llama-modchip runs a 284B DeepSeek on 64GB of RAM, no GPU
An Apache-2.0 llama.cpp patch fits a near-frontier MoE on a desktop at ~10 tok/s — while Google doubles TPU inference density and Goldman pulls Claude Code.

Copy markdown
A 284B model on a 64GB desktop
llama-modchip quantizes a 284B DeepSeek-V4-Flash MoE to ~1.71 bits (56.6 GiB) and runs it entirely in CPU RAM — no GPU, no swap — at ~43 tok/s prefill and 9–15 tok/s decode on a single Ryzen 9. It's Apache-2.0, ships a patched llama.cpp server plus prequantized GGUFs on Hugging Face, and stock llama.cpp, Ollama and LM Studio reject the format, so you need its binary.
Google doubles TPU inference density
Google launched TPU 8i Max — two 8i chips per module, doubling compute, SRAM and HBM bandwidth — aimed at the memory wall in long-context and multi-agent inference, plus an 8tx variant with hardware-level confidential inference for regulated teams. It also open-sourced TPU Sync, which tiers KV cache off-chip to cut time-to-first-token by 90%+, with vLLM and SGLang backends. No cloud pricing yet.
Goldman pulls Claude Code from its traders
Goldman Sachs restricted Claude Code and GitHub Copilot to roughly 12,000 engineers, cutting off the traders and bankers who'd adopted them — a signal that big regulated shops are reining in who gets agentic coding tools, and how. Multiple outlets reported it; Goldman hasn't spelled out its reasoning.