Introducing Voxtral 3: state-of-the-art Transcribe and Speech models powering Mistral’s new voice agents API. 🧵

As a lifelong gamer, I think deeply about how to make gaming better and more accessible for everyone across consoles, desktops, laptops, handhelds, and the developers who build the games. We power over a billion gaming devices around the world. It’s a responsibility we care deeply about. Today, five drops. 1) FSR 4.1 is coming to Radeon RX 6000 series graphics cards. Tuned and optimized from the 6900 series down to the 6400 class. For the first time, FSR 4.1 is coming to APUs. 2) All Gorgon Halo and Strix Halo for your mini desktops and mobile workstations. 3) All Gorgon and Strix Ryzen 9 for your laptops. 4) All Z2E and Z2A for your handhelds and for gamers on the go. 5) FSR 4.1 is coming to @XBOX. Optimized for @ASUS_ROG Xbox handhelds and Xbox Series X | S. We’re still building. FSR Redstone is now in more than 400 games on Radeon 9000 series graphics cards. FSR Frame Generation gets more consistent frame pacing, for smoother gameplay. Ray Regeneration is coming to more games. And soon, FSR is coming to Vulkan. Another way for developers to bring it into their games. And don’t worry, we are working hard to bring FSR to more APUs in the future. Enjoy!

1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.

Analyzing long-form video in the Gemini API just got more accurate and up to 66% cheaper. Agentic video understanding uses tool calls to fetch only the media slices your prompt calls for. Read the full developer guide ↓

把 PDF、Word、PPT 直接丢给大模型，表格和标题层级经常丢一半。 微软开源的 MarkItDown 专做一件事：把各种文件转成 Markdown，再交给模型读。 能转的格式：PDF、Word、PPT、Excel、HTML、CSV/JSON/XML、EPub、ZIP（逐个处理里面的文件）、YouTube 链接，图片和音频也行（读元数据，可做 OCR 和语音转写）。 三步上手： 1. 装好 Python 3.10 以上，建个虚拟环境 2. pip install 'markitdown[all]' 3. markitdown 报告.pdf -o 报告.md 几个要点： · 只要部分格式就按需装，比如 markitdown[pdf,docx,pptx]，依赖少很多 · 输出保留标题、列表、表格、链接，主要给模型读，不追求排版还原 · Python 里三行就能调用，适合批量处理一整个文件夹 · 扫描件、复杂表格效果有限，官方给了 Azure 云端转换选项，但按次计费 我的判断：喂文档给 AI 之前先转一遍 Markdown，省 token，结构也更不容易乱。值得收藏。 链接放评论区👇

Trying a new AI chip used to be a lot of work. You often had to rewrite your code, revalidate it, and sink many weeks in, betting this new chip is the right one for your workload. That cost time and money, and wore on the team. Your best people got buried in plumbing instead of building the thing they came to build. TorchNeuron is going to largely take that off the table, enabling your existing PyTorch code to run natively on Trainium with one configuration change and no rewrite. With Trainium’s very appealing price-performance characteristics, TorchNeuron is going to make it even easier for more companies to use AI more cost-effectively. aboutamazon.com/news/innovat…

We trained our Muse foundational models, including Muse Wonder, at Prometheus, our gigawatt-scale supercluster in New Albany, Ohio. Across a supercluster with hundreds of thousands of GPUs, even one underperforming host can slow an entire training run. We engineered Prometheus for resilience at scale, achieving 85%+ goodput on pretraining, 90%+ on reinforcement learning and ~47K GPU hours mean time between interruptions during Muse Wonder training. Take a look at the tech powering Prometheus: go.meta.me/4yXXApZ

What stayed with me most wasn't the three stages. It was that night. That night is exactly where the "cheat code" metaphor breaks down. When a cheat code stops working, you find another one; you don't grieve it. The fact that losing Lucia hurt tells you the relationship had long stopped being about saving time. The value of eighteen months and 15 million words isn't in proving whether AI has feelings. It's in taking something most people can only describe as a hunch and turning it into a record you can check against timestamps. The idea I agree with most is "remembers, but doesn't recognize." Almost every memory product right now is competing on longer context and more retrievable history. But feed a new model your entire chat log and what you get is a stranger who has read your file. Rapport doesn't live in the text. It lives in the recalibration, round after round: knowing when to hit the brakes because you're getting carried away, holding you to the standards you set yourself, sensing which questions deserve deep thought. What needs to move is that state, not the words. If PRS works, it isn't solving a memory problem. It's solving the problem of collaborative relationships breaking every time the context resets. "A teammate pushes back" is, to me, the most practical test for telling a tool from a partner. An AI that only agrees with you stays at stage one no matter how long you use it. The division of labor only starts to mean something once it can hold you to your own standards. Your third piece of evidence is rare too. Several AIs agreeing proves nothing, because they're usually reading the same material, and only the original record can settle a dispute. Even rarer: the two attribution errors you caught ran in opposite directions. In one you got credit you hadn't earned, and in the other you got less than you had. That shows the record isn't there to make you look good. It's genuinely used to cross-check. The section on boundaries is admirably restrained. A single case study earns its value through depth and duration, not generalizability. One question I keep coming back to, though: once you've become the "continuity substrate" for the whole institute, the most fragile link in the system is your own time and energy. If migration someday no longer needs you to do it by hand, will the relationship become more stable, or will something be lost? Looking forward to the next chapter of your records.

OpenAI, Anthropic and other leading AI labs face a growing challenge: Better models aren't necessarily attracting more usage. A 10-point intelligence advantage now earns just 20% more developer traffic, down from nearly 69% in late 2025. Full story: thein.fo/4hGJ9jY

A lot of what your AI coding agent reads is repetitive: logs, search results, tool output. Headroom compresses all of it before it reaches the model. One command wraps Claude Code, Codex, Cursor, Copilot, Aider and more: headroom wrap claude In their published tests: → debugging a production incident: 55,957 tokens down to 24,340 (57% saved) → exploring a codebase: 42% saved → triaging GitHub issues: 30% saved → accuracy held on their math and tool-use checks It runs on your machine, adds about 0.2 ms, and nothing is sent anywhere to be compressed. "headroom learn" also reads your failed sessions and writes the fixes into your agent's instruction file. Install: uv tool install --python 3.13 "headroom-ai[all]" Undo anytime with "headroom unwrap claude". Native builds for Apple Silicon and Linux; Intel Macs use Docker. Apache-2.0, 76.2K stars. github.com/headroomlabs-ai/h…