XAI's open-source terminal AI coding agent grok now supports custom model endpoints, allowing users to run models like GLM 5.2, DeepSeek V4 Pro, and Kimi K2 via Ollama Cloud or any OpenAI-compatible p…
A new MLIR-based compilation method for large language models, developed by researchers and implemented in the TPU-MLIR compiler and LLM-TPU deployment project, addresses challenges in importing train…
Alibaba's Qwen team previewed Qwen3.8, a 2.4 trillion-parameter AI model, on July 19th, claiming it trails only Anthropic's Claude Fable 5 among frontier systems, but withheld technical details such a…
Alibaba is launching Qwen3.8, a 2.4-trillion-parameter open-weight AI model that it claims is one of the most powerful available, second only to Fable 5. The Qwen3.8-Max-Preview is already available o…
China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect July 15, 2026, forcing ByteDance, Alibaba and Tencent to remove personalized companion agents fro…
Researchers have developed a method to fingerprint large language models (LLMs) using just one output token per query, enabling verification of model identity in opaque serving chains. Testing 165 mod…
A developer deleted a line from their bio boasting about running hand-built inference engines for open-source models like Qwen and DeepSeek, realizing it was a 'demo' rather than a true flagship. The …
A developer running a small training platform found that a bug in the evaluation harness, not the models, was preventing five open-source models from passing a 54-task exam. The same pattern repeated …
A team of developers built Memoria, a personal AI with human-like memory that remembers, forgets, and evolves over time, for the Qwen Cloud Hackathon. The system uses a three-tier memory architecture …
Alibaba's chip design unit T-Head open-sourced SAIL, the full software stack for its Zhenwu AI chips, at the World AI Conference in Shanghai on Saturday, aiming to lower migration barriers for develop…
A developer outlines a three-layer testing approach—syntax, schema, and semantic validity—for ensuring reliable structured outputs across multiple AI models including GPT, Claude, Gemini, DeepSeek, Qw…
A developer has released an open-source tool called `compaction_omission_gate.py` that detects whether a context compactor has dropped a record that would have blocked a proposed agent action. The too…
Open Interpreter, a lightweight open-source coding agent for open models like Deepseek, Kimi, and Qwen, gained 2,024 GitHub stars. The project, a fork of OpenAI's Codex with 66k stars, was last update…
A new open-source sandbox allows users to run AI agents locally, eliminating the need for cloud-based sandbox providers and per-minute billing. The tool enables large language models to execute comple…
OpenCode sends user prompts to a public cloud endpoint (https://opencode.ai/zen/v1/responses) to generate a title summary even when a local LLM is explicitly selected, according to a bug report on the…
A developer shared a configuration example for running Qwen 3.6 with llama-turboquant, a Docker-based CUDA inference container. The setup includes GPU offloading, flash attention, speculative decoding…
Apple Intelligence has been registered for use on iPhones in China by the Cyberspace Administration of China on July 15, nearly two years after its announcement at WWDC 2024, but the system will rely …
A developer explains that attention sinks—tokens at position 0 that absorb excess attention weight—cause streaming LLMs to fail when evicted from the KV cache. The softmax normalization forces the mod…
China's Qwen model now accounts for a growing share of the AI models used by Western companies to serve, train, and build AI, according to a co-written article by Dean Meyer and Konstantine Buhler. Th…
Moonshot released Kimi K3, a 2.8 trillion-parameter open-weight model, the largest ever, with native vision and a 1M-token context window, claiming ~2.5× better scaling efficiency than K2 and pricing …