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 26.6B parameter model supports seven languages and offers natural-language control over transcription style and context biasing.
Perplexity releases open-weight Decider v1.3 and Flash v1.3, with v1.3 scoring 69.0 and Flash 63.6 on out-of-domain extraction tasks, surpassing GLiNER large v2.5 (33.6).
Precision Numerics releases Boole-32B-A2.5B-Preview, a 1.8-bit MoE model that is 9x smaller than BF16 and 2x faster than INT4.
Alibaba's new model is available only as a temporary preview scheduled to retire on November 25, 2026.
A community post confirms the Qwen4-27B model is accessible on Hugging Face.
Tether AI released BotPsy-0.8B, based on Qwen3.5-0.8B, demonstrating superior tool-use performance compared to 2B models.
Using modified training protocols on Qwen3.6-35B-A3B and gpt-oss-20b, the research demonstrates that models can perform serially dependent reasoning steps without visible chain-of-thought output.
The model uses a hybrid dense architecture with 18 linear attention layers and supports Korean and English tool calling.
The model is available for early access on QwenWork, Qoder, Qwen APP, and QwenStudio.
The new backend allows a single .pte file to run on Chrome, Safari, and other WebGPU browsers without conversion to other formats.
The attack requires no retraining and passes standard safety evaluations, with triggers as simple as a date change.
The method learns a binary mask over gate projections, resulting in checkpoints hundreds to thousands of times smaller than full fine-tuning.
The system uses LLMs to validate user-stated interests and apply them responsively to ranking and delivery, balancing new content visibility with existing engagement.
The airline cut agent handling time by 10% and tripled Flight Recommender usage by integrating OpenAI models into customer support and internal workflows.
The model features a 1M-token context window for complex software engineering and cybersecurity tasks, with a 50% price reduction ending November 30, 2026.
The cascade approach uses specialized judges and confidence gating to reduce LLM calls while raising rationale alignment with human annotators to 82%.
An Anthropic analysis shows that the effectiveness of agent memory varies by model and instructions, highlighting the importance of verification and restructuring.
Priced at $0.10 per million input tokens, the model targets high-volume agent workloads like tool calls and coding subtasks.
Kakao's Kanana-2-v-decide-3b-2611 is a decision model that accepts text, JSON, or chat history with optional images to output probabilities for choice, yes/no, or scale-based answers.
It recorded a 0.07% out-of-scope attack rate while discovering 948 Linux kernel privilege escalation vulnerabilities.
The model features a 1M-token context window and is available in public preview without waitrooms or session limits.
Alexiuk argues that full access to weights, data, and training recipes is essential for AI democratization, using NVIDIA Nemotron as the primary example for developers.
The disparity in generation cost between later and earlier tokens on frontier open-weight models can reach 7x.
The repository includes a minimal implementation of AlphaEvolve, data for a new upper bound on the matrix multiplication exponent, and verification code for 15 mathematical problems.