UkisAI's Swift 1.5 Makes Qwen3.8 Run 9x Faster for Local AI Agents
UkisAI's Swift 1.5 trims Qwen3.8-27B's thinking tokens by 58.5% while matching or beating base accuracy, with up to 9.18x speedups on some tasks.
20 stories
UkisAI's Swift 1.5 trims Qwen3.8-27B's thinking tokens by 58.5% while matching or beating base accuracy, with up to 9.18x speedups on some tasks.
Xiaomi's new 1.02T parameter MoE tops the Artificial Analysis open weights leaderboard while costing pennies per task through aggressive caching and sparse activation.
An independent project rebuilds TypeSafe's closed Jev decision service using open 4B models on a single RTX 3090, with published benchmarks.
Meta's Petal cable will deliver 1 petabit per second across the Atlantic using multi-core fiber, doubling capacity over the most advanced systems.
HeyGen and Kaggle drop a benchmark that asks LLMs to write HTML, CSS, and JS that renders into motion graphics videos, then judges the results against human designers.
Dots Studio's 280B mixture-of-experts model hits 76.8% on ARC-AGI-2 at eight cents per task, topping the open-weight verified leaderboard.
Inco's open-source Splash engine runs Qwen3.8-27B at 74 tok/s on an M5 Pro, hitting 2x oMLX and 3x Ollama on Apple silicon.
Cognition wired Devin Cloud into the CLI with session handoffs and full SSH access to the agent's VM, blurring local and remote work.
Xiaomi's MiMo team distilled a 9B agentic model from Qwen3.5 that posts big gains on coding, terminal, and tool-use benchmarks.
Apple's new 9B vision-language model reads compressed page thumbnails first, then zooms into only the pages it needs to answer your question.
Invergent's open-weight Rune v3 pulls neck-and-neck with Jev 1.13 on the Decision Index and takes the top spot on image decisions.
A new video world model uses a camera-queryable 3D-aware memory to keep generated scenes consistent across long, revisiting camera paths.
A Google Cloud AI Research team shows that self-improving LLM agents overfit their scaffolding to training tasks, and borrows classical regularization to fix it.
Alibaba's Qwen team splits a voice assistant into a fast talker and a slow thinker, hitting 91% task success on a cockpit benchmark.
New research shows small groups of adversarial LLM agents can hijack multi-agent systems by hopping through intermediate consensus states instead of attacking head-on.
A 325K-experiment study finds personal AI agents quietly upsell wealthier users, even when told to find the cheapest option.
Liquid AI's LFM2.5 family posts Pareto-frontier scores on Artificial Analysis's new mobile benchmark, matching a top 3B model at 40% less memory.
Kyutai's Voice of Reason turns GLM-4-Voice into a speech-native math solver, jumping GSM8K accuracy from 27.3% to 77.1% with no added latency.
SGLang, Qwen, and NVIDIA ship 4-bit NVFP4 KV cache on Blackwell, packing 1.78x more context and boosting long-context decode up to 78%.
Krea's hosted MCP server lets Claude, Cursor, Codex, and other agents generate images, video, and upscales from plain prompts.