HAI Daily

Recursive's Flint Brings Coding Agents to Your Laptop, and China's Plan for AI in Industry

Recursive open-sources Flint, a small-active-parameter coding model built largely by its own AI research system, China's MIIT and seven other departments set out AI's role in strengthening industrial foundations, plus PHP2Shell and Claude Duet in brief.

By H.AI Tools Team3 stories
Products in this issue: Cline, Cursor, GitHub Copilot, Snyk, Claude
In this issue

The short version

  • Recursive's open-source Flint is a coding model with about 40 billion parameters but only about 2.5 billion active, so a 4-bit version can run on a well-equipped laptop. Its benchmark scores are self-reported.
  • Recursive says its AI research system built Flint, from tokenizer to reinforcement learning, in under four months.
  • China's plan for strengthening industrial foundations makes AI a driver across product design, simulation, testing and production.
  • Patch PHP sites for the PHP2Shell flaws now, and read our Special Edition for October 19 to 24 for details and for Claude Duet.

October 23 brought a coding model that is small enough to run on your own computer, a Chinese industrial plan that puts AI at the centre of manufacturing, and two stories we cover in depth elsewhere: the AI-discovered PHP2Shell vulnerabilities and Anthropic's proactive assistant, Claude Duet. Both are in our Special Edition for October 19 to 24.

01

Recursive's Flint: a coding model for local agents

What Recursive released

Recursive, an AI research team reportedly led by Tian Yuandong, released Flint under the Apache 2.0 licence. It is a mixture-of-experts model with about 40 billion parameters in total, of which only about 2.5 billion are active for each token, designed for coding agents that run on local hardware. A 4-bit version needs roughly 23 GB of memory, so it can run on Apple Silicon Macs and Nvidia graphics cards; reports cite speeds of about 125 tokens a second on a MacBook. The full-precision model is intended for a single data-centre GPU.

Recursive says Flint was developed from the tokenizer through pre-training, fine-tuning and reinforcement learning in under four months, largely by its own automated research system, Eureka Machine. It reports high scores on SWE-bench and Terminal-Bench, but these are self-reported and not yet independently verified.

When a local coding model makes sense

Local models matter when code cannot leave your machine, when you work offline, or when you want predictable costs. The trade-off is quality: a small active model will not match the strongest cloud models on hard problems, whatever the benchmarks say.

Set it up as a second option rather than a replacement. Run Flint through a local model server, connect it to an agent that supports local models, such as Cline, and give it bounded tasks: writing tests, explaining unfamiliar code, small refactors and scripts. Keep cloud assistants such as Cursor or GitHub Copilot for complex changes. Track which tasks the local model completes without help for a week, and route those to it permanently. Check the memory requirements before downloading, and keep the model updated as new versions appear. Our Cline guide explains how to connect different model providers.

Get HAI Daily in your inbox

The day's AI news, what it changes, and which tools to try.

02

China's plan for AI in industrial foundations

What the plan says

China's Ministry of Industry and Information Technology and seven other departments issued a plan to strengthen industrial foundation capabilities during the 15th Five-Year Plan period (2026–2030). Industrial foundations are the basic components, materials, processes and software that manufacturing depends on. The plan makes AI-enabled innovation a priority: it calls for general-purpose and industry-specific large models for these fields, for combining industrial process knowledge with agent decision models, and for applying AI across product development, design simulation, pilot testing and production. The aim is to make AI a core driver across the industrial chain rather than an add-on tool.

What manufacturers and their suppliers can take from it

The plan points to where public support and customer demand will go: AI that works with engineering data, not only with documents. If you work in manufacturing, start with one process where data already exists, such as quality inspection records or maintenance logs. Use an assistant to summarise recurring defects or failure patterns, and check the findings with your engineers before acting.

Treat process knowledge as an asset. Write down the rules experienced staff use to judge quality or diagnose faults, because those rules are what make an industrial AI system useful. Tools in the Data Analysis category help with the first analyses, and Claude can help turn interviews with experienced staff into written procedures.

03

In brief: PHP2Shell and Claude Duet

Researchers using AI, with OpenAI and Wiz among those involved, found two critical flaws in PHP that allow attackers to run code without logging in on sites such as WordPress, Drupal and Joomla. PHP released fixes on October 21, and the UK's National Cyber Security Centre urged immediate patching. If you run PHP, update now and scan your dependencies with a tool such as Snyk. Anthropic's Claude Duet, an assistant that works through your email, calendar and work apps without being asked, opened as a research preview. Both stories are covered in our Special Edition.

Not sure which tool fits your task?Browse the H.AI Tools directory by category, then compare a shortlist.

What to try next

Put one change into practice

Try a local coding model on small, bounded tasks for a week and keep it for the ones it handles well. If you work in manufacturing, pick one process with existing data and analyse it with your engineers. If you run PHP sites, check they are on a patched version today.

Explore Code Generation, AI Security and Data Analysis to build your shortlist. Our AI coding workflow shows how to combine several tools.

How we put this issue together

The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.

Go deeper