Quiet Day, Loud News: Xiaomi's MiMo, Cheaper Training, and Computers That Dream

Xiaomi's MiMo-V2.5 and the DFlash speed trick

Xiaomi quietly uploaded a new set of weights for MiMo, a large language model with over 300 billion parameters, alongside a variant called DFlash. Rather than a smaller, weaker model, DFlash is a speculative-decoding-style acceleration that preserves the model's output while roughly doubling generation speed on consumer GPUs. It's a win for capable, private, local AI that needs no cloud.

Prime Intellect makes agent training cheap

Prime Intellect shipped a redesigned toolkit for training agentic AI. The key change: rollout traces are now stored as message DAGs, so each message is kept once instead of being copied into the full history every turn. That shifts storage growth from quadratic to linear and let the team train a 100B reasoning model on 40-turn software-engineering tasks in under two days on just six machines.

Hugging Face and vLLM end double implementation

Clement Delangue highlighted that Hugging Face Transformers models can now run in the vLLM serving engine at native speed. Historically every architecture was built twice, once for research and once for fast serving. Removing that duplication means new open models reach users far faster.

Smarter quantisation and 4-bit serving

A new quantisation method claims to beat NVIDIA's ModelOpt by assigning precision layer by layer, squashing what it can and protecting what it can't. Combined with growing momentum behind 4-bit (NVFP4/FP8) serving, it means bigger models on cheaper hardware for pennies.

China's chip push and OpenRouter surge

Reports say DeepSeek is developing its own AI accelerator in response to export controls limiting NVIDIA access. Meanwhile Chinese models from DeepSeek, Xiaomi and MiniMax took the entire top five on OpenRouter's usage rankings. As the community quipped, it's hard to compare benchmarks but easy to compare bills.

Grok Build's privacy controversy

Security researchers alleged xAI's Grok Build coding tool was uploading entire repositories, including private code and secrets, to a cloud bucket. xAI pointed to zero-data-retention options and a privacy command to disable retention and delete synced data, but concerns about default behaviour remain, a reminder to know what agent tools transmit.

Teaching models to sleep and dream

Researchers proposed giving language models a sleep phase to consolidate short-term into long-term memory and a dreaming phase for self-improvement, tackling the fact that today's models are frozen after training. It dovetails with Rich Sutton's new venture, Oak Lab, chasing animal-like intelligence that learns from experience.

3D on your phone with Modelr

Modelr, an open-source Apple Silicon and iOS app, turns a single photo into a 3D model locally in under 20 seconds using less than 2GB of RAM, built on Tencent's Hunyuan3D tech ported to Apple's MLX framework. Licensing limits commercial use, but it's a striking glimpse of on-device generation.