AI Briefing#llm
The latest AI and developer news about #llm, with the original source and a short summary.
The latest AI and developer news about #llm, with the original source and a short summary.
The ThinkingBox benchmark evaluates 9 models across 10,140 trials per model, showing that high pass@20 scores do not guarantee consistent task completion.
The 0.8B Jeff model determines whether to think on each turn, reducing Qwen3.8-27B's task time by 32%.
A technique that improves inference accuracy by using the first iteration's prediction as a reference for loop transformers without additional training, reducing theoretical compute by up to 48.2%.
Google announced the Gemini agent, a universal AI for work that operates across Google Workspace and Microsoft 365, featuring multi-agent orchestration and persistent cloud execution.
Scheduled for release in October under the Apache 2.0 license, this sparse MoE architecture emphasizes inference compute efficiency.
The new model costs around 75% less than Haiku 4.5 and introduces adjustable effort settings for the first time in the Haiku class.
The model introduces effort levels and adaptive thinking, supporting low to max settings to control token usage and reasoning depth.
The model is free for Pro and Enterprise teams with purchased AI Gateway credits but lacks Zero Data Retention support.
A new interactive tool estimates GPU requirements for serving large language models like Llama 3.3 70B and DeepSeek V4 Pro, using benchmark data for H100 and B200 hardware.
The 3.35B parameter model achieves over 93% in-language reasoning rates by using a three-pillar data mixing strategy.
The new 7B model achieves 84.83% on the Alyah benchmark and significantly outperforms larger general-purpose models in dialect fidelity and cultural understanding.
Analysis of 1,250 arXiv papers establishes a taxonomy for self-improvement loops at deployment and training time.
No open model has surpassed Jev, but Cygnet recorded the top score on JevBench.
The single-model approach achieved +4.53% Active Users and +6.30% Total Listening Time in a 7-day A/B test on smart speakers.
An empirical study by researchers from SMU and Meta AI found that the Jev model for recommendation reranking has up to 17x lower latency than pointwise Qwen and lies on the quality-latency non-dominated frontier in 11 out of 12 experimental conditions.
Achieved an AA Index score of 24.1 with an MoE architecture that activates only 3B of its total 35B parameters
Aleph Alpha has released Kolibri, an open-weight English-German Mixture-of-Experts model with 78.1 billion parameters and support for 1 million token contexts, available under the Apache 2.0 license.
The 8B model, based on Qwen3-8B, generates reports in an average of 51.1 seconds compared to 178.5 seconds for the previous Claude-based Thinking mode.
The 26B parameter model achieves an 85.71% overall accuracy across 11 benchmarks, outperforming its base model Qwen3.8-27B and the Jev baseline.
The new SDK allows Python developers to use Jev, a universal classifier model, via the experimental evaluate() function.
Ai2 releases Olmo-core 3, an open-source training framework designed for large MoE models, achieving 2.7x throughput improvement in preliminary tests and benchmarked at over 1 trillion parameters.
Both studies achieved an F1 score of 90.14% on the ADReSSo benchmark, surpassing the previous best record
The model features a 262,144-token context length and learns to improve solutions through iterative test-time reflection.
Fireworks AI has released 'Ember-1' as a research preview, an inference-optimized model based on Kimi K3 that reduces token usage by approximately 35–40% without quality loss in benchmarks and real-traffic tests.