Tencent open-sources Hy4: 770B MoE, 1M context, Apache 2.0
Weights and an FP8 build are already on Hugging Face, or call it at ~$0.83/M on Vercel — the fifth open-weight frontier drop in nine days.
Copy markdown
770B total, 49B active — under Apache 2.0
Tencent's Hy4 Preview is a 770B-parameter MoE (256 routed experts, 8 active plus 1 shared) that fires only 49B per token, with a native 1M-token context — not RoPE-scaled. It ships under Apache 2.0 with no field-of-use or geographic limits, so you can build on it commercially.
It leads open models on SWE-Bench Pro
Hy4 posts 65.7 on SWE-Bench Pro, 82.9 on SWE-Bench Multilingual, 85.4 on Terminal-Bench 2.1, and 92.3 GPQA Diamond — front of the open pack on agentic coding. Tencent's own eval calls it roughly interchangeable with GLM-5.3 and Kimi K3, but this one you can self-host.
$0.83 in, $2.50 out — with a 20x cache discount
It's already live on Vercel's AI Gateway at $0.834/M input and $2.501/M output, and cached input is just $0.042/M — a 20x break that adds up in agent loops replaying the same context. That's roughly 40% under GLM-5.3 on input, though DeepSeek V4 Pro is still cheaper.
The catch: it over-thinks, and it's slow so far
Tencent admits Hy4 'over-verifies its own work' and over-reasons on hard tasks; early hosted throughput was around 36 tokens/sec at P50 with ~86% availability — rough for latency-sensitive agents. It defaults to 'high' reasoning, so switch to 'no_think' when you don't need the deliberation.
Weekend build: drop the RAG, feed the whole codebase
A native 1M window plus open weights means you can pour an entire mid-size repo or a book-length corpus into one prompt and skip the retrieval plumbing. Run the FP8 build on vLLM or SGLang behind an OpenAI-compatible endpoint; its sparse-attention IndexCache keeps long-context serving affordable on your own GPUs.
Elsewhere: five open-weight drops in nine days
Hy4 caps a run — DeepSeek V4-Flash-Vision (Aug 21), MiniMax M3 (Aug 24), GLM-5.3-Flash and Qwen3.8-Flash (Aug 26). 1M context, multimodality, and permissive licenses are now table stakes in the cheap tier; the costs that actually vary are reasoning effort, cache hit rate, and provider choice — one workload showed an 11x price spread across hosts.