A local MCP server plugin for Astah Professional lets AI agents design and explain UML models, generate code from models and back, and build diagrams from hand-drawn sketches via programmatic tool calling.
Lumae is a free native macOS screen recorder and screenshot tool whose recordings AI agents can edit over MCP, with automatic zoom, smooth cursor, motion blur, captions, and annotation callouts.
DeepSeek is reportedly preparing a Shanghai STAR Market listing with CITIC Securities as sponsor, targeting an A-share debut in 2026-2027, following a June round of about $7.2B at roughly $56B post-money.
Replacing dense random orthogonal matrices with sign-randomized DCT transforms makes Kashin-decomposition-based weight quantization far cheaper while staying competitive with OPTQ, QuIP and QuIP-RG at 4-bit per channel, and remaining numerically stable where QuIP diverges on Pythia-6.9B and Mistral-7B.
A calibration-free scoring method based on latent spectral geometry and KS distances outperforms PruneNet's 45M-parameter RL policy by 1.57pp on LLaMA-2-7B at 25% sparsity and beats WikiText-2-calibrated SliceGPT by up to 6pp.
NSD makes LLMs diverge from their own flawed reasoning traces instead of imitating ground-truth solutions, countering on-policy self-distillation's tendency to suppress uncertainty and self-correction on complex reasoning tasks.
Progressive audio-encoder compression with cross-scale distillation cut macro-average error on Qwen3-ASR-0.6B from 5.61% to 5.27% across ten Chinese-English benchmarks while the language-model backbone stayed frozen.
Fused representation compression reduces the compute, memory and communication demands of running multimodal LLMs on resource-constrained edge platforms across vision, text and telemetry modalities.
A Qwen3.6 fine-tune from the dougalldeepmind experimental line was uploaded Sept 12 with safetensors weights, part of the ongoing delib/sonnet series.
A second in-window upload from dougalldeepmind focuses on agentic advice and reasoning traces; minor but genuine new weights for the reporting window.
Daily build b10937 extends the noshuffle row-alignment rule in the OpenCL backend to q4_K, q5_K and q8_0 quantizations, previously limited to q6_K.
A configurable add-on for local AI coding agents ships 21 role agents across four seniority tiers, 26 portable workflow skills and an escalation protocol, working with Claude Code, Codex, Gemini CLI, Cursor, Windsurf, OpenCode and GitHub Copilot.
An AI-first browser testing and automation platform managed entirely through MCP, with low token consumption, a scheduler, CI webhooks and a Chrome extension that records events for agents.
A free, open-source pipeline that runs entirely on your machine, handling development from issue ticket to reviewed pull request using AI coding agents.
This week's hardware roundup finds VRAM capacity alone cannot overcome severe software and inference-engine bottlenecks when benchmarking Qwen 3.8 27B, alongside coverage of the splintered compute economy and GPU/memory pricing pressure.
Nvidia offered RTX 5090 ($1,999), RTX 5080 ($999) and RTX 5070 ($549) Founder's Edition cards at original launch MSRP to Verified Priority Access attendees, amid the 2026 GPU price premium environment.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
