NVIDIA's open Nemotron 3.5 Transcribe: self-host 40-language ASR
A 30B model (4B active) with 8-speaker diarization now leads open ASR — alongside fresh open image/video weights and OpenAI's Decisions API GA.
Copy markdown
Self-host ASR that outscores the paid APIs
NVIDIA open-weighted Nemotron 3.5 Transcribe (30B total, ~4B active per token), which tops the Hugging Face Open ASR leaderboard at a 5.33% word error rate while covering 40 languages, translating 31, and labeling up to 8 speakers across 13 languages. Weights are on Hugging Face and run on vLLM today, so you can swap paid transcription plus diarization into meeting-notes or dictation tools for the cost of a GPU.
A 4K image model that fits in 8GB of VRAM
Black Forest Labs released open weights for FLUX 3 Image Klein — 5.6B params with native 4K generation, bounding boxes and pixel-perfect editing — and it runs on a single consumer GPU with as little as 8GB of VRAM. Free commercial use is capped to smaller businesses, but for local projects you get near-frontier image generation with zero API spend.
Open 4K video with multishot, straight into ComfyUI
Lightricks' open-weight LTX-2.5 is a 22B audio+video model doing native multishot and 4K HDR, with Fast/Pro/Ultra modes and a ready ComfyUI workflow. Chain it after Klein and you can take a storyboard to a finished short entirely on local hardware.
A $0.10 routing brain for your agents
OpenAI's Decisions API, powered by GPT-6 Luna, is generally available at $0.10 per million input tokens with no output-token charge, now handling up to 400 questions per request plus native JSON and image inputs. It's purpose-built to classify, route, prioritize, or risk-score an agent's next action before it runs — a cheap guardrail layer for any agent pipeline.
The #1 model you still can't buy
Google's Gemini 4 Carbon tops Artificial Analysis's Intelligence Index at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57, but access is gated to Fairwind cyber-defense partners with no public API ID or GA date. The top score you can actually ship on today is Opus 5.5; build against gemini-4-argon and swap the model ID if Carbon opens up.