Researchers introduced SCX Router, a lightweight GLiClass-based model for zero-shot selection among LLM inference endpoints, achieving an aggregate top-1 score of 0.707 on a 1,000-task subset versus 0…
PulsarForge, a CPU-only LLM inference engine written in C11, runs a 744-billion-parameter GLM-5.2 MoE model on a 2018 laptop with 32GB RAM and a USB SSD, achieving 9.3 seconds per token interactive sp…
Perplexity released Hybrid Compute for Mac, enabling its Perplexity Computer agent to split tasks between cloud frontier models and a local on-device model, with an on-device privacy gate deciding wha…
Researchers introduced OCGQuant, a post-training quantization method that pairs outlier channels with low-magnitude companions to reduce quantization error in NVFP4 low-bit inference, achieving the lo…
Researchers introduced REAL-Q (Real-time E2E-loss Aligned LLM Quantization), a post-training quantization method that uses dynamic Block-wise Gradient Descent to reduce end-to-end KL divergence by up …
A developer's attempt to run an AI coding agent entirely offline on a laptop using the open-source pi.dev agent CLI, Ollama, and Alibaba's Qwen3 27B model succeeded in building a 1,086-line Tetris gam…
Researchers propose self-specialized teacher distillation (SSTD), a two-stage procedure that trains a domain teacher from a base model and distills its token distribution to a student on self-sampled …
A developer's experiments with an agentic terminal and a fine-tuned model failed because they were measured with self-built instruments that lied, leading to the development of a context engine called…
A new arXiv paper (submitted Aug 26, 2026) reveals that self-evolving LLM coding agents can be poisoned through a self-propagating worm attack called EvoMal, which exploits the agents' habit of imitat…
Four reproducible parser failures in vLLM 0.26.0, 0.27.1, and 0.28.0 can drop or garble tool calls and reasoning content while still returning HTTP 200, according to the Ingot team's tests on CPU with…
Researchers fine-tuned a generalist LLM based on Qwen3 for hate speech mitigation, unifying 36 English hate speech datasets, and achieved state-of-the-art performance on in-domain benchmarks with subs…
Appier AI Research and National Taiwan University introduced SMITH (Schema-grounded Multi-task Iterative Tool Honing), a reinforcement learning framework that jointly trains tool creation and tool use…
Researchers propose SMITH (Schema-grounded Multi-task Iterative Tool Honing), a reinforcement learning framework that jointly trains tool creation and tool use in a single policy, enabling a 4B Qwen3 …
A new arXiv study (2608.21345v1) presents the first stage-wise model size analysis of self-refinement pipelines, testing 6 model sizes of Qwen3 and 4 model sizes of Gemma 3 across 5 benchmarks. The re…
LLM tokenizer variance means identical dollar-per-million token prices can produce materially different bills, because each vendor's tokenizer segments the same input into different token counts. Anth…
Researchers introduced Skill Entropy RL, a reward that quantifies the difficulty of switching between reasoning skills, and showed it more than doubles the accuracy of Qwen3 models on the Skill²-Bench…
A developer building tetanus, a Rust-based pentesting automation tool with C2 and collaboration features, is seeking help integrating AI automation into the tool and asks whether guardrails can be dis…
A developer's four-month home-lab test found that self-hosting open models costs about 7× more than using hosted APIs, cutting their bill from roughly $3,050 to $420 per month by switching to API acce…
A developer building a retrieval-augmented generation (RAG) pipeline discovered an accidental prompt injection triggered by retrieved text from a book about large language models. The developer fixed …
A new arXiv study (2608.18271v1) finds that performance gains from on-policy self-distillation (OPSD) in Qwen3 models (1.7B to 8B) do not consistently show that the student learned privileged referenc…