AI news, October 22, 2026: OpenAI models head to Google Cloud, Recursive Flint runs on laptops, Perplexity goes local on Mac
Five stories from October 20โ21, 2026 (US time) for solo makers deciding where to run models and which tools to build on. Several items are about running AI on your own hardware; speed and benchmark figures below are the vendors' own unless the source says otherwise.
OpenAI will offer its models to developers through Google Cloud
CNBC reports that OpenAI's GPT models are coming to Google Cloud's model garden and to Google's Antigravity coding tool in the coming weeks. With Amazon already hosting OpenAI models since April, after Microsoft's exclusive license ended, this puts OpenAI on all three of the largest US cloud providers.
Why it matters for makers: if your project already runs on Google Cloud, you may soon be able to call GPT models under the same account and billing instead of managing a separate provider.
Source: CNBC
Recursive releases Flint, an open-weight coding model built for laptops
Recursive, the startup led by former Salesforce chief scientist Richard Socher, released its first model, Flint: a mixture-of-experts coding model with 40 billion total parameters, of which 2.5 billion are active per token. The weights are on Hugging Face under the Apache 2.0 license and work with llama.cpp, Ollama, LM Studio, vLLM and SGLang. Recursive says it generates more than 125 tokens per second on an M4 Pro MacBook Pro and supports up to 512K tokens of context, though it recommends staying within 256K. The company also says its automated research system helped build the model in under four months and that Flint needs more step-by-step guidance than large paid coding assistants.
Why it matters for makers: a free, permissively licensed coding model that runs offline can cut API costs for routine coding tasks and keep private code on your machine.
Sources: Recursive, TechCrunch
Perplexity Computer can now run fully on a Mac
Perplexity added a fully local mode to Computer in its Mac app, where both the model and the agent run on the device, with Qwen 3.6 or PPLX Pro (a post-trained Qwen) as model options. With permission it can work with local files and apps such as Notes, Mail, Messages and Numbers, and users decide when a task may reach out to the cloud. Perplexity recommends 36GB or more of memory, and the mode is available to Pro, Max and Enterprise subscribers.
Sources: Perplexity blog, 9to5Mac
Mistral launches Voxtral 3 speech models and a voice agent SDK
Mistral released Voxtral 3, a family of speech-to-text and text-to-speech models plus a voice agent SDK built on LiveKit that chains listening, reasoning, tool calls and speaking in one real-time pipeline. Mistral says transcription covers 50+ languages and speech output 35+ languages and 50+ dialects, with Korean among the voice preset languages. Listed prices start at $0.0035 per minute for batch transcription with Voxtral Mini Transcribe 3, $13 per million characters for Voxtral Mini TTS 3, and $0.04 per minute for the voice agent API with Mistral Small 4.
Why it matters for makers: per-minute pricing for a complete voice agent makes it easier to estimate the cost of a phone or web voice assistant before building one.
Source: Mistral AI
DeepSeek open-sources DeepInfer, the inference framework behind its production service
DeepSeek published DeepInfer on GitHub under the Apache 2.0 license, describing it as the inference framework that powers its large-scale production services. This release supports offline inference for DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. The README notes that a KV cache system, prefill/decode disaggregation and load balancing, which a complete serving setup needs, are not included.
Source: GitHub: deepseek-ai/DeepInfer