DeepSeek V4 Flash on a Single AMD MI300X
An open-source project got the full 304-billion-parameter DeepSeek V4 Flash running on a single AMD MI300X GPU, with no multi-GPU sharding or weight offloading — the whole model fits in the card's 192GB of HBM3. The team fixed a mismatch between AMD's FP8 number format and the OCP-standard FP8 DeepSeek shipped upstream, tuned matrix-multiply kernels for the card's 304 compute units, and added speculative decoding, reaching 168.6 tokens/second on a single request and roughly 830 tokens/second aggregate under heavy concurrent load. It's a rare case of a frontier-scale open-weights model getting a ground-up AMD serving recipe instead of an NVIDIA port — at roughly half the hardware cost of an equivalent NVIDIA setup.
Sources & depth
FLUX 3 Video
Black Forest Labs launched FLUX 3 Video, generating up to 20-second clips in HD or Full HD from a text prompt or a starting image, with native dialogue, sound effects, and ambient audio produced alongside the picture rather than dubbed on afterward. It handles lip-synced dialogue in 13-plus languages and supports multi-scene sequences plus a fast draft mode for iterating before a final render. BFL says human raters preferred its output over rival video models on both text-to-video and image-to-video prompts; it's live now through the BFL API and select partners, with no pricing announced yet.
Sources & depth
Shieldstral
Mistral released Shieldstral, a 3-billion-parameter safety classifier that scores text and images against safety policies written in plain language at inference time instead of a fixed set of moderation categories. Ask it "does this promote violence?" and it returns a single calibrated probability rather than a canned label, so the same model can be repointed at a cybersecurity research tool's policy or a mental-health platform's policy without retraining. Mistral says it matches or beats guard models up to seven times its size on text safety, refusal detection, and multimodal benchmarks; it's released under Apache 2.0 and runs on a single 16GB GPU.
Sources & depth
A unified API for AI model routing
Google Cloud's API Gateway now offers model routing in public preview: a serverless layer that takes OpenAI-compatible requests and routes them to whichever backend model a developer configures — Gemini 3.5 Flash-Lite, Claude Opus 4.7, and OpenAI's open OSS-GPT-120B are supported at launch — without hardcoding endpoints or running a manual proxy. Routing rules live in an OpenAPI spec, and the Gateway transcodes each request into the target model's native schema automatically. It plugs into Google's Gemini Enterprise Agent Platform for governance, giving teams a vendor-neutral front door for swapping models without touching application code.
Sources & depth
Deploy local agents everywhere with LFM2.5-2.6B
Liquid AI released LFM2.5-2.6B, a 2.6-billion-parameter model built to run autonomous agents entirely on-device — phones, laptops, edge hardware — with no cloud round-trip. It fits under 2.5GB of memory, handles a 128K-token context, and hits 220 tokens/second on an Apple M5 Max CPU; Liquid trained it with reinforcement learning inside real agent harnesses (OpenClaw, Hermes Agent) rather than supervised fine-tuning alone, and says it beats models four to eight times its size on instruction-following and tool-use benchmarks.
Sources & depth