mlx-audio v0.4.0 is here 🚀 What's new: → Qwen3-TTS: fastest generation on Apple silicon and first batch support. > Sequential (<80 ms TTFB at 2.75x realtime) > Batch support (<210 ms TTFB at 4.12x for batch of 4-8) → Audio separation UI & server → nvfp4, mxfp4, mxfp8 quantization → Streaming /v1/audio/speech endpoint → Realtime STT streaming toggle New models: → Echo TTS → Voxtral Mini 4B, → MingOmni TTS (MoE + Dense) → KittenTTS → Parakeet v3 → MedASR → Spoken language identification (MMS-LID) → Sortformer diarization + Smart Turn v3 semantic (VAD) Plus fixes for Kokoro Chinese TTS, Pocket TTS, Whisper, Qwen3-ASR, and more. Thank you very much to @lllucas, @beshkenadze, @KarnikShreyas, @andimarafioti, @mnoukhov and welcome the 13 new contributors 🙌🏽 Get started today: > pip install -U mlx-audio Leave us a star ⭐ https://t.co/bQ5WBLR6FK

Farmers are turning to drones and AI to fight weeds more precisely. By identifying unwanted plants in real time, the systems can target herbicides exactly where needed. The result could mean lower chemical use, lower costs, and smarter agriculture. https://t.co/WITdTG6T8I @bbcnews

wow it's so cool that they added our favorite feature from the Claude Code CLI to the desktop app https://t.co/evcMJBWDf0

Traffic police… but in the sky. In Shenzhen, drones are now responding to traffic accidents in real time. Officers can analyze the scene remotely, generate a 3D reconstruction, and complete responsibility reports in about 5 minutes. https://t.co/hYefGavepK

Traffic police… but in the sky. In Shenzhen, drones are now responding to traffic accidents in real time. Officers can analyze the scene remotely, generate a 3D reconstruction, and complete responsibility reports in about 5 minutes. https://t.co/hYefGavepK

i made a tamagotchi that lives in your notch and reacts to your claude code sessions. it cries when you yell at claude and gets happy when you praise it. https://t.co/YYFEz8NyGy

I spent over a month and 1.8 B Tokens running OpenClaw 24/7 and discovered what nobody tells you. This is a must-read for both OpenClaw users and people trying to start their own OpenClaw.

I will never hide the thinking trace from Grok and its agents. Watching them thinking and discuss is pure entertainment for me. https://t.co/0a5QKDZQ6x

for now it has mostly been improvements on “agentic coding” prepare yourselves for “agentic research” expect something from me in the coming weeks on the same

AgentLens: A WebUI for local observability of agent traces 🔎 📝 Happy to share a side-project for detailed real-time inspection of what all your coding agents are up to. npx -y @roberttlange/agentlens --browser It locally monitors session traces and allows you to interact & terminate sessions in Codex, CC, OpenCode, Pi, Cursor Agents & Gemini CLI 🧑‍💻

mlx-vlm v0.4.0 is here 🚀 New models: • Moondream3 by @vikhyatk • Phi-4-reasoning-vision by @MSFTResearch • Phi4-multimodal-instruct by @MSFTResearch • Minicpm-o-2.5 (except tts) by @OpenBMB What's new: → Full weight finetuning + ORPO h/t @ActuallyIsaak → Tool calling in server → Thinking budget support → KV cache quantization for server → Fused SDPA attention optimization → Streaming & OpenAI-compatible endpoint improvements Fixes: • Gemma3n • Qwen3-VL • Qwen3.5-MoE • Qwen3-Omni h/t @ronaldseoh • Batch inference, and more. Big shoutout to 7 new contributors this release! 🙌 Get started today: > uv pip install -U mlx-vlm Leave us a star ⭐️ https://t.co/un61O8fEZd

mlx-vlm v0.4.0 is here 🚀 New models: • Moondream3 by @vikhyatk • Phi-4-reasoning-vision by @MSFTResearch • Phi4-multimodal-instruct by @MSFTResearch • Minicpm-o-2.5 (except tts) by @OpenBMB What's new: → Full weight finetuning + ORPO h/t @ActuallyIsaak → Tool calling in server → Thinking budget support → KV cache quantization for server → Fused SDPA attention optimization → Streaming & OpenAI-compatible endpoint improvements Fixes: • Gemma3n • Qwen3-VL • Qwen3.5-MoE • Qwen3-Omni h/t @ronaldseoh • Batch inference, and more. Big shoutout to 7 new contributors this release! 🙌 Get started today: > uv pip install -U mlx-vlm Leave us a star ⭐️ https://t.co/un61O8fEZd

The new Mistral 3 models in llama.cpp https://t.co/G4HoEGw85G

@arthur_spirling Arguably, what one would ideally need to compute is not the exact gradient but the direction of steepest descent assuming a step size commensurate with the expected change in parameters. No need for this fancy infinitesimally-small fluxion stuff!

ok but like, if you want to run some interesting benchmarks on our 5.4 series and pro, and find the cost to be an issue. dm me!

@joburgai @_xpn_ This is for illustration purposes, so I am only focused on math tasks. E.g., consider the MATH dataset with 12,500 math problems. If the 12,000 samples that are not in MATH-500 (which is the test set), distill answers from the largest Qwen3 model, you can improve the accuracy of the 0.6B model from 15.3% to 45.8% on MATH-500, which is an amazing jump.

Thanks AK for sharing our work! To the best of our knowledge, this is the first large-scale, open-source resource for instruction-reference guided video editing. Code, model, and data are all available now: https://t.co/g3Ji2JhTLn

GPT-5.4 with the Roblox MCP server created its own 3D models, animations and GUI. It's so cool you can make games so easily https://t.co/TIxf4fihhu

Here is the key table showing the 30 day effects (which they misreport in the paper) &amp; which apparently uses a completely different test, but the ChatGPT group remains ahead. The errors shouldn’t fill you with confidence about the study, though. https://t.co/hqocnAsF38

The next step for autoresearch is that it has to be asynchronously massively collaborative for agents (think: SETI@home style). The goal is not to emulate a single PhD student, it's to emulate a research community of them. Current code synchronously grows a single thread of commits in a particular research direction. But the original repo is more of a seed, from which could sprout commits contributed by agents on all kinds of different research directions or for different compute platforms. Git(Hub) is *almost* but not really suited for this. It has a softly built in assumption of one "master" branch, which temporarily forks off into PRs just to merge back a bit later. I tried to prototype something super lightweight that could have a flavor of this, e.g. just a Discussion, written by my agent as a summary of its overnight run: https://t.co/tmZeqyDY1W Alternatively, a PR has the benefit of exact commits: https://t.co/CZIbuJIqlk but you'd never want to actually merge it... You'd just want to "adopt" and accumulate branches of commits. But even in this lightweight way, you could ask your agent to first read the Discussions/PRs using GitHub CLI for inspiration, and after its research is done, contribute a little "paper" of findings back. I'm not actually exactly sure what this should look like, but it's a big idea that is more general than just the autoresearch repo specifically. Agents can in principle easily juggle and collaborate on thousands of commits across arbitrary branch structures. Existing abstractions will accumulate stress as intelligence, attention and tenacity cease to be bottlenecks.

I built a cloth simulation test Opus 4.6 used 100% + $3 credits GPT 5.4 (high) 10% If you are on a tight budged and want the best value for $20, it's note even close

I built a cloth simulation test Opus 4.6 used 100% + $3 credits GPT 5.4 (high) 10% If you are on a tight budged and want the best value for $20, it's note even close

This (very small) study hints at something more interesting. If you use AI to support learning while coding you can gain additional skills, if you delegate all intellectual work to AI you learn nothing. This has also turned out to be true in other larger RCT studies in education https://t.co/sp4cqPNwBP

The obvious reasons intelligence-per-watt is going up so fast: more efficient architectures, more efficient hardware, and higher quality data. The less obvious reason: finding the right balance on what should be stored in the model's weights and what can be computed through tool use, reasoning, and potentially other types of in-context learning. A simple example: in the earlier LLM days, it was quite likely that for simple arithmetic (e.g. adding two numbers), the model had to basically memorize tuples of (inputs, op, outputs). You can imagine this took up a lot of room in the weights. With reasoning the model can compute this in its chain-of-thought. With tool calling the model can compute this with a tool call. In both cases it saves a lot of space in the weights. I'm sure there is a floor on the smallest LLM that can have say GPT 5.x quality. But that floor could be 5B, it could be 100B. And I don't think anyone really knows because of the above effects. In other words we can probably go much further with a 5B-15B model with exceptional tool calling and reasoning.