Tencent's Hy4 adds vision and computer use — API live, weights soon
The 770B flagship now sees images and drives Blender, CAD and game engines at ~$0.83/$2.50 per 1M. Plus a Factory CLI fix and Snowflake's auto-router.

Copy markdown
A 770B model that sees and runs software, for cents
Tencent shipped the full Hy4: a 770B / 49B-active MoE with a 1M-token context that now takes image input and operates GUI and command-line apps — it was trained to drive Blender, CAD and Unreal/Unity/Godot. It's live today via OpenRouter, OpenCode and Tencent Cloud at ~$0.83 in / $2.50 out per 1M tokens, with open weights promised "in the coming days." On Tencent's own 579-task finance eval it tops open-weight rivals like DeepSeek-V4.1 and Kimi K3.1 Pro — self-reported, so verify before you switch.
Factory's CLI stops choking on big tool sets
Factory AI shipped CLI and Desktop v1.2.0: non-interactive runs no longer slow to a crawl when hundreds of tools are loaded — the same tool-scaling wall Copilot's CLI hit at 512 functions. A window whose view crashes now reloads itself instead of going blank, and revoking an API key finally requires confirmation. If you run Droid headless in CI behind a fat MCP roster, the perf fix is the one that matters.
Snowflake's gateway now picks the model for you
At Expedition 2026, Snowflake added dynamic model routing to its Cortex AI Gateway: it auto-selects the cheapest model that clears your quality bar per call, with cost governance and agent-abuse detection built in, plus more open-weight models on tap. If your data and agents already live in Snowflake, it's a built-in version of the cost-routing layer teams usually wire up by hand.