LLM / Model Trends

Kolibri, a MoE model supporting English and German, has been released, specializing in mission-critical sovereign operations.

• •It is a Mixture-of-Experts model with 78B total parameters and 3B active parameters.
• •It supports a context length of up to 1 million tokens.
• •It is provided as open weights under the Apache 2.0 license.

An article exploring the hypothetical impact of AI agents operating at one million tokens per second, noting that bottlenecks shift to testing, evidence, and judgment.

• •The speed of one million tokens per second is a hypothetical scenario, not a measured value.
• •The article examines the changes that would occur if AI agents generated tokens at ultra-high speeds.
• •It analyzes that as processing speeds increase, testing, evidence, and judgment become the new bottlenecks.

AstaBrief, an 8B parameter open-weight model for generating scientific reports with citations, is available in Asta's fast mode.

• •AstaBrief was released as an 8B parameter open-weight model.
• •The model is specialized in generating scientific reports with citations.
• •It is available in Asta's Fast mode and can also be downloaded and run on self-hosted infrastructure.

An explainer covered ds4, a C-language inference engine by Redis founder antirez, including documentation and benchmarks for running models like DeepSeek V4 locally.

• •ds4 is a C-language inference engine supporting DeepSeek V4/V4.1, Qwen3.8 Flash Next, and GLM 5.x.
• •It supports operation on different hardware environments such as Metal, CUDA, and ROCm.
• •Documentation, benchmark results, and setup instructions for ds4 have been published.

A report on using GLM 5.3 Flash for coding for one month detailed 2 billion tokens of usage and factors affecting model selection and costs.

• •There was a total of 2 billion tokens of AI usage during one month of coding.
• •The report analyzed elements influencing model selection criteria and cost structures.

The calibration of AI model Jev's probabilistic predictions was verified by generating physically realistic probability distributions.

• •Jev's performance was tested by requiring the generation of physically realistic probability distributions.
• •The evaluation assessed how well the model's probability outputs align with actual physical phenomena.

OpenAI announced that ChatGPT will be reset at 2:00 AM JST on October 4 due to performance delays with GPT-6.1 Sol.

• •OpenAI's Thibault Sottiaux announced the reset on X on October 2.
• •The reset is intended to address performance delays associated with GPT-6.1 Sol.
• •The scheduled time is 2:00 AM JST on October 4 (October 3 US time).

ELYZA, a KDDI subsidiary, released a free, fully domestic AI model with enhanced performance based on the open model LLM-jp-4.

• •ELYZA, an AI development company in the KDDI group, announced the AI models ELYZA-Thinking-1.0-llm-jp-4-33b and ELYZA-Thinking-1.0-llm-jp-4-32b-a3b.
• •These models are based on the open LLM-jp-4 series developed by the National Institute of Informatics.
• •Post-training was conducted on the base models to enhance their performance.

Decision models that return classification or routing judgments without generating text are gaining attention, following Typesafe AI's 'Jev'.

• •Decision models that return only classification or routing judgments without generating any text are gaining momentum.
• •Typesafe AI's 'Jev' became the catalyst for this category.
• •Compatible models and supporting implementations have been emerging in succession following 'Jev'.

Inference engine 'Strata' released to run Alibaba's Qwen3.8-Flash-Next on gaming PCs at 60-95 tokens per second.

• •Qwen3.8-Flash-Next, with 125B total and 6B active parameters, runs on a single-GPU gaming PC.
• •The 'Strata' inference engine enables high-speed execution at 60-95 tokens per second.
• •Developer Niko1221 released it on GitHub under the MIT license.