SolidCraft is a new open-source CAD app written in pure Rust (MIT or Apache-2.0), with a built-in MCP server so a coding agent such as Claude Code can model parts in it. Its README reports that 75 of 76 public reference parts match their expected volume, area and centre of mass, and calls the app early but usable for simple parts. Would you use a part an AI agent modelled without opening it yourself first? github.com/bherbruck/solidcr…

Anthropic stopped using git for its monorepo. The number behind it: about 2,500 merged changes a day in May, and 30,401 on 9 September. From their engineering post on Scion, the replacement: • CI runs on the order of 25 million checkouts a day. Fetched naively from git, that would be over 600 PB of traffic a day • at the peak they ran several hundred internal git mirrors just to keep up • Scion's footprint is 200x smaller in CPU and memory than those mirrors, each cache tier holds a hit rate over 99%, and checkouts take seconds instead of minutes • seven people built it, a new code review tool and the migration in five months Their own limit: Scion is "by no means a suitable replacement for git for general use". The design only makes sense for enormous monorepos, with tens of thousands of engineers or swarms of agents writing code. On prior art, the post credits Google's Piper and Meta's Sapling. Rain, who helped build Meta's open-source Mononoke server, argues that Scion's "derived data" design follows Mononoke's far too closely for the post never to name it. Rain is fine with the copying; the complaint is the missing credit. anthropic.com/engineering/re… sunshowers.io/posts/anthropi…

Your coding agent probably doesn't need a bigger context window. It needs a cleaner one. In September 2025, Anthropic called it likely that, for the foreseeable future, context windows of all sizes will be subject to context pollution, at least when you want the strongest agent performance. Most of its Claude Code best practices rest on one constraint: the window fills up fast, and performance degrades as it fills. One long session: Task A → "investigate X" → back to A → correction ×2 → same-session review Where it breaks: • "investigate X", unscoped: Claude reads hundreds of files, filling the context • correction ×2: the context is cluttered with failed approaches • same-session review: Claude reviews code it just wrote • the whole session: as the window fills, Claude may start "forgetting" earlier instructions What fixes it in Claude Code: 1. Send the investigation to a subagent. Subagents run in separate context windows and report back summaries. 2. After two failed corrections, /clear and write a better prompt with what you learned. A clean session with a better prompt almost always outperforms a long session with accumulated corrections. 3. Review in a fresh context, where Claude won't be biased toward code it just wrote. Same session, fixed: Task A → subagent investigates X → back to A → correction ×2 → /clear + better prompt → fresh-context review Not just Claude: Chroma's Context Rot report (July 2025) tested 18 models, closed and open-weights, and found performance degrades as input length increases. Open-source Aider (Apache 2.0) has the same moves: /drop files "that don't need any more changes", and when it gets stuck, /clear "to discard the chat history and make a fresh start". Long sessions are fine when you're deep in one complex problem and the history is valuable. Which do you use least: a subagent, /clear or a fresh-context review? code.claude.com/docs/en/best… anthropic.com/engineering/ef… aider.chat/docs/usage/tips.h… #ClaudeCode

If a model running on your own machine feels slow, check where it actually loaded before you blame the model. ollama ps The PROCESSOR column tells you: • 100% GPU: the whole model sits in the graphics card's memory • 100% CPU: all of it is in system memory • a split like 48%/52% CPU/GPU: part of it is on the GPU, the rest in system memory Models also stay loaded for a while after you use them. ollama stop unloads one now, and keep_alive (or OLLAMA_KEEP_ALIVE for the whole server) sets how long they stay: 0 unloads right after a response, a negative number keeps the model loaded. docs.ollama.com/faq

Mistral has released an open-weights speech-to-text model built for Arabic: Voxtral Mini 4B Realtime Arabic, under Apache 2.0. It transcribes as the audio arrives instead of waiting for the whole recording, the shape you need for live captions or a voice assistant. Mistral says it covers Arabic dialects and Modern Standard Arabic, and training focused on code-switching, where speakers move between languages in the same conversation. Egyptian, Moroccan, Gulf and Levantine Arabic are all in its language list. The numbers are Mistral's own: at a 480 ms delay it averages 8.82% character error rate across seven Arabic benchmarks. That still trails Mistral's Voxtral Transcribe Arabic at 7.91%, by 0.91 points. The card also says scores may differ under other normalization methods, so compare vendors with care. It's about 4.4B parameters with BF16 weights, so roughly 9 GB of weights, and the card's way to serve it is vLLM: vllm serve mistralai/Voxtral-Mini-4B-Realtime-Arabic Which Arabic dialect would you test it on first? huggingface.co/mistralai/Vox…

If your team builds on Claude, there's a date to put in the calendar: Anthropic's updated Usage Policy takes effect on November 12. Anthropic says most of the changes clarify rules that already existed. Three parts are still worth a manager's ten minutes. The first covers high-risk uses. When Claude's output can affect someone's health, legal rights, finances, livelihood or access to essential services, there has to be a qualified human in the loop, someone with the authority to review and change Claude's recommendations, and the person affected has to be told AI was used. That requirement isn't new, but the section now lists which kinds of recommendations are covered and which aren't. A good moment to check where your product sits. The second covers physical actions. If Claude drives hardware that acts on its own and could injure someone, a qualified operator must be able to watch it and stop it, and the equipment must hold a safe state if Claude disconnects. The third is surveillance. Tracking people without their consent is prohibited, and Claude can't be used to decide or recommend who to investigate, arrest or charge. Tracking people agreed to, like fraud monitoring, stays permitted. The clause getting the headlines, a ban on sustained, needless cruelty toward the models, is meant only for extreme cases, not ordinary frustration or model testing. For most teams the work is a short review of where their use cases sit before the 12th. anthropic.com/news/2026-usag…

When NVIDIA announced RTX Spark on 7 October, it named Qwen 3.8 Flash Next, a 125-billion-parameter model, as one it can run. An open-source project already runs that model on an ordinary gaming PC. Strata (MIT, more than 18,000 GitHub stars) lists what it takes: • a graphics card with 12 GB of video memory (NVIDIA RTX 20 series or newer, or one of the listed AMD Radeon cards) • 32 GB of RAM for the coding version, 64 GB for every size • about 80 GB of free disk, for a download of about 70 GB Why it fits: Qwen's model card says only 6 billion of the 125 billion parameters are active for each token (about ¾ of a word), and Strata downloads compressed versions made by ISTA-DASLab and others. On a 12 GB RTX 5070 with 64 GB of RAM, the project measured 53 to 94 tokens per second in a short chat. Its own numbers, on two PCs. The catch for companies: the weights come under the Qwen Community License 1.0, not an open-source licence, and an AI-as-a-service or AI work-assistant business needs a separate licence from Qwen for commercial use. github.com/Niko1221/Strata

If you use the Gemini app with a personal Google account, the model menu changes on October 9. Google's help page says the change starts that day for people without an AI subscription (AI Plus members get an email with their own date). The table lists: • No plan: Flash-Lite • AI Plus: Flash-Lite and Flash • AI Pro: Flash-Lite, Flash and Pro • AI Ultra: Flash-Lite, Flash and Pro Pro and Ultra also keep Deep Think, the option for maximum parallel reasoning. Two details matter more than the names. On every available model you can pick low, medium or high effort, and higher effort uses more of your limit. And since the May change, limits are compute-based: they refresh every 5 hours up to a weekly cap, and long chats, media generation and Deep Research use more. Plus gets 2x the standard limit, Pro 4x, and Ultra 5x or 20x Pro's, depending on the subscription. For a team, that turns the plan into a budget: effort level is the dial worth teaching people to use. support.google.com/gemini/an…

Cloudflare released two open-weight "decision models" on Wednesday, Clef and Clef-flash, under Apache 2.0. They don't write text at all. You give them a situation and a list of allowed answers, and they return a probability for every option. How it works, per the model card: Clef reads the state (text, JSON, images or video) and scores every allowed option of every question in a single forward pass. No generated text means nothing to parse. Both sit on Qwen: Clef is post-trained from Qwen3.8-27B, Clef-flash from Qwen3.5-9B. Cloudflare's own numbers: • Median latency 209.3 ms for Clef and 38.8 ms for Clef-flash, vs 524.1 ms for Typesafe AI's Jev, the model that introduced the idea • Classifying a website took Clef 2.2 s, vs 4.7 s for gpt-oss-120b in the same workflow • On Typesafe's 4 workflow evals it beats Jev on 3 and loses on agent trace observability (68.5 vs 71.6) Where it fits: routing tickets, triaging requests, labelling content. Any place an LLM is being asked to pick from a fixed list. The caveats: every benchmark here is Cloudflare's, on a leaderboard Cloudflare hosts, and on that table Clef trails Jev on When2Call (72.37 vs 80.97). The card's only hardware note is a test on a single H200 data-centre GPU. blog.cloudflare.com/clef-dec…