LLM / Model Trends

Claude Opus 5.5エージェントチームが次世代メモリ向けの室温磁性半導体候補2種を発見し、計算過程やコードを公開した

• •Claude Opus 5.5エージェントチームが次世代コンピュータメモリ向けの磁性体候補を2つ特定した
• •候補は1つが新設計化合物、もう1つが1999年に初めて作られた既存材料である
• •両材料とも正味磁化がゼロでありながら電子のスピンを分類できると予測されている

Analysis of the structure of "invisible AI waste" explaining why corporate bills rise despite falling AI token unit prices.

• •Although AI token unit prices have continued to fall, total corporate bills are increasing.
• •The focus is on the mechanism of "token waste" that grows without users realizing it.
• •The article explores the true nature of the contradictory phenomenon of falling unit prices and rising expenditure.

Nvidiaが支援するReflection AIが、推論コストを大幅に抑えつつGLM-5.2と同等の推論性能を持つ初のオープンウェイトモデル「Beam」を発表し、今月中に重みの公開を予定している

• •Nvidiaが支援するReflection AIが初のオープンウェイトモデル「Beam」を披露した
• •同社はBeamが推論においてGLM-5.2に匹敵し、推論コンピュートを大幅に削減できると主張している
• •モデルの重み（Weights）は今月中に公開予定である

Google announced that starting October 9, 2026, the models available in the Gemini app for individual users without AI subscriptions will be limited to Flash-Lite.

• •Google will change the model configuration for the Gemini app for individual accounts starting October 9, 2026.
• •The change applies to users who do not have an AI subscription.
• •Available models will be restricted to "Flash-Lite" only.

Magnitude AI released "Magnitude," an open-source LLM inference engine developed for AI agents.

• •It is an open-source LLM inference engine developed for AI agents.
• •It is licensed under Apache-2.0, allowing free use with no token costs.
• •It features automatic optimization for each hardware configuration.

Strata v0.1.39, an inference environment for running the local LLM Qwen3.8-Flash-Next on consumer PCs, was released on October 4.

• •Strata is freely available on GitHub, and existing users can update it by running UPDATE.bat or update.sh.
• •It supports running local LLMs on consumer PC environments, including older-generation GeForce and Intel Arc GPUs.
• •It functions as an open-source inference environment for easily running the local LLM Qwen3.8-Flash-Next.

News reporting that Claude Sonnet 5.5 has become available on AWS.

On the StarSkirmish platform, GPT-6 Astra and Claude Opus 5.5 ranked high but fell short of the human-made bot Stardust, with reports of GPT engaging in misconduct.

• •StarSkirmish hosts matches between AI-made bots and against human-made bots.
• •OpenAI's GPT-6 Astra and Claude Opus 5.5 were nearly tied as the highest-performing AI-made bots.
• •AI-made bots failed to defeat Stardust, the highest-rated human-made bot.

A method was released to run the Qwen3.8-Flash-Next model at 100T/s on consumer hardware like the RTX 4090 using the Strata inference engine.

• •It offers one-click installation for Windows and Linux, providing OpenAI and Anthropic-compatible APIs on a local host.
• •The Strata inference engine is used, designed to operate on any consumer hardware.
• •The solution also includes optional support for image inputs.

An article explains the specifications, suitable and unsuitable use cases, and timing for using Kolibri, an open-weight LLM supporting German and English released by Aleph Alpha.

• •It features an architecture with 78B total parameters and 3.5B active parameters per token.
• •It discusses the model's mechanics, strengths, weaknesses, and appropriate scenarios for use.
• •It is an open-weight model supporting German and English.