claude Opus 4.8 and Opus 5.

It stacks layers instead of parameters. The decoder has 32 layers, but layers 4 through 27 run twice with shared weights, so a forward pass does 56 layers of work while the checkpoint stays at 9B. Their own eval numbers: 71.38 on BIRD-Dev and 13.72 on AutomationBench, both above Qwen3.5-27B. Terminal Bench is still behind.
