Nemotron 3.5 Transcribe: NVIDIA's open 40-language ASR, laptop-ready
Free, commercial-use, 8-speaker diarization — plus Liquid's six on-device 2B models, a looming Gemini 2.5 reroute, and FLUX.3's open image editor.

Copy markdown
40 languages, 8 speakers, one chip
NVIDIA open-weighted Nemotron 3.5 Transcribe — ~30B total but only ~4B active per step — transcribing 40 languages, translating 31 to and from English, and splitting up to 8 speakers. It tops Hugging Face's open ASR leaderboard at 5.33% English word error rate and runs from Jetson boards and laptops up to Blackwell, free for commercial use on pre-recorded audio up to an hour long.
Liquid's six 2B models run fully offline
Liquid AI released LFM3: six open-weight 2B-class models spanning text, vision, audio, omni, and decision-making on one shared backbone, with a 256K-token context. The text model fits in about 1GB at 4-bit — small enough to fine-tune a specialist and ship it on-device with no new hardware.
Report: your Gemini 2.5 calls reroute Dec 1
Google's API release notes reportedly set December 1 to retire gemini-2.5-pro and gemini-2.5-flash, silently routing requests to gemini-3.8-flash and gemini-3.5-flash-lite at the replacements' prices and defaults. If you call 2.5 in production, test now — dropped sampling params and stricter function-calling can break requests, not just reprice them.
FLUX.3 Image Klein ships open weights
Black Forest Labs' FLUX.3 Image Klein is a ~5.6B open-weight model for multi-reference image editing and composition, light enough to run locally and drop straight into ComfyUI — a weekend unlock for consistent-character edits and product compositing.