Recursive, a startup built around an automated AI research system called the Eureka Machine, released its first product on October 21, 2026: a coding model named Flint. The timing lands Flint in the middle of the busiest month for open-weight coding models this year, with Google’s Gemini 4 Argon, DeepSeek’s V4.1 Flash 1017, and Inception’s Mercury 3 all expanding access within the same three-week window. For developers choosing between a local, license-free model and a metered API call, October 2026 is turning into the month the choice actually matters.
Flint ships as a mixture-of-experts model with roughly 40 billion total parameters and 2.5 billion active parameters per inference pass, released under the Apache 2.0 license. Recursive reports a score of 89.7% on SWE-bench Pro v2, a benchmark that tests whether a model can resolve real software issues pulled from live repositories rather than isolated coding puzzles. The company is pitching Flint specifically as a model built to run locally and fast, not as a cloud-hosted alternative to the subscription assistants most developers already use.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Recursive Launches Flint as a Fast, Local Coding Model
Richard Socher, CEO and cofounder of Recursive, introduced the model directly, describing it as “Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine,” in the public launch announcement. The framing is deliberate: Flint is not positioned as a general chatbot or a reasoning model that needs a large cloud cluster behind it. It is built to sit on a developer’s own hardware and answer coding questions without a network round trip.
That local-first angle matters because most of the coding-tools conversation in 2026 has been about scale, not efficiency. Reflection AI’s Beam, released as a 501-billion-parameter open-weight model with 23 billion active parameters, pushed in the opposite direction: bigger weights, more capability, and a model too large for most individual developers to run on a laptop. Flint’s 40B/2.5B footprint is a different bet entirely, chasing speed and local deployment over raw scale.
Inside the Eureka Machine: How Recursive Built Flint
Recursive’s pitch rests on the idea that Flint’s development pipeline was itself heavily automated. Socher told Wired that “nearly every aspect of the entire model pipeline was actually done and improved upon by AI,” describing the Eureka Machine as a system that iterates on its own research process rather than relying purely on human-directed experimentation. If accurate, that would mark one of the more concrete public claims yet that an AI lab is using AI to design and refine its own models at a pipeline level, not just to generate training data or review code.
Recursive has not published a full technical paper alongside the launch, so independent verification of the Eureka Machine’s actual contribution to Flint’s benchmark results is not yet available. The claim should be read as a company description of its own process rather than a peer-reviewed finding. What is verifiable is the model itself: its license, its parameter count, and its benchmark score on SWE-bench Pro v2, all of which are now public and reproducible by anyone who downloads the weights.
SWE-Bench Pro v2: What Flint’s 89.7% Actually Measures
SWE-bench Pro v2 grades a model on whether it can produce a patch that actually resolves a flagged issue in a real codebase, factoring in context most benchmark suites strip away: existing tests, dependency chains, and multi-file changes. A score in the high 80s puts Flint in contention with some of the stronger coding-focused releases this year, though Recursive’s number comes from the company’s own reporting rather than a third-party leaderboard run, which is worth keeping in mind before treating 89.7% as a settled, independently audited figure.
The benchmark context also matters because it arrives weeks after other coding-focused releases set their own bars. JetBrains’ Mellum 2.1 open coding model reported a jump to 47% on its own SWE-bench run when it launched, a very different architecture and target use case from Flint, but a useful reminder that benchmark scores move fastest at the smaller end of the model-size spectrum right now, where a single architectural change can shift results by double digits.
Gemini 4 Argon Goes Fully Public for Paying Developers
Flint’s launch did not happen in isolation. Google opened Gemini 4 Argon to every paying API customer on October 13, after initially restricting access to vetted cyber-defense partners. The model is positioned for complex, long-running software-development work, the kind of multi-step task that requires a model to hold context across many tool calls rather than answer a single prompt. Access now runs through paid developer tiers, including Google AI Studio and Android Studio, putting Argon squarely inside the same commercial channel as the rest of Google’s developer tooling.
Gemini 4 Argon already has a public track record on this site: the model was reported to have completed a full Minecraft playthrough without external help in roughly 27 hours, and it placed on the most recent general LLM leaderboard within 1.7 points of Anthropic’s Opus 5.5. Its appearance on Terminal-Bench 4.0, where Opus 5.5 posted a 65.15% score, gives developers a second, independently run data point beyond Google’s own marketing claims about long-running agentic reliability.
DeepSeek V4.1 Flash 1017 and Mercury 3 Expand the AI Gateway Race
Two more releases rounded out October’s coding-model wave. DeepSeek V4.1 Flash 1017, an update to the existing V4.1 Flash line, reports improvements in reasoning, coding, agentic task handling, and image understanding, and it is distributed through AI Gateway infrastructure rather than as a single standalone chat interface. That distribution choice matters for teams running model-routing setups, since it means Flash 1017 can sit alongside other models behind the same gateway rather than requiring a separate integration. The model follows an earlier V4.1 Flash release that scored 77.0 on a DeepSwe benchmark and shipped with its own search API, giving DeepSeek a second major coding-adjacent update inside a matter of weeks.
Inception’s Mercury 3 became newly available through AI Gateway around the same period, and the company reports throughput above 1,000 tokens per second across a range of benchmarks, a figure aimed squarely at latency-sensitive agentic workflows where a model has to make many small decisions in sequence rather than generate one long response. For teams running coding agents that issue dozens of tool calls per task, raw generation speed becomes as important as benchmark accuracy, since a slow model multiplies wait time across every step of an agentic loop.
Comparing the New Coding Models Released This Month
The table below lines up the five models that moved in October 2026, covering architecture, license, and the headline benchmark or performance figure each company has published.
| Model | Developer | Parameters (total/active) | License | Headline figure | Status as of Oct. 21, 2026 |
|---|---|---|---|---|---|
| Flint | Recursive | 40B / 2.5B (MoE) | Apache 2.0 | 89.7% on SWE-bench Pro v2 | Just released, Oct. 21 |
| Beam | Reflection AI | 501B / 23B (MoE) | Apache 2.0 | Open-weight at scale | Previously released |
| Gemini 4 Argon | Undisclosed | Proprietary API | 65.15% (Opus 5.5 on same leaderboard, Terminal-Bench 4.0) | Opened to all paying API users Oct. 13 | |
| DeepSeek V4.1 Flash 1017 | DeepSeek | Undisclosed | MIT | 77.0 on prior DeepSwe run | Updated release via AI Gateway |
| Mercury 3 | Inception | Undisclosed | Proprietary API | 1,000+ tokens/sec reported | Newly on AI Gateway |
Open Weights vs. Proprietary APIs: Who Each Model Is For
License terms are doing a lot of work in this comparison. Flint, Beam, and DeepSeek V4.1 Flash 1017 are all distributed under permissive open-source licenses, meaning a company can download the weights, host them on its own infrastructure, and avoid per-token billing entirely once it has covered the hardware cost. Gemini 4 Argon and Mercury 3 remain behind metered APIs, trading that upfront control for Google’s and Inception’s managed infrastructure, uptime guarantees, and continuous model updates without any local deployment work.
| Model | License/Access | Distribution channel | Best fit |
|---|---|---|---|
| Flint | Apache 2.0, open weights | Direct download | Local, offline coding on modest hardware |
| Beam | Apache 2.0, open weights | Direct download | Enterprise self-hosting with large GPU budgets |
| Gemini 4 Argon | Proprietary, paid API | Google AI Studio, Android Studio | Long-running agentic development workflows |
| DeepSeek V4.1 Flash 1017 | MIT, open weights | AI Gateway | Budget-conscious teams running model routers |
| Mercury 3 | Proprietary, paid API | AI Gateway | High-throughput, latency-sensitive agent loops |
Historical Context: From Copilot-Style Assistants to Open Coding Models
Three years ago, the coding-assistant market was dominated by a handful of subscription products layered on top of a small number of proprietary foundation models. That structure has been eroding steadily through 2026. GitHub Copilot retired six of its own model options this fall in favor of consolidating around newer releases, a sign that even the incumbents are under pressure to keep pace with how fast the underlying model layer is changing. Meanwhile, Cursor’s AI agents now merge 56% of all pull requests on the platform, evidence that agentic coding has moved from a novelty feature to the default way a large share of code actually ships.
Flint’s arrival fits a broader pattern that started gaining real traction in mid-2026: smaller, cheaper, open-weight models closing the gap with proprietary flagships on narrow, well-defined tasks like code repair, even as the largest proprietary models keep extending their lead on broad, general reasoning. The split is becoming structural rather than temporary. Teams that need a model to handle one job well, like resolving GitHub issues, increasingly have an open option that costs nothing per token. Teams that need a model to plan, reason, and operate autonomously across a long task still lean toward the paid frontier models.
Why Mixture-of-Experts Keeps Showing Up in New Releases
Flint and Beam both use a mixture-of-experts design, where only a fraction of the model’s total parameters activate for any given input. The appeal for a coding-focused model is straightforward: a developer gets access to a larger overall knowledge base without paying the full inference cost of a dense model that size. Flint’s 2.5-billion-parameter active footprint is small enough to plausibly run on a well-equipped workstation GPU, which is precisely the audience Recursive says it is targeting with a “fast local” pitch rather than a cloud-hosted one.
Market Impact: Pricing Pressure on Developer Tool Subscriptions
Every open-weight model that reaches competitive benchmark scores adds pressure on the subscription pricing that coding-assistant vendors have relied on. If a company can run Flint or DeepSeek V4.1 Flash 1017 on its own infrastructure and get results close to a paid API, the calculus for budget-conscious engineering teams shifts quickly, especially at the scale of hundreds or thousands of developer seats. That dynamic does not eliminate demand for proprietary models like Gemini 4 Argon, which still hold an edge on long, open-ended agentic tasks, but it does compress the price a vendor can charge for narrower, well-defined code-repair work where an open model can increasingly do the job.
The practical effect shows up in how engineering leaders plan tool budgets for 2027: expect more teams to run a mixed stack, routing routine code-fix and test-generation tasks to a cheap or free open-weight model while reserving paid frontier access for the hardest planning and architecture work. That is exactly the kind of split that AI Gateway-style routing infrastructure, now hosting both DeepSeek V4.1 Flash 1017 and Mercury 3, is built to support.
Security Risk: What the OpenCode Console Breach Means for Agent Platforms
The same week also delivered a reminder that agentic coding infrastructure carries its own exposure. OpenCode disclosed a Console breach on October 17 affecting 1.4% of Console users who had linked external AI providers to their accounts. The disclosure warned that third-party API keys stored in affected workspaces may have been exposed and could remain usable until revoked, a detail that matters specifically because coding agents are often configured with standing credentials to multiple model providers at once, multiplying the blast radius of any single platform compromise.
For teams adopting any of the new models covered here, whether through AI Gateway routing or a direct API integration, the OpenCode incident is a practical argument for rotating provider API keys on a schedule and limiting how many platforms hold live credentials at once, rather than treating a coding-agent dashboard as a low-risk piece of developer tooling.
Cloudflare’s Clairic and the Limits of “Type-Valid” Output
Cloudflare placed a model called Clairic into Workers AI preview on October 19. Rather than generating free-form text, Clairic extracts typed field values from text and up to four images, a narrower and more structured task than most coding models attempt. Cloudflare’s own disclosure included an unusually direct warning: Clairic’s type-valid outputs are not necessarily correct, meaning the model can return a value that matches the expected data type while still being factually wrong, which makes independent validation and application-level safeguards necessary before any output reaches production.
That caveat is worth generalizing beyond Clairic specifically. Every model covered in this piece, from Flint’s SWE-bench score to Mercury 3’s throughput claim, is reporting a number that describes performance under specific test conditions, not a guarantee of correctness in every production scenario a developer might throw at it.
Pi 1.0 Reaches Stable Release for Agentic Coding Harnesses
Rounding out the month’s infrastructure news, Pi reached its 1.0 stable release as a minimal AI coding-agent harness. The release adds stable Codemode and native Model Context Protocol support, along with virtual models and deferred tool loading for agent workflows. A harness like Pi sits between a raw model and a finished developer product, giving teams building their own agentic coding tools a lighter-weight foundation than a full commercial platform, and its native MCP support means it can plug into the same tool ecosystem that larger agent frameworks already use.
What Recursive’s Launch Signals About the Next Wave of Coding Models
Recursive’s pitch for Flint, that the Eureka Machine let AI improve nearly every stage of the model’s own development pipeline, is the kind of claim that will likely shape how competitors describe their own next releases, regardless of how much independent scrutiny it receives in the coming months. Socher’s framing of Flint as “a fast local coding model” rather than a general-purpose assistant also signals where Recursive thinks it can win: not against Gemini 4 Argon’s long-horizon agentic reasoning, but against the cost and latency of running any cloud model at all for routine coding work.
Predictions: Where the Open Coding-Model Race Goes Next
- Expect at least one more major lab to publish an “AI-designed its own training pipeline” claim similar to Recursive’s Eureka Machine narrative before the end of 2026, as labs compete on process transparency as much as benchmark scores.
- Mixture-of-experts architecture will keep appearing in new open-weight coding releases through early 2027, since the cost savings on inference are too attractive for compute-constrained teams to ignore.
- More coding-assistant vendors will quietly add routing logic that sends simple tasks to cheap open models like Flint or DeepSeek V4.1 Flash 1017 and reserves paid frontier access like Gemini 4 Argon for complex, multi-step jobs.
- Agent-platform security incidents similar to the OpenCode Console breach are likely to recur as more coding tools store standing credentials for multiple AI providers in one place.
- Independent benchmark verification of self-reported scores like Flint’s 89.7% on SWE-bench Pro v2 will become a bigger point of scrutiny as more vendors report their own numbers without third-party leaderboard confirmation.
Frequently Asked Questions
What is Flint and who made it?
Flint is a coding model released on October 21, 2026, by Recursive, a startup built around an automated AI research system the company calls the Eureka Machine. It is the company’s first publicly released model.
How big is Flint and how does it run?
Flint uses a mixture-of-experts architecture with roughly 40 billion total parameters and 2.5 billion active parameters per inference pass, a design meant to let it run quickly on local hardware rather than requiring a large cloud cluster.
What license is Flint released under?
Flint is released under the Apache 2.0 license, meaning the weights can be downloaded and used commercially without per-token licensing fees.
How does Flint’s benchmark score compare to other coding models?
Recursive reports an 89.7% score on SWE-bench Pro v2. That figure is self-reported by the company rather than independently verified on a third-party leaderboard at the time of release.
Is Gemini 4 Argon available to all developers now?
Yes. Google opened Gemini 4 Argon to every paying API customer on October 13, 2026, after initially limiting access to vetted cyber-defense partners. It is available through Google AI Studio and Android Studio.
What happened in the OpenCode Console breach?
OpenCode disclosed on October 17 that a Console breach affected 1.4% of users who had linked external AI providers to their accounts, warning that third-party API keys in affected workspaces may have been exposed and could remain usable until revoked.
What does Cloudflare’s Clairic model do?
Clairic, placed into Workers AI preview on October 19, extracts typed field values from text and up to four images rather than generating free-form text. Cloudflare has warned that its type-valid outputs are not necessarily factually correct.
What is Pi 1.0?
Pi 1.0 is a minimal AI coding-agent harness that reached stable release this month, adding stable Codemode and native Model Context Protocol support along with virtual models and deferred tool loading.
