dextune

I happily picked up two used Nvidia Tesla P40 GPUs. I knew the P40 was old and ridiculously overbuilt. What I did not expect was how long the GPUs would become once the coolers were attached. Tried putting them in the case. Nope. The cooler end would not fit. I thought about it for about 10 seconds. Turns out the bottleneck was not the human. It was the case. So I grabbed a hammer and pliers and destroyed the front section of the case. After that both P40 GPUs fit perfectly. Put the cover back on and you cannot even tell from the outside. Apparently hardware fits if you hit it hard enough. Good experience.

SUPIR Image Restoration and Upscaling SUPIR is a photo restoration framework from CVPR 2024 that uses generative diffusion models to recover details from degraded images. Built on SDXL and LLaVA, it supports image restoration and upscaling with a default output resolution of at least 1024 pixels. Two checkpoints are provided. SUPIR-v0Q targets general image quality, while SUPIR-v0F preserves more original details in lightly degraded images. It supports local inference through Python and Gradio, with a lower memory configuration using 12GB for diffusion and 16GB for LLaVA. github.com/Fanghua-Yu/SUPIR

TensorFold Desktop macOS only TensorFold Desktop 1.0.0 is the first stable Mac app release for Apple silicon on macOS 13.3 or later. It is signed and notarized by Apple. Ask TensorFold adds window selection, capture preview and local text extraction in a compact input bar. The app stays in the menu bar when its window closes. It bundles engine 1.2.0, with 1.2.1 available as an optional update. Rollback preserves the app and model files. Windows and Linux builds remain previews. No Linux installer ships with this release. github.com/ashhart/TensorFol…

QwenPU QwenPU compiles a 32 bit RISC-V CPU and vector unit into Qwen3.5-4B, so one LLM can execute another during its forward passes. It runs Karpathy's 260K parameter stories260K as a C program. The added weights are compiled, leaving Qwen's original weights unchanged. The vector unit cuts trace tokens per story token from 225,000 on Qwenputer's NPU to 53,000, producing the same story. On an L4 GPU executor, the author reports about 3,300 trace tps and 16 seconds per story token, down from 41. github.com/divagr18/QwenPU

RustXNG RustXNG 0.26.0 is a self hosted Rust metasearch engine for AI agents, combining multiple search sources through a local JSON API. On the SealQA hard set of 254 questions, the developer reports 61.42% reference page retrieval in the top 10 results, with a median response time of 1.53 seconds. It can fetch page text for grounding and reduce model input through result limits and domain filters. Prebuilt binaries are available for Linux, macOS and Windows. github.com/sahenjp/RustXNG

Fibo 2.0 Edit Turbo 2-Step Edit Turbo Model Fibo 2.0 Edit Turbo has a new IDM LoRA that reduces image editing from 4 steps to 2 without CFG. akhaliq trained the adapter on 9,000 UltraEdit editing triplets at 1024x1024 for 550 steps on 4 x H200. The frozen teacher and generator branches share one base model. Across 128 MagicBrush eval samples, the repository reports L1 improving from 0.0479 to 0.0415 and CLIP image similarity from 0.9180 to 0.9290 compared with the original teacher. huggingface.co/akhaliq/Fibo-…

AI Phone ai-phone supports Apple Silicon Macs, including MacBook, Mac mini, iMac, and Mac Studio, but not Intel Macs, Windows, or Linux. Setup offers 2B for 8GB RAM, Qwen3.5 4B for 16GB, and 9B for 32GB. It handles incoming and outgoing SIP calls. Parakeet v2 recognizes speech and Phonon 31M generates it, both locally through MLX. On an M5 Pro with 24GB, the developer measured 1.50s median first audible response across 20 calls and 51.9ms median first TTS chunk. All five interruption tests resumed speech, with four stopping within 300ms. github.com/oxnw/ai-phone

uzu uzu is a Rust based inference engine for running AI models locally inside applications using Metal and Apple unified memory. Its high level API handles model downloading, session creation and inference configuration. Unified model configurations simplify adding support for new models. Rust, Swift, Python and TypeScript interfaces support streaming chat, structured outputs and tool calling. The engine also supports cloud models through the same API, with traceable computations to verify local inference against reference implementations. github.com/trymirai/uzu

PhotoCraft Lightweight photo and PSD editor PhotoCraft is an open source image editor built in Rust. It handles layers, masks, brushes, vectors, text and PSD files offline, following Photoshop workflows. It provides 68 tools and 16 adjustment layers, plus editable filters and layer styles. Its GPU compositor uses wgpu. In the project's PSD tests, 307 of 309 files keep the same rendered result after opening and saving. A registry of 500+ commands drives the UI, CLI and MCP server for scripted and agent editing. The project remains in early alpha. github.com/storytold/photocr…

Drama 3 Fish Audio's Drama 3 is a 35B MoE speech model with 3B active parameters, now serving through SGLang-Omni. The model supports up to 16 speakers, voice cloning and editing earlier takes. At 256 concurrent streams on H200 GPUs, moving the vocoder to a second GPU raises throughput from 2,422 to 4,160 acoustic frames per second. Median time to first audio drops from 3,403 to 2,146 ms. On one H200, it reaches 1,687.3 frames per second at 64 streams. A single stream starts audio in 88.9 ms median. huggingface.co/fishaudio/dra…

Project Maya Project Maya v1.1.49 fixes startup failures after DeepSeek Maya-XS v1 to v2 updates and recovers about 161GB of disk space. The failure came from checking an old copy of the N-gram tables instead of the shared tables. Running setup once replaces that copy with a link to the shared version. Qwen CYBER-FROST also runs on 8GB GPUs again after memory errors in v1.1.48. On an RTX 3050 8GB, the first request and speed calibration now complete, with generation speed about the same or faster. github.com/mw00/project-maya

PiVo-1.4M Empero Research released PiVo, an English TTS model with two voices and just 1,401,562 parameters, including the vocoder. It runs entirely on CPU using NumPy and ONNX Runtime, generating 16kHz audio without a network connection. The acoustic model has 584,965 parameters, while a FARGAN derived vocoder handles speech synthesis with another 816,597. Empero reports a 0.532% word error rate on 152 held out test utterances transcribed by Whisper small.en. huggingface.co/empero-ai/PiV…

Hy-SimulMT-30B-A3B Tencent Hunyuan released Hy-SimulMT-30B-A3B for live Chinese and English speech translation, reporting about 3 seconds of StreamLAAL on RealSI long audio. Its RealSI-Sentence score reached 84.7, compared with 84.4 for Qwen3.8-LiveTranslate. The release includes FP8 and GGUF checkpoints, plus SimulInfer for inference and SimulTrain for continued training. Speech output and real time voice cloning require a separate TTS service. huggingface.co/tencent/Hy-Si…

DeepTutor DeepTutor is an open source AI tutor that lets you study with your own documents, ask questions, and test what you have learned with quizzes. It brings lessons, notes, reading materials, and learning progress into one place, while remembering context across study sessions. Version 1.7.1 adds preview desktop apps for Mac and Windows that bundle the software needed to run the workspace. The update also adds narrated video lessons, flashcard exports to Anki, and calendar exports. github.com/HKUDS/DeepTutor

Elephant-Goldfish Model Thousands of Google engineers use Elephant-Goldfish Model to design and review software before writing code. The open source plugin separates design drafting from independent review using AI agents. Elephant drafts specs with full project context. Fresh Goldfish subagents review the draft and references without prior chat history, exposing missing assumptions and implementation risks. It provides skills for brainstorming, design reviews and detecting drift between docs and code. It supports Antigravity, Claude Code and Codex. github.com/google/elephant-g…

MLX-Serve MLX-Serve v26.11.1 speeds up DeepSeek-V4.1-Flash inference on Macs, especially long prompt processing on M5 hardware. According to the release notes, decoding is also faster under concurrent requests. Conversation caches now persist even after a reply is interrupted. The update adds custom model plugins that can be built, tested and installed with signer verification. Plugins can draft ahead and process images. NVIDIA Nemotron 3.5 Transcribe is also supported, with speech recognition in 40 languages, translation, speaker labels and word timestamps. github.com/ddalcu/mlx-serve

Xiaomi-OCR-0 Xiaomi released Xiaomi-OCR-0, a 0.8B vision language model for document parsing, built from Qwen3.5-0.8B and trained on roughly 170M OCR samples. It scores 95.24 on Real5-OmniDocBench using a two stage pipeline, compared with 92.29 for OvisOCR2. Across five OCR VQA benchmarks, the mean rises from 72.9 for Qwen3.5-0.8B to 83.2, exceeding Qwen3.5-2B at 80.9. The model uses Q-Mask text anchoring and mixed task reinforcement learning. Local inference supports vLLM and SGLang. github.com/xiaomi-research/X…