1 The week2 Build3 Launches and opportunities1T params Image: Mistral AI Open weights Mistral has published the weights of Large 4, a multimodal mixture-of-experts model it calls the strongest open model for coding, under a licence that restricts hosted-service use.
For a small team, this is a frontier-scale model that can be run, fine-tuned or inspected on infrastructure you choose, rather than rented from a single vendor. It is not a laptop model: Mistral's quantised version runs on four Nvidia B200 or B300 GPUs, so most builders will reach it through Mistral's API, OpenRouter, Microsoft Foundry or Hugging Face's inference providers. Mistral says three further weeks of reinforcement learning lifted its Terminal-Bench 4.0 score from 28% to 41%; that figure is the company's own, and independent results are still to come. Check the licence before building a product on it: it is an MIT derivative that limits model-as-a-service offerings, which matters if you plan to resell hosted access. Fine-tuning is supported through TRL.
Parameters About 1T total, 49B active
Context window 1M tokens
Licence Mistral Ecosystem Open-Weight License
Quantised version runs on Four Nvidia B200 or B300 GPUs Why it matters A downloadable frontier-scale model gives small teams an alternative to closed APIs, provided the licence terms and hardware costs fit their plans.
Part 1 of 3 The week The week’s story for builders, the rest of the week in brief, and the licences, security issues and breaking changes to watch.
Image: Google Google DeepMind announced Gemini 4 Carbon, built for long-running agentic engineering work, two weeks after Argon. It scores 59 to lead Artificial Analysis's Intelligence Index v5 and launches at $2 per million input tokens and $10 output, with access going first to cyber defenders.
Input price $2 per million tokens
Output price $10 per million tokens
Cached input 95% off input price
Intelligence Index v5 59 Why it matters Builders running long agent loops get a new top-ranked option, though access is staged and the launch price is introductory.
Image: SiliconANGLE OpenAI, Google, Microsoft and Anthropic OpenAI, Google, Microsoft and Anthropic say they will honour a shared robots.txt signal, drawn from IETF AI Preferences work, that lets a site refuse generative AI training with one entry. OpenAI says existing crawler controls remain; Microsoft is adding a separate setting for AI grounding.
Basis IETF AI Preferences working group
Rollout OpenAI and Microsoft in coming months; Google next year Why it matters If you publish content or run crawlers, one rule now governs training access across major providers, separately from search indexing.
Image: OpenAI OpenAI now lets eligible ChatGPT Pro and Business subscribers claim monthly API credits, reported at up to $500, by linking the subscription to an API organisation they own. Credits renew with the subscription, are spent before purchased credits, and exclude Codex use through API keys.
Maximum monthly credit $500
Eligible plans ChatGPT Pro and Business
Excluded Free trials and complimentary subscriptions
Not covered Codex usage through API keys Why it matters Subscribers already paying for ChatGPT can prototype apps and agents on the API without a separate payment method, within rate limits.
Image: Hugging Face IBM released Granite Switch 4.1.1, a backwards-compatible patch to its modular model that composes a base model with swappable adapters on one endpoint. It adds an updated speech-recognition model covering 10 languages and an improved decision adapter, while the base model is left untouched.
Model Granite Switch 4.1.1 30B
Speech languages 10
Migration None; backwards compatible with 4.1 Why it matters Adapter-based updates let teams take new capabilities without migrations, which suits small deployments that cannot afford frequent model swaps.
A tampered Claude Code extension, version 2.1.313, was available on Cursor's marketplace on 29 October and tried to install a persistent SSH backdoor; at most 5,000 users were affected. Anthropic says it was not involved in publishing it, and version 2.1.314 omits the malicious code.
Malicious version 2.1.313
Clean version 2.1.314
Exposure window About 11 hours
Users affected At most 5,000 Image: BankInfoSecurity Why it matters If you installed it in Cursor, updating is not enough: check for added SSH access and rotate any credentials on the machine.
Image: Unite.AI Cloudflare says a mismatch between GLM-5.3's 1,048,576-token context window and a smaller prefill cache pool triggered an SGLang bug on Workers AI. From the model's launch on 27 August, 155 responses to very long prompts contained data from other customers' sessions; Cloudflare worked with SGLang on remediation.
Context window 1,048,576 tokens
Prefill KV pool 681,280 tokens
Affected responses 155 Why it matters Anyone self-hosting with SGLang in disaggregated setups should check that KV cache pools exceed context length on prefill and decode.
Image: GitHub Gist A developer's analysis found Claude Code sends each background agent's entire transcript to its model roughly every 30 seconds to generate a 3-5 word status label. The author estimates this uses about 15–20% of tokens; a cheaper short mode exists but is off by default.
Interval ~30 seconds per background agent
Estimated token share ~15–20%
Short mode ships in 2.1.310, off by default Why it matters Heavy users of background agents may hit usage limits sooner than expected; the analysis lists settings that stop the behaviour.
Part 2 of 3 Build The machines to run it on, open models to run yourself, trending repositories, changes to tools and APIs, and one thing to build this weekend.
Image: Macworld Apple has begun shipping the Mac Studio with M5 Ultra and 512GB of unified memory, about a month after the rest of the range. Macworld puts the minimum price at $22,299, which is $11,500 more than the same machine with 256GB.
Minimum price $22,299
Chip M5 Ultra
Premium over 256GB $11,500
Two 256GB units $21,598 Why it matters It brings 512GB of memory for large local models to one desktop, though Macworld notes two 256GB machines cost less.
Image: The Manila Times (GlobeNewswire) BrainChip has launched a development kit that combines a Raspberry Pi 5 with its AKD1500 M.2 board for low-power, always-on AI. The event-based chip can learn new classes on the device, such as a keyword or gesture, without cloud retraining, and the kit is on sale now.
Host board Raspberry Pi 5
Accelerator AKD1500 M.2 board
Availability Now, through the BrainChip Shop
Production interfaces PCIe or SPI Why it matters Raspberry Pi builders get an off-the-shelf route to always-on, on-device learning that carries over to production hardware.
An NVIDIA blog post on Hugging Face describes MonsterHouse, a talking garage-door projection built for Halloween. It chains Nemotron speech recognition, the Nemotron 3.5 Bolt E4B 4GB model, Magpie text-to-speech and Audio2Face-3D, and runs entirely on a single workstation as a self-contained system.
Speech model nemotron-speech-streaming-en-0.6b, q8_0 GGUF
Language model Nemotron 3.5 Bolt E4B 4GB
Hardware Single workstation Why it matters It offers a worked recipe for a fully local voice agent, combining open speech, language and animation models on one machine.
SheepSeb's tinyworldformer runs two small networks on a Raspberry Pi Pico 2 W: a looped-transformer world model that dreams each game frame and an actor that plays it. The world model is int8-quantised, the firmware is no_std Rust, and a WebAssembly build runs in browsers.
Hardware Raspberry Pi Pico 2 W (RP2350)
Games Five
Quantisation int8 world model, 32-bit float actor Image: GitHub Why it matters It is a compact, readable worked example of training, quantising and exporting PyTorch models to embedded Rust on a microcontroller.
Image: GitHub makecindy/cindy, an Apache-2.0 desktop and mobile client for AI agents, gained 1,465 stars in a day on GitHub Trending, a DEV Community post recorded. It wraps agents such as Claude Code and Pi, lets sessions continue across devices, and offers an optional managed model service.
Licence Apache License 2.0
Stars in a day 1,465
Clients Desktop and mobile Why it matters Builders want one interface over several agent harnesses; check what the hosted service and device linking send off your machine.
Image: Hugging Face mattt/AnySpeechRecognizer Mattt's AnySpeechRecognizer keeps Apple's SpeechAnalyzer interface from macOS 26 and iOS 26 but lets apps swap in NVIDIA's Parakeet or Alibaba's Qwen3-ASR on Core ML, and adds speaker diarisation. Its sibling AnyDecisionModel returns typed yes-or-no, multiple-choice and rating answers with probabilities.
Platforms macOS 26 and iOS 26
Engines Parakeet, Qwen3-ASR
Requirement Apple silicon for Parakeet, Sortformer and MLX Why it matters App developers can transcribe and classify on the device without a cloud API, keeping familiar Apple code paths.
Teach a 2B model to read receipts into JSON Liquid AI's LFM3 release includes a fine-tuning notebook that trains LFM3-VL-2B to turn receipt photos into structured JSON. You can run it in a free Colab session and then use the result on your own machine through tools that support LFM3.
Open the LFM3 post on Hugging Face and launch the 'Read receipt photos into JSON' Colab notebook. Gather twenty or so receipt photos of your own and write the JSON you want for each. Run the supervised fine-tuning cells on LiquidAI/LFM3-VL-2B with your examples added. Compare outputs from the base and tuned models on receipts held back from training. Export the tuned weights and try them locally in llama.cpp, Ollama, LM Studio or MLX. Part 3 of 3 Launches and opportunities The week’s new startups, what launched, and what is open to enter, with its deadline.
Dream is a San Francisco startup founded by two former Palantir executives, and it has come out of stealth with $17M in seed funding. It builds what it calls built world models. These are trained on 100 million square feet of permitted US buildings and used to design and engineer buildings.
Round Seed
Amount $17M
Investors Andreessen Horowitz (a16z), SV Angel, XYZ, Zigg Capital
Training data 100 million square feet of permitted US buildings Image: TechCrunch Why it matters World models are moving beyond games and robotics into regulated trades, and that opens room for specialist tools built on domain data.
Image: EU-Startups Stockholm-based Pixel Systems was founded in 2026 and has raised €2.7 million in a seed round led by Unconventional Ventures. Its software detects AI-generated and manipulated images and documents. The company says its internal tests show a 99.98 percent detection rate, a figure that has not been independently verified.
Round Seed
Amount €2.7 million ($3.1 million)
Lead investor Unconventional Ventures
Pre-money valuation approximately €18 million ($21 million) Why it matters As generated documents get cheaper to make, builders handling uploads, claims or identity checks will need ways to verify them.
Image: EU-Startups Madrid-based Talky Labs has raised €1.5 million in pre-seed funding. Its AI learns each company's unwritten accounting rules from historical records and from its finance teams. HelloWorld, Silence VC and Motive Force took part, and the startup says it handles 86% to 93% of documents without human intervention.
Round Pre-seed
Amount €1.5 million
Investors HelloWorld, Silence VC, Motive Force
Claimed automation between 86% and 93% of accounting documents Why it matters Finance back-office automation is drawing steady early-stage money in Europe, which suggests it is a busy but funded niche for small teams.
Opera has launched WebAI, which lets websites and browser extensions call AI models running on the user's own device. These include open-weight models on Opera's inference engine. Page data need not leave the machine and developers pay no per-request cloud costs. Users switch the feature on in settings.
Runs on Opera browser, on the user's device
How to enable Settings → Opera AI → Enable Web AI API
Documentation next.opera.com/web-ai Image: Opera Why it matters Web developers get a way to add AI features without running servers or sending user data to a cloud provider.
Trigger.dev has released inbound webhooks in public beta on both its cloud service and its open-source project. Events from Slack, Stripe, GitHub and other services can now start AI agents directly. It also added new observability features, and the launch was listed on Product Hunt's daily leaderboard.
Status Public beta
Availability Trigger.dev Cloud and the open source project Why it matters Linking agents to everyday tools like Slack or Stripe events is often fiddly plumbing that small teams would rather not build themselves.
Simantic is part of Y Combinator's F26 batch and has launched on Hacker News. Its firmware simulator, written from scratch in Rust, lets AI coding agents test embedded code without flashing physical boards. The team says it prioritised accuracy, and it also offers an AI firmware-review bot for GitHub repositories.
Built in Rust
Backer Y Combinator (F26) Why it matters Hardware is where coding agents struggle most to check their own work, and a fast test loop could make them usable there.
SR008 applications close this Sunday, Nov 1, at 11:59 PM PT
Andreessen Horowitz (a16z speedrun) Andreessen Horowitz's speedrun accelerator is taking applications for its SR008 cohort. It invests up to $1M per startup, aims to close within two weeks and offers $10M+ in partner credits. The residency runs in San Francisco from February to April 2027.
Why it matters Solo founders and small teams building on AI can still get early capital and credits, but the deadline is very close.
November 2, 2026, at 8:00 PM PT
Y Combinator Y Combinator is accepting applications for its Winter 2027 batch, which runs in person in San Francisco from January to March 2027. Each accepted startup receives $500,000. The on-time deadline is Monday evening Pacific Time, and decisions are expected by December 11.
Why it matters It is one of the largest standard cheques open to very early teams, though it requires being in San Francisco in person.
