LLM / Model Trends

NVIDIA released Nemotron 3.5 Bolt 74B-A9B 24GB, a model designed for local agent AI that fits within 24GB of memory.

• •NVIDIA released the open models "Nemotron 3.5 Super VL" and "Nemotron 3.5 Bolt 74B-A9B 24GB" on Hugging Face on October 15 (US time).
• •The new model is designed for local agent AI and operates within 24GB of memory.
• •A smaller model, "Nemotron 3.5 Bolt E4B 4GB", is scheduled for release on October 23.

Explains a technique to convert regular JavaScript code into a neural network without training, making it fully differentiable and runnable in a browser without a GPU.

• •It converts code into a neural network without a training process, ensuring it is exact and fully differentiable.
• •The design allows execution within a browser environment without requiring a GPU.
• •It directly processes regular JavaScript code as the conversion target.

The performance of Anthropic's lightweight model 'Claude Haiku 5.5' was verified across three Japanese tasks by comparing it with 'Sonnet 5.5' and 'GPT-6 Luna'.

• •The models compared were 'Claude Haiku 5.5', 'Sonnet 5.5', and 'GPT-6 Luna'.
• •It showed performance close to higher-tier models in creating boilerplate documents.
• •Responses fluctuated in tasks requiring answers based on an understanding of one's own stance.

PrismMLが3値化LLM「Ternary Bonsai 2」に2B/4B/9Bモデルを追加し、フル精度の9分の1以下のサイズで元モデルに近い性能を維持

• •10月15日(現地時間)にパラメータ数2B、4B、9Bの3モデルが追加された
• •モデルサイズはフル精度の9分の1以下でありながら元モデルに近い性能を保つ
• •Apache 2.0ライセンスでHugging Faceにオープンウェイトとして公開されている

TypeSafe AI released Jev 1.14.0, an automation-focused AI model that enables image input and context length expansion via experimental features.

• •TypeSafe AI announced the new version 'Jev 1.14.0' of its automation-focused AI model series 'Jev', known as System One Model, on October 14.
• •The model ID is 'jev-1.14.0' and is accessible via the alias 'jev-preview'.
• •The API pricing is $0.042 per million input tokens, with output tokens being free.

Liquid AI released the LFM2.5-24B-A2B inference model for AI agents as open weights, with API pricing starting at $0.10 per million input tokens.

• •Liquid AI announced the LFM2.5-24B-A2B inference model on October 13 and released open weights on Hugging Face.
• •The model is under the LFM Open License v1.0, requiring a commercial license for companies with over $10 million in annual revenue.
• •API pricing via Liquid Console is $0.10 per million input tokens, $0.05 for cached input, and $1.00 for output.

Ricoh developed the multimodal LMM 'Ricoh LMM_35B-A3B' based on Alibaba Cloud's Qwen3.6, enhancing Japanese reasoning performance via reinforcement learning.

• •Ricoh developed the multimodal large language model 'Ricoh LMM_35B-A3B' for high-precision understanding of Japanese documents and charts.
• •The model is based on Alibaba Cloud's Mixture of Experts (MoE) structure, 'Qwen3.6-35B-A3B'.
• •Ricoh utilized its own reinforcement learning and curriculum learning techniques to improve reasoning performance in Japanese.

Meta and Apple partnered to jointly announce Meta's latest open-weight AI model, Muse Spark Open, which runs locally on Apple silicon.

• •Meta and Apple partnered to announce the latest model from the major social media company.
• •The new model is an open-weight model named Muse Spark Open.
• •The model is designed to run locally on Apple silicon.

Meta announced the Muse Spark Open AI model, which is capable of running on two Apple M5 Ultra Mac Studios using the latest MLX framework.

• •Meta positions the Muse Spark Open AI model as the highest-performing US open-weight model to date.
• •The model is capable of running on two Apple M5 Ultra Mac Studios using the latest MLX framework.
• •Meta claims the model is the first in its class to support desktop execution.

Google is updating its AI search results to display inline links more frequently at the top of responses, moving away from traditional link lists to a customized view.

• •Google has updated AI Overviews and AI Mode to make inline links appear higher up in responses more often.
• •The new format provides answers in a customized view rather than a traditional list of links.
• •This change addresses concerns that the previous format might reduce transparency regarding sources.