Velvet2-3B-Speech expands multilingual speech AI
Almawave Labs' Velvet2-3B-Speech handles speech recognition, translation, summarization, emotion recognition, and spoken question answering across six European languages. The open-licensed release follows the earlier Velvet speech model and adds the summarization and question-answering tasks. — huggingface.co
Nemotron quantizations tighten fidelity at small size
New Nemotron 3.5 Bolt E4B quantizations reproduce the original model's output more closely than same-size rival files, while running on a small GPU or a CPU. — huggingface.co
Russian-tuned draft head speeds up Qwen3.8-27B
A fine-tuned draft head for Qwen3.8-27B — a small model that proposes token sequences a larger verifier checks — is tuned toward Russian-domain data and lifts output throughput on retail and coding workloads at practical concurrency, while trailing the baseline on some general English benchmarks. — huggingface.co
NorthCUA-Mini drives desktop and browser automation
NorthCUA-Mini, a 35-billion-parameter open-weight model, turns screenshots and interaction history into actions for desktop and browser automation, pairing with a dedicated runtime harness that executes GUI actions. — huggingface.co
GemmaTranslate-v3-12B now ships in compact local builds
New quantized builds of the GemmaTranslate-v3-12B translation model run fully on local hardware across quality tiers from compact to near-original, needing no cloud API. — huggingface.co