A 40B coder that runs on a laptop, Walmart opens checkout to AI agents, and AI factories ship in containers

Recursive published Flint on 21 October as open weights on Hugging Face under the Apache 2.0 licence, which allows commercial use without paying or asking permission. Flint is a mixture-of-experts model: the Hugging Face card lists 39.8 billion total parameters, of which roughly 2.5 billion are active for any given token. Each MoE layer holds 256 experts, and a router selects six of them per token plus one shared expert. The documented context window is 262,144 tokens shared between prompt and output, and Recursive advises staying inside 256,000 for best results. The model is text-only, and its knowledge stops in June 2026.

The point of the design is speed on small machines. Because compute scales with the 2.5 billion active parameters rather than the full 40 billion, Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64GB Nvidia RTX Spark at short contexts. Four-bit builds come to roughly 23GB, and Recursive targets machines with 48GB of memory so there is headroom for a long context. The full BF16 checkpoint is about 80GB and runs on a single Nvidia B300 or H200 per replica. Flint works with Ollama, LM Studio, llama.cpp, vLLM and SGLang, and there is a free hosted endpoint on OpenRouter.

On benchmarks, Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on what it calls SWE-bench Pro v2, and 74.2% on SWE-bench Verified. The company's comparison table puts Flint ahead of other sub-50-billion-parameter MoE models — it lists Qwen 3.6 at 73.8% and Poolside's Laguna XS 2.1 at 38.6% on the same SWE-bench Pro v2 measurement. Recursive's own claim is narrower than the headline suggests: that Flint leads MoE models under 50 billion total parameters on agentic coding and 'rivals' dense models of 25 to 40 billion parameters while running more than five times faster.

Flint is also a claim about how models get built. Recursive says an internal automated research system it calls the Eureka Machine proposed ideas, ran experiments and chose next steps across data, optimisers, training recipes, infrastructure and inference, and that a small team of humans and agents built Flint from scratch in under four months. Recursive credits it with four specific findings, including settling on a 40B/2.5B-active shape, a 4-bit NVFP4 output layer that scores within 0.7 points of the 16-bit model, and a single Metal megakernel covering the whole decode step on Apple M4 and M5 chips. Humans supervised throughout and could stop the process.

The caveats are real. Every benchmark figure above is Recursive's own, averaged over at least five attempts with sampling at temperature 1.0, and 'SWE-bench Pro v2' is Recursive's own measurement, which the company says cannot be compared with the public SWE-bench Pro leaderboard. By Recursive's own table a dense Qwen 3.8 beats Flint on all three tests, and Terminal-Bench is Flint's weakest result. Recursive itself warns the model 'may need more guidance than the larger closed-source coding models'. Recursive emerged from stealth in May 2026 with more than $650 million at a $4.65 billion valuation, led by GV and Greycroft with Nvidia and AMD participating, and is run by Richard Socher and Tim Rocktäschel.

Why it mattersA permissively licensed coding model that runs offline on a high-end laptop removes both the per-token bill and the data-leaves-the-building objection that keeps many firms from using AI on their code.

Walmart published the announcement on 21 October. The core idea is that an AI agent gets its own identity inside a customer's Walmart account rather than borrowing the customer's. Walmart creates an Agent Profile automatically the first time an agent shops through the gateway; the customer then signs in to link it, sets a spending limit, and can delete it whenever they want. Walmart's own analogy is adding a cardholder to a credit card. Purchases made by that agent show up separately from the customer's other orders. Walmart says Agent Profiles arrive 'in the next month', which puts them around late November.

The second piece is the Walmart Agent Gateway, an interface built on the Model Context Protocol — the open standard for connecting AI models to external tools — so that agents query live product and fulfilment data instead of clicking through web pages, which Walmart calls slow and error-prone. The gateway carries verified agent identities, controlled permissions, category restrictions and rate limits. It is launching as an invitation-only beta for companies and developers at walmart.com/agents, and Walmart warns that access may be suspended and features may change. Agents can find a customer's usual items without receiving their broader purchase history.

Walmart also published rules. By default customers must approve agent purchases, and Walmart says it may require extra approval even for purchases that sit inside an agent's spending limit. Agents must not take actions beyond what the shopper intends. They do not receive payment-card information and cannot add or change a payment method. They may access only the information a task needs and the customer allows, and they must identify themselves as agents rather than pass as human. Agents that break the terms get blocked from the site. On price, Walmart's line is that agents see the same prices a person would, with offers and fees shown before purchase.

The context is a land grab over who controls agent-driven shopping. Walmart already has conversational commerce integrations with Meta's Muse, OpenAI's ChatGPT and Google's Gemini, alongside its own assistant Sparky, and the gateway extends agent shopping to Sam's Club as well. Amazon has taken the opposite line, blocking Meta's Muse agent from its store in September according to coverage of the announcement. Walmart's calculation appears to be that if customers are going to shop through assistants anyway, it would rather be the retailer with a supported, rule-bound path than the one being scraped.

Several things remain unclear. Walmart has not said who gets into the beta or how long invitations take, and the terms are one-sided by design: Walmart can suspend access, change features and block agents at its discretion. There are no published figures yet on how many customers use conversational commerce or what agent-placed orders look like, so the commercial size of this is unknown. Walmart's own post also misexpands MCP as 'model text protocol', a small sign of how new this plumbing still is. And returns remain Walmart's normal process, including when an agent does something the customer did not intend.

Why it mattersThis is the first big US retailer to publish identity, permission and spending rules for AI agents rather than block them, which gives developers a legitimate surface to build shopping tools against.

Nvidia announced DSX POD at GTC Berlin on 21 October. Rather than assembling an AI data centre on site, partners build it inside containers in a factory: computing, networking, rack-to-rack cabling, rack-level power distribution and liquid-cooling piping are all installed and functionally tested before the container ships. On arrival the pod connects to the site's power, cooling and network. Nvidia's claim is that this cuts on-site deployment 'from six months to as little as 14 days'. Chief executive Jensen Huang framed it in terms of revenue timing, saying an AI factory creates value when it starts producing tokens, not when construction is complete.

The pods are built around Vera Rubin, Nvidia's current platform, whose NVL72 configuration pairs 36 Vera CPUs with 72 Rubin GPUs. Tech Times reports the platform is in full production across more than 350 factories in 30 countries, and that DSX cooling supports fully liquid-cooled operation at coolant temperatures up to 45°C. Nvidia named 17 delivery partners, including Dell Technologies, HPE, Lenovo, Schneider Electric, Vertiv, Bull, Foxconn's Ingrasys, GIGABYTE, Quanta, Pegatron, Wistron, Eaton, Delta, Crusoe, DayOne, Armada and Anduril. Partners plan worldwide deliveries in early 2027.

The problem this targets is electricity, not silicon. Nvidia's own write-up on the DSX platform says more than $300 billion in equipment is backlogged and more than 200 gigawatts of projects sit in US interconnection queues. The DSX software stack is aimed squarely at squeezing more compute out of a fixed power budget: DSX MaxLPS reallocates power dynamically, DSX Flex matches workload demand to grid conditions, DSX Sim builds a digital twin of the facility, and DSX Exchange links compute to building systems. Nvidia cites partner results such as Nscale reporting 49% higher inference throughput within a 264kW budget — a vendor-reported figure, not an independent test.

The 14-day number deserves care. It covers connecting a pod at a site that is already prepared, and Nvidia is explicit that site preparation — power, cooling, networking and the rest — happens separately. Securing a grid connection is routinely the slowest part of building a data centre, and containerising the racks does nothing to speed that up. Nvidia also did not publish a per-pod rack count, GPU count, power draw or price, so the pods cannot yet be compared on cost or density with conventional builds. Coverage noted that one supplier expansion Nvidia announced the same week, IBIDEN's substrate plant, only ramps in the first quarter of 2028.

Taken with the rest of the week's GTC Berlin announcements — substrate capacity targets, a doubling of testing capacity, co-packaged optics expansion — the direction is industrial rather than scientific. Nvidia is trying to turn AI data centres into a manufactured product with a predictable lead time, and to make the grid, the building and the racks behave as one system. For everyone downstream, the practical consequence is that compute capacity may arrive in new places faster than local skills, planning permission or electricity supply can keep up.

Why it mattersIf AI data centres ship as validated containers, capacity becomes a procurement decision with a known lead time, and the bottleneck moves almost entirely to power and people.