Rumble-2-Flash brings a 0.5B multilingual speech model to open tooling
Reson8 has released Rumble-2-Flash on Hugging Face under Apache 2.0, offering a 0.5-billion-parameter speech-recognition model for more than 40 languages with native Transformers and vLLM support. Its compact size and claimed throughput make local transcription attractive, but language-level accuracy, call-centre robustness, batching assumptions and privacy controls need independent validation.
Amazon Bedrock brings explicit prompt caching to open-weight models
Amazon Bedrock now supports explicit prompt caching for compatible open-weight models, alongside its existing implicit mode. Developers can mark reusable prompt prefixes, inspect cache reads and writes through the Chat Completions API and reduce repeated processing in agents. Savings depend on model-specific pricing, prefix stability, minimum lengths and cache lifetime.
NVIDIA packs Nemotron 3.5 Bolt into a 24GB local model
NVIDIA has released Nemotron 3.5 Bolt 74B-A9B-24GB, a native two-bit multimodal model designed to fit within 24GB of memory. It combines a 74-billion-parameter, nine-billion-active architecture with a 256K context window, image and video inputs, tool calling and local vLLM deployment, targeting capable agents on a single workstation-class GPU.
Google Cloud gives Reflection Beam day-one agentic serving support
Google Cloud has published day-one deployment support for Reflection’s open-weight Beam model, pairing AI Hypercomputer infrastructure with llm-d, vLLM, NVIDIA Dynamo and SGLang. The provider reports substantial throughput and latency gains for long-running coding agents, while offering recipes for A4X Max and G4 systems through its infrastructure repositories and Model Garden.
Mistral Large 4 brings a trillion-parameter open-weight model to preview
Mistral Large 4 is now in public preview as a one-trillion-parameter, natively multimodal model with 49 billion active parameters. Mistral positions it for coding, scientific computing, enterprise knowledge work and agentic tasks, with open weights due later this month and API access already available through Mistral Studio.
NVIDIA releases an open telecom model and fine-tuning recipe
NVIDIA has announced Nemotron 3 Large Telco Model, a 30-billion-parameter open model fine-tuned by AdaptKey for network configuration and customer-incident workflows. The 6 October article also releases an end-to-end NeMo fine-tuning recipe. Operators gain more customisation and deployment control, but still need strong privacy engineering, evaluation and production governance.