Tencent open-sources Hy4: a 770B coding model that ties Opus 5
Apache-2.0 weights, 1M-token context, and coding scores level with the frontier — but 770GB of VRAM to self-host. Plus Codex can now rewrite MCP results.
Copy markdown
The drop: 770B open MoE, 49B active, 1M context
Tencent open-sourced Hy4-preview under Apache 2.0 — no field-of-use or geography clauses — with full and FP8 weights on Hugging Face, ModelScope and GitCode. It's a 770B-parameter Mixture-of-Experts (256 routed + 1 shared expert) that fires just 49B parameters per token, with a 1M-token context window.
Coding scores that reach the frontier
Hy4 posts 85.4 on Terminal-Bench 2.1 — level with Claude Opus 5 — and lifts DeepSWE from Hy3's 28.0 to 64.3, with 65.7 on SWE-Bench Pro and 74.1 on Toolathlon-Verified. For agentic coding you now have an openly-licensed model in the same league as the closed frontier.
The price — and why you probably can't self-host it
Hosted API runs about $0.83 per million input tokens and $2.50 output ($0.04 cached) via Tencent Cloud TokenHub and OpenRouter, with two weeks free through WorkBuddy and CodeBuddy. But the FP8 weights alone need ~770GB of VRAM — past a single 8×H100 node — so 'open weights' here means a hosted endpoint or 8×H200 / multi-node vLLM, not your laptop.
Elsewhere: Codex 0.151.0 lets extensions rewrite MCP tool results
OpenAI's Codex CLI 0.151.0 now lets extensions inspect or replace MCP tool results before they reach the model — a clean hook for redacting secrets, caching, or reshaping tool output. It also adds configurable grace periods for discovering tools from optional MCP servers.
Elsewhere: Claude Code 2.1.251 patches symlink and plugin path-traversal holes
The 2.1.251 release fixes file-tool symlink vulnerabilities and plugin path-traversal issues, and adds PreModelSwitch / PostModelSwitch hooks plus live Remote Control streaming for foreground subagents. Worth updating if you run untrusted plugins.
Elsewhere: Gemini 3.5 Transcribe ships a dedicated speech-to-text API
Google made gemini-3.5-transcribe GA — high-accuracy STT across 85+ languages with speaker diarization, word-level timestamps and custom vocabulary — plus a Live variant that streams transcription over WebSockets. Enough to build a meeting notetaker or live-caption tool this weekend.