Recursive released Flint on October 21, 2026, a free coding model designed to run locally on developers' computers under Apache-2.0 licensing. The company claims its Eureka Machine automated research system helped a small team build the model in under four months.
Flint uses a mixture-of-experts architecture with roughly 40 billion total parameters and approximately 2 billion active parameters per token. Mac deployment requires an M4 Pro, M5 Pro, or M5 Max with at least 48GB unified memory. The model supports a 262,144-token context window and reportedly generates about 131 tokens per second on compatible hardware, though Recursive's benchmarks remain vendor-reported and unindependently verified.
Recursive released Flint on October 21, 2026, offering a free coding model designed to run on a developer’s own computer. The company says its automated research system, the Eureka Machine, helped a small team build the model from scratch in under four months, according to TechCrunch’s launch report.
The immediate benefit is practical: local coding assistance can work offline, keep prompts and code on the machine, and avoid cloud inference charges. The published Mac setup also has a substantial entry requirement: a supported Apple chip and at least 48GB of unified memory.
Flint is a concrete model release with documented operating boundaries. Recursive’s wider ambition, building AI that improves the process of making AI, remains a separate claim worth evaluating on its own evidence.
What Flint’s Free Local Release Offers
Flint’s weights are available through Recursive’s Hugging Face repository under Apache-2.0. That license allows use, modification and redistribution subject to its terms, giving developers room to integrate the model into their own tools and workflows.
Local inference removes the need to send every coding prompt to a hosted model. It also replaces per-token service charges with costs borne by the developer: hardware, storage, electricity and setup time. “Free” describes access to the model, not a cost-free deployment.
Privacy depends on the surrounding workflow. Keeping Flint’s inference server local helps, but a coding harness can still invoke networked tools or services. Developers handling confidential repositories should check where tool calls, logs and other integrations send data.
My view is that this control is Flint’s clearest attraction. A developer with suitable hardware can choose where inference happens and how the model connects to a repository. That is useful independently of whether Flint matches the strongest paid coding assistants.
Small Active Parameter Counts Still Need Memory
Flint uses a mixture-of-experts, or MoE, architecture. Different parts of the model handle different inputs, so each generated token activates only a fraction of its total parameters.
The release materials describe that active fraction differently. Recursive’s homepage lists roughly 2 billion active parameters, while TechCrunch reports 2.5 billion. Both describe a model with about 40 billion parameters overall; the supplied evidence leaves the active-count discrepancy unresolved.
Work with Zeniteq
Promote your app with us
Reach readers exploring AI tools and workflows. Discover sponsorship and advertising options for your app.
BF16 uses approximately two bytes per parameter, which explains why roughly 39.8 billion parameters produce around 79.6GB of tensor files. Sparse activation does not make the remaining weights disappear from the model package.
That creates an important buying and setup distinction. The large BF16 checkpoint and the optimized 48GB Mac workflow describe different deployment conditions. Developers need the appropriate runtime and weight format, plus room for working memory and the conversation cache. The documented evidence does not specify enough detail to calculate every optimized configuration’s footprint.
I’d use the documented 48GB requirement, not the active-parameter count, when deciding whether an existing Mac is a sensible candidate. “About 2 billion active parameters” describes computational sparsity; it is a poor shortcut for estimating the whole deployment’s memory needs.
The Coding Harness Does the Executing
Recursive documents Flint for repository understanding, code generation, edits and tests. Its local-use guide draws a useful boundary: the model generates text and tool requests, while a compatible runtime and coding harness execute those requests and return results.
That distinction explains how an agentic coding session works. Flint can request a file read, propose an edit or ask for a test command. The harness provides the connection to the filesystem and shell.
For supported Macs, Recursive recommends its optimized rec inference engine and the Pi coding harness. After installing rec, users launch it from their project directory:
rec launch pi
The command starts the model server and configures Pi. Recursive also documents rec launch opencode for developers who prefer OpenCode. These are documented launch commands, not commands tested for this article.
The harness boundary matters for permissions and supervision. A local model that can request shell commands still needs a carefully configured execution environment. Local inference changes where the model runs; it does not remove the consequences of a mistaken edit or an unsafe command.
Context Length Depends on the Local Configuration
Recursive’s hosted Flint configuration supports a 262,144-token context window, with an output budget of up to 65,536 tokens. Those numbers should not be treated as a promise for every laptop deployment.
The context contains more than source files:
The prompt and conversation history.
Tool definitions and returned results.
The model’s reasoning and final answer.
For local inference, available memory, quantization and cache settings determine how much fits. Recursive recommends retaining the context and output settings supplied by the rec launcher. Users serving through llama.cpp should set --ctx-size and configure their harness to match.
Flint’s reasoning is always on, according to the documentation. Its output allowance therefore needs room for both reasoning and the visible answer. A response ending with finish_reason: length may need a larger permitted output budget or a shorter conversation.
A large advertised context window is useful only when the deployment can support it. Repository work can fill that budget with tool output and conversation history before a developer has explicitly supplied much code.
About 58–69 tokens per second for models including Gemma 4 and Nemotron 3.5 Lightning
Throughput within the company’s comparison
SWE-Bench Verified
Ahead of other models below 50 billion parameters
Performance on real-world software bug-fixing tasks
These are vendor-reported results, not independently reproduced measurements. The opened evidence lacks complete runtime configurations, exact SWE-Bench scores and enough evaluation detail to reproduce the comparisons.
Generation speed and coding success also answer different questions. Tokens per second indicate how quickly output arrives once generation is underway. A completed bug fix may require repository searches, several edits, tool execution and repeated tests. Faster text generation helps, but it cannot predict total task completion time by itself.
The benchmark lead is bounded by the stated model-size category. TechCrunch also reports that Alibaba’s Qwen3.8 27B scored higher on Recursive’s own average across six tests, illustrating how rankings can change with the chosen evaluation.
Recursive acknowledges that Flint may need more guidance than large paid coding assistants and recommends breaking larger tasks into smaller steps. That suggests a sensible starting point: scoped fixes with clear acceptance criteria, such as a failing test, a contained refactor or a specific feature change. It gives developers a way to judge usefulness against their actual work without assuming one leaderboard result guarantees broad autonomy.
Eureka’s Role Is an AI-Assisted Development Claim
Recursive says the Eureka Machine contributed across Flint’s development, from pretraining through post-training and inference. TechCrunch reports that CEO Richard Socher attributed contributions including a sparser model design, low-bit training and custom kernels to the system.
Those contributions address distinct parts of model development. Architecture affects which computations occur, lower-bit representations can reduce storage and computation demands, and custom kernels can speed up underlying operations. Together, they offer a plausible route to a more efficient local model.
The company’s claim is broader than using an assistant to write occasional code. Eureka reportedly proposes ideas, runs experiments and uses the results to choose subsequent work. Recursive says this helped its small team deliver Flint in under four months.
The missing measurement is how much improvement came specifically from Eureka. The published reporting does not quantify researcher time saved or provide a comparable development run without the system.
Recursive’s homepage describes its ambition as a “self-improving superintelligence that automates the scientific method.” Flint provides a bounded example of AI-assisted model development under human oversight. Demonstrating a sustained recursive improvement loop would require evidence across successive cycles, showing how one generation improves the research system and the next model.
For developers, the shipped boundary is already clear enough to evaluate: Flint generates code and tool requests, a harness executes actions, and people supply goals, permissions and guidance. Its local-use value can stand on that foundation while Recursive works toward its larger research ambition.
Frequently Asked Questions
4 questions
1
Is Flint Free for Commercial Use?
Flint is released under Apache-2.0, which permits commercial use, modification and redistribution subject to the license’s terms. Local hardware and operating costs still apply.
2
What Mac Does Flint Require?
Recursive’s optimized rec workflow documents an M4 Pro, M5 Pro or M5 Max with at least 48GB of unified memory. Those requirements apply to that supported deployment route.
3
Can Flint Run Commands by Itself?
Flint generates tool requests. A compatible coding harness executes them and returns results, so permissions and execution controls belong to the surrounding setup.
4
Does Flint Prove Recursive Self-Improvement Works?
The published evidence supports Recursive’s claim of AI-assisted development with human oversight. It does not demonstrate a sustained, autonomous cycle of models improving their own successors.
