Tencent released Hy4-Preview on Hugging Face and ModelScope on Thursday — a 770 billion parameter Mixture-of-Experts model activating 49 billion parameters per token across 78 layers with 256 routed experts. The model's custom attention mechanism, Gated DeepSeek Sparse Attention with IndexCache, enables cross-layer sparse index reuse and represents a notable architectural departure from standard transformer attention. Benchmark scores — 92.3 on GPQA Diamond, 65.7 on SWE Bench Pro, 64.3 on Deep SWE — place it at the competitive open-source frontier tier, particularly for software engineering tasks. One million token context window, Apache 2.0 licence, available in standard and FP8-quantized versions. Tencent explicitly describes the release as a preview with "real headroom left in both pre-training and post-training." The timing — released into a week in which the primary open-weight hosting and model development infrastructure is being absorbed by Nvidia and Stripe via the Hugging Face and Poolside acquisitions — is worth noting. Hy4-Preview lands on a commons that may, within months, no longer be a commons.
LLMs
Google Releases Gemini 4 Argon — Tops Frontier Benchmarks and Targets Autonomous Cybersecurity
30 September: Google's Gemini 4 Argon leads the Vals AI benchmark index over OpenAI GPT-6 Astra and …