Tasks and inputs
Flint is listed for Text generation, Coding & Code Generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
Recursive Flint is an open-weight 39.8B MoE coding-agent model with 2.5B active parameters, free to try via OpenCode and OpenRouter.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Free tiers only (OpenCode, OpenRouter :free). No paid public API; direct access limited to select partners.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Recursive Flint was trained from scratch, which is rarer than it sounds in a year of fine-tunes. It is Recursive's mixture-of-experts model for coding agents: 39.8B parameters, about 2.5B active per token, released with open weights under Apache-2.0. That active count is tiny. Recursive cites over 125 tokens per second on an M4 Pro MacBook Pro.
So the question Flint raises is simple. How much agentic coding can a laptop-class active footprint do?
These are publisher-reported. The Terminal-Bench result sits well below the large frontier models, while the SWE-bench scores are strong. Expect it to be better at fixing code in a repository than at long shell sessions.
Flint has 49 layers. Thirty-five use sliding-window attention over 513 tokens and fourteen use full attention, a mix that keeps long-context memory affordable. There are 256 routed experts with six chosen per token. It supports reasoning and OpenAI-style tool calls, and its stated knowledge cutoff is June 2026.
| Item | Detail |
|---|---|
| OpenRouter ID | recursive/flint-v1.0:free |
| Hugging Face | RecursiveSI/flint |
| Parameters | 39.8B total, ~2.5B active |
| Context | 262,144 tokens (card) |
| Licence | Apache-2.0 |
| Recommended output budget | 32,768+ tokens for hard problems |
| Price | Free tiers only |
Today it is free. OpenCode hosts Flint-40B, capacity-limited, and OpenRouter offers a free endpoint. Direct API access from Recursive is limited to select partners. For production, self-hosting is the realistic route: the bundled SGLang kit runs one replica on a single H200 or B300.
Context first. Recursive's announcement mentions up to 512K, with a recommendation to stay within 256K; the model card documents a 262,144-token configuration. Use 262,144. Output limits differ too: one registry says 262,144, while OpenRouter caps completions at 235,929 tokens. Dates vary: the blog post is dated 18 October 2026, while the Hugging Face repo and OpenRouter listing say 21 October.
Give it enough room first. Recursive recommends at least 32,768 output tokens on hard problems, and many harnesses default much lower. Then run your own task suite twice, once with Flint and once with your current model, and record solved tasks, retries and wall-clock time. If Flint solves most of what your larger model solves, the cost of a self-hosted single-GPU replica is very hard to beat. Compare it with Qwen3 Coder Next and other code generation models.
Compare the model on the work you need to complete.
Flint is listed for Text generation, Coding & Code Generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 262.14K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Gemini 4 Hydrogen is Google's fast, lower-cost Gemini 4 model for everyday coding and agentic execution, with a 1M context, xhigh thinking and a free tier.
OpenAI
GPT-6 Cyber is OpenAI's restricted security model for authorised vulnerability investigation, reproduction and fix verification, for Daybreak Red customers.
particle-ai
Particle-1 is Particle's first software-engineering model, an RL-tuned DeepSeek V4.1 Flash that aims to solve coding tasks with fewer tokens.
Microsoft
MAI-Cyber-1.1-Flash is Microsoft AI's vulnerability-discovery model, available only inside its MDASH scanning harness to verified security customers.
Deepreinforce
Ornith-1.5 122B is an open MIT-licensed MoE coding and agent model with a 262K context, reasoning by default and single-80GB-GPU serving.
MiniMax
MiniMax M3.1 Flash is MiniMax's GA open-weight multimodal MoE for coding agents: 428B total, 23B active, 1M context, from $0.15 per 1M input tokens.