LLM / Model Trends

Local LLM inference engine Strata now supports Ryzen AI Max/Max+ series, enabling Qwen3.8-Flash-Next operation with 128GB of memory.

• •Version 0.1.40 of the open-source local LLM inference engine Strata was released on October 6.
• •Experimental support for AMD's Ryzen AI Max/Max+ series was added.
• •The engine can now run Qwen3.8-Flash-Next by utilizing 128GB of memory.

Google announced image generation model Nano Banana 2.1, setting API image output fees to nearly half of the previous model.

• •The new model is based on Gemini 3.6 Flash and supports up to 1 million tokens.
• •Benchmark tests recorded scores surpassing previous models and the Pro version.
• •It will be gradually provided in Search's AI Mode and the Gemini app.

高速な実験、ローカル開発、イノベーションに最適化された小型のオープンソース意思決定モデル「Strands Decider 2B」が公開された

• •Strands Decider 2Bは小型のオープンソースモデルである
• •高速な実験、ローカル開発、イノベーションのために最適化されている
• •特定の用途に特化した意思決定モデルとして位置づけられている

Mistral AI released a preview of 'Mistral Large 4,' a 1.05-trillion-parameter MoE model, with open weights planned for release by the end of the month.

• •The MoE model has 1.05 trillion total parameters and supports image recognition and switching between reasoning and instruction following.
• •It demonstrated high performance in cybersecurity and coding, achieving a top record in vulnerability reproduction tests.
• •The model weights are scheduled to be released as open weights by the end of October.

Jev駆動型のSRE診断パイプラインのワークフローと、105件の診断中76.2%が合格した結果、成功と失敗のトレースを比較分析した記事

• •Jev駆動型のSRE診断パイプラインが105件の診断のうち76.2%を合格した
• •ワークフローの詳細と、成功および失敗したトレースの分析が行われている
• •インタラクティブな時間とコストの比較データが提供されている

Google has released a new model, Gemini 4 Argon, which supports output of up to one million tokens.

OpenAI revealed that it has published 722 mathematical results generated by its frontier models on GitHub, many of which include formal proofs using the Lean proof assistant.

• •722 mathematical results generated by internal frontier models have been published on GitHub.
• •Many of the results include formal proofs using the Lean proof assistant.
• •Referencing recommendations from the independent organization AGMAI, summaries of computational complexity and reasoning processes are also disclosed.

Explanation of how generative AI operates probabilistically and the hallucination risks arising from the lack of fact-checking capabilities.

• •Generative AI works by predicting the next likely word based on probability, which causes answers to fluctuate.
• •While it has become intelligent through massive data, the absence of fact-checking features creates hallucination risks.
• •Principles for enterprises to safely utilize AI are also mentioned.

The article explains the proposal of UniEvo-VL, a self-evolving framework where multimodal models learn from self-feedback, and the use of self-correction in test-time computation.

• •It leverages the ability of integrated multimodal systems for generation and understanding to learn from their own feedback.
• •It utilizes the model's own self-criticism as privileged information without relying on external teacher models.
• •It improves the model through constructive self-correction feedback during the test-time computation stage.

An article sharing technical progress and achievements of AI in the field of mathematics.