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.
Kanzhun released the 33B dense model Nanbeige4.3-33B under the Apache 2.0 license.
Post-trained on DeepSeek V4.1 Flash, the model targets software engineering tasks with shorter reasoning and fewer tokens.
The 1.7B model scores 59.8 on SWE Verified and 85.3 on tau2-bench, closely trailing the 4B variant's 62.8 and 86.9.
The upgraded 560B-parameter model features a 1M-token context window and improved performance on reasoning and automation benchmarks.
Toss significantly reduced LLM invocation costs by applying Selective Neural Approximation (SNAP).
The system separates LLM interpretation from code-based calculation to ensure accurate daily reports for LINE OA and Display Ads.
The framework uses MoE expansion and RL-based distillation to consolidate short-term context into long-term parameters, achieving 79.2% accuracy on AIME 24/25.
The method uses self-generated question-answer pairs and reasoning traces to boost performance across text and multimodal domains while retaining capabilities in untrained areas.
The system wraps deterministic processes in AI to extract errors and provide focused diagnosis, replacing legacy regex rules that left 50% of failures unclassified.
The 1.58-bit MoE model achieves 172 tok/s decode and 7000+ tok/s prefill speeds in both Stock llama.cpp and the PrismML fork.
The model is designed for high-volume agentic tasks, achieving a 76.5% pass rate on SWE-Bench Pro v2-Hard in its initial 0.1 release.
The method, presented at ICML 2026, generates observation and action models via LLMs to reconfigure the agent-environment interface, outperforming human-designed baselines in locomotion and manipulation tasks.
The 33B model serves as a teacher for upcoming smaller agents and outperforms Qwen3.8-27B on Terminal-Bench 2.1 but trails on SWE-Bench Pro.
A training-free method using five coordinated model instances raises open-weight Qwen3.8-27B accuracy from 66.8% to 92.4% on hard coding tasks.
Nine frontier and near-frontier models use difficult psycholinguistic features more often than human writing, attributed to optimization for non-communication goals.
OpenAI donated the Open Responses specification to AAIF to promote the standardization of LLM calls.
A new framework identifies scientifically plausible research directions that existing communities are unlikely to propose, recovering 2025 combinations at 35.9 times chance.
The system uses BookLLM to profile characters with 97% accuracy in English and 94% in Spanish, enabling distinct and consistent voices across full-length books.
You can control local and cloud LLMs identically by changing only the import statement.
The model is now accessible for use with agents within the OpenClaw platform.
The new models improve coding, 3D modeling, and financial analysis while maintaining a 1M-token context window and image input capabilities.
The 14.8B model averages 66.7% on five DNA benchmarks, outperforming Qwen3-14B by 27 points and running on a single 80 GB GPU.
Open-source engine Kapitsa matches Recursive's native performance on short contexts and leads by 1.30x at 128K context on Apple Silicon.
PyMuPDF 2.0 integrates three packages to convert PDFs into LLM-ready data without a GPU.