69 picksLast 30 days 48 itemsCollected 216 items
NewGoogle Research has introduced an architecture that attaches a multi-token-prediction (MTP) head to the frozen Gemini Nano v3 model, speeding up on-device generation without changing the backbone weights.
Why it matters: Google attaches an MTP head to frozen Gemini Nano v3 and reports concrete methods and gains under on-device memory and energy constraints.
Google DeepMind has released DiffusionGemma, an experimental open-source model that uses text diffusion to generate entire chunks of text in parallel. It generates text up to 4× faster on specialized GPUs, exceeding 1,000 tokens per second on a single NVIDIA H100 and 700 tokens per second on an RTX 5090.
Why it matters: Google DeepMind's open text-diffusion model comes with concrete speed and hardware figures, helping readers assess the feasibility of local interactive inference.
Andrej Karpathy says Claude Fable 5 shares the same underlying model as Mythos but adds safety safeguards, and reaches state-of-the-art results by some margin on almost every benchmark.
Why it matters: Drawing on firsthand use, Karpathy calls Fable 5 a release worthy of a major version-number jump and describes concrete changes on long-horizon tasks.
Google DeepMind has released Gemma 4 12B, an agentic multimodal model aimed at laptops, positioned between the edge-focused E4B and the 26B MoE. It is the company's first medium-sized model with native audio input.
Why it matters: Gemma 4 12B removes multimodal encoders and feeds visual and audio input directly into the LLM, giving readers a view of hardware requirements for local multimodal agents.
In an ICLR paper, Google Research proposes ReasoningBank, an agent-memory framework that extracts high-level, structured reasoning memories from successful and failed experiences. Its code is open-source.
Why it matters: The paper describes a closed loop for extracting reasoning memories from successes and failures, with success-rate and step-count results on WebArena and SWE-Bench-Verified.
Google has released TurboQuant, a compression algorithm that quantizes KV cache to 3 bits without training or fine-tuning and without loss of model accuracy, reducing KV-memory use by at least 6×.
Why it matters: Google reports TurboQuant's algorithm details and benchmark results, helping readers assess the usefulness of KV-cache compression for long-context workloads.
Mistral AI has released Mistral Small 4, the next major version of the Mistral Small family and its first model to combine Magistral's reasoning, Pixtral's multimodal, and Devstral's agentic-coding capabilities in one model. It is licensed under Apache 2.0.
Why it matters: Mistral has combined reasoning, multimodal, and coding capabilities in one open small model, with adjustable reasoning effort and deployment requirements to help assess implementation costs.
Mistral has introduced several new features in Le Chat, including a preview of Deep Research mode, voice mode powered by the new Voxtral model, multilingual reasoning supported by the Magistral reasoning model, Projects for organizing conversations, and image editing in partnership with Black Forest Labs.
Why it matters: Mistral has brought deep research, voice, reasoning, and image editing into Le Chat at once, giving readers a view of the assistant's expanding capabilities.
Mistral AI has released its first reasoning model, Magistral, in two versions: the open 24B-parameter Magistral Small and the enterprise Magistral Medium.
Why it matters: Mistral's first reasoning model offers both open weights and an enterprise version, with AIME2024 results and a training paper available for comparison.
