Embedding a deterministic RISC-V CPU inside Qwen 3.5 weights
Divyansh Agrawal compiled a functional RISC-V core into Qwen 3.5 4B, using Gated DeltaNet layers as exact RAM for arithmetic and file storage without external tools.
Open-weight models, on-device AI and running models on hardware you control.
Divyansh Agrawal compiled a functional RISC-V core into Qwen 3.5 4B, using Gated DeltaNet layers as exact RAM for arithmetic and file storage without external tools.
Hugging Face introduces LFM3, a new generation of open-weight models optimized for on-device deployment across text, vision, and audio modalities.
A new policy proposal suggests granting US AI labs antitrust safe harbor in exchange for predictable open-weight model releases, aiming to reduce reliance on Chinese alternatives.
A new open-source Linux kernel module unlocks full BF16 and FP16 instruction rates on consumer Nvidia RTX 30-series cards, offering significant throughput gains for local LLM inference.
Moonshot launched Kimi K3.1 Pro, a 2.8 trillion parameter mixture-of-experts model with open weights arriving November 4, 2026.
A new Hugging Face Space lists which popular models run unmodified on AWS Trainium2, providing scripts and compatibility data for developers.
Hugging Face releases OlmoEarth v1.3, offering state-of-the-art open-weight satellite embeddings and precomputed global data to simplify geospatial AI tasks.
Hugging Face researchers show how replacing general-purpose vocabularies with cost-weighted, domain-specific tokens reduces memory bandwidth usage and latency in small language models.
Recursive launched Flint, a free open-weight coding model that runs locally on laptops. The company claims its automated Eureka Machine system helped build the model in under four months.
IBM has updated its open-weight Granite Speech models, introducing a new 2B parameter model for accurate multilingual transcription and translation, alongside faster and more robust versions of existi
Nvidia launched Kumo Relational, an open-weight model that predicts outcomes directly from relational databases without task-specific training or feature engineering.
DeepSeek releases DeepInfer, its production inference engine, supporting Nvidia Blackwell and Huawei Ascend chips to lower compute costs for developers.
TorchNeuron introduces a native PyTorch backend for AWS Trainium, enabling eager execution and broader model support without compiler-specific workarounds.
Deepseek released V4.1-Flash-1017, an open-weight model that outperforms larger rivals in coding benchmarks while significantly reducing per-task costs.
Remek releases basal-omni-1.0-ultra, a 31B decision model on Gemma 4 that handles text and images with calibrated probabilities for automated routing.