Hunyuan Hy4 preview is live: 770B total / 49B active, 1M context. But ¥6 in / ¥18 out per M tokens — that's a real jump from Hy3's pricing.
Hunyuan Hy4 preview is live: 770B total / 49B active, 1M context
More posts
hy-4 and glm-5.3-flash now work through the reverse proxy
. workbuddy CLI picked up both too. Haven't touched hy-4 yet, but glm-5.3-flash has been my daily driver for a day. Follows instructions tightly, fast, and just a hair pricier than muse-spark-1.2-contributor-free. Best value model right now, IMO.
grok-bot's server is worth way more than the bot on it.
The bot eats through its weekly quota in no time, but at least it doesn't reset the device. Server's in San Jose and the IP quality is trash. I set up a game accelerator, got 400ms. Bare direct connect to US-west games is only 300ms. The "accelerator" was slower. Also tried spinning up a VM. The box is itself a KVM VM with no hardware virtualization, so it's software emulation only. Too slow to bother.
Meituan shipped LongCat-2.5
. 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.
SenseTime open-sourced SenseNova Loop-9B-Preview.
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.
