🚀 RELEASES & LAUNCHES
- Tencent open-sources Hy4 preview MoE with 770B parameters
Tencent open-sourced Hy4 preview, a 770B-parameter MoE model.- Built for coding agents and productivity work across software, gaming, finance, security, and research.
- Hy4 preview has 770B total parameters, 49B active per token, and a 1,048,576-token context window.
- The model is available through WorkBuddy, CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub, OpenRouter, and Hugging Face.
- API pricing is $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cache-hit tokens.
- Tencent’s internal test scored Hy4 preview 2.99/4.00 across 203 engineering tasks, while reported inference throughput improved 31.8% over baseline.
🔬 RESEARCH
Micron says HBM uses 3× the wafer area of DDR5 per bit
Micron reported that HBM requires roughly three times the wafer area of DDR5 for equivalent capacity.- The finding applies to AI accelerators, data-center GPUs and other systems using HBM.
- At Hot Chips 2026, Micron compared HBM4 dies with 256 banks to DDR5’s 32 banks.
- Parallel data paths, power delivery and TSVs drive HBM’s roughly 3× wafer-area requirement.
- Micron said the ratio is not improving, tightening the relationship between AI memory demand and commodity DRAM supply.
Sources 📎 Coverage: tomshardware.com · 👁 via r/LocalLLaMAReverse-engineered AX8850 runtime runs GGUFs 1.5× faster than vendor stack
A llama.cpp backend now runs GGUF models on the AX8850 NPU without conversion.- Targets the M5Stack LLM-8850 card with Axera’s AX8850 NPU, hosted by a Raspberry Pi 5.
- Runs Qwen3-0.6B GGUFs directly through llama.cpp without compiling models into vendor engines.
- The AX8850 card provides 24 TOPS and 8GB of LPDDR4x memory.
- The reverse-engineered runtime is reported to run 1.5× faster than the vendor runtime’s 13.5–14.5 tokens per second.
- Direct GGUF support removes the vendor stack’s per-model conversion requirement for local NPU deployments.