What it is
Flint is a mixture-of-experts coding model from Recursive, the London-incorporated lab led by former Salesforce chief scientist Richard Socher, released on 21 October 2026. It has 39.8 billion parameters but uses about 2.5 billion per token, and was trained from scratch for coding agents: investigating codebases, fixing bugs, writing features and tests, with reasoning and OpenAI-compatible tool calls. The documented context is 262,144 tokens, shared between prompt and output.
Recursive reports 54.9% on Terminal-Bench 2.1, 74.2% on SWE-bench Verified and 89.7% on SWE-bench Pro v2, measured by itself. It also reports more than 125 tokens a second on an M4 Pro MacBook Pro and more than 150 on a 64 GB RTX Spark with its own rec engine. The 4-bit builds take 22.6 to 26.0 GB and need 48 GB of unified or GPU memory; the BF16 checkpoint is about 80 GB and is served on one H200 or B300 with Recursive's SGLang kit. Recursive says its automated research system, the Eureka Machine, designed much of the model.
The weights are under Apache 2.0. A free, rate-limited endpoint runs on OpenRouter; Recursive's own API is open to select partners only.
Our review
Flint is a coding model sized for the machine on your desk, and its speed is the claim to take most seriously. With about 2.5 billion of 39.8 billion parameters active per token, Recursive reports more than 125 tokens a second on an M4 Pro MacBook Pro and more than 150 on a 64 GB RTX Spark, both at short contexts on its own rec engine. For an agent making hundreds of tool calls, that throughput matters more than a few benchmark points.
The benchmarks need care. All are Recursive's, and the 89.7% is on Scale's cleaned-up SWE-bench Pro v2, where the leading agents now score above 90%, so it cannot be compared with the original Pro scores other labs report, nor read as stronger than its 74.2% on SWE-bench Verified. The card adds that its Terminal-Bench run counts web-retrieved answers as failures, an adjustment rivals may not make. Recursive's own launch post concedes Flint may need more guidance than large closed models and suggests breaking work into subtasks.
The memory figures are honest but tight. The 4-bit builds are 22.6 GB (MLX), 22.9 GB (NVFP4 GGUF) and 26.0 GB (Q4_1 GGUF), and Recursive asks for 48 GB of memory to run the full window, so a 32 GB laptop is not the target. Thirty-five of its 49 layers see only a 513-token sliding window, which keeps the cache small but leaves 14 full-attention layers to carry all long-range recall; the card calls the context a capacity limit, not a promise of accuracy.
The lab behind it is not a hobby project. Recursive came out of stealth in May with $650 million in funding, according to i-scoop, and says the Eureka Machine helped with architecture, data, training recipes and kernels in under four months. A free BF16 endpoint on OpenRouter lets a team try Flint before downloading 80 GB, but the paid API is partner-only and unpriced. In i-scoop's reading of Recursive's chart a larger dense Qwen model still scores higher, which makes Flint the faster local option rather than the stronger one.
Where it earns its place
A backend developer on a 48 GB M4 Pro MacBook Pro
Run a coding agent on a private codebase without sending code to an API
Pull the 23 GB Ollama build and point OpenCode, Cline or Claude Code at it (`ollama launch claude --model flint`). There is no per-token bill and nothing leaves the laptop; expect to split larger tasks into steps, as Recursive itself advises.
A platform engineer with one H200 or B300
Serve a shared coding model to an internal team behind an OpenAI-compatible endpoint
Recursive's Docker SGLang kit runs the 80 GB BF16 checkpoint on one GPU per replica; add replicas as users grow. Transformers modelling code is not yet included, so SGLang or vLLM is the supported path.
An engineering manager at a 20-person software agency
Find out whether local models can cut the team's coding-assistant API spend
Trial the free, rate-limited OpenRouter endpoint on the team's real tickets first, then price 48 GB laptops against current API bills. The weights are free, but with the paid API partner-only there is no hosted fallback price to budget around.
Before you commit
- Setup
- `ollama run flint` pulls 23 GB; on a supported Mac, Recursive's rec engine downloads the model and wires up Pi or OpenCode. llama.cpp needs a release with Flint support, and GPU serving uses the bundled SGLang Docker kit.
- Integrations
- OpenAI-compatible chat and tool calls; documented harnesses are OpenCode, Pi, Cline and OpenHands, and Ollama can launch Claude Code against it. The card says to keep returned reasoning_content in conversation history.
- Data and privacy
- Run locally, nothing leaves the machine. The OpenRouter endpoint is served by Recursive; the card says hosted terms and data practices are separate from the weight licence, and we found no published retention terms.
- Cost at scale
- Weights free. Locally, a Mac or GPU with 48 GB of memory for the 4-bit builds; on servers, one H200 or B300 per BF16 replica (about 80 GB). Hosted: free and rate-limited on OpenRouter; no paid per-token price published.
- Limits
- Text only; prompt and output share the 262,144-token window. Hard problems want an output budget of at least 32,768 tokens and long reasoning can exhaust it; knowledge cutoff is June 2026.
What's good
- Over 125 tokens a second on an M4 Pro, per Recursive
- 4-bit builds of 23 to 26 GB
- Apache 2.0, plus a free OpenRouter endpoint
- Ollama, llama.cpp, LM Studio, SGLang and vLLM at launch
What's not
- All benchmarks self-reported
- SWE-bench Pro v2 score not comparable with other labs' Pro
- Needs 48 GB of memory for the full context
- Paid API partner-only, no published price
Best for
Developers who want a fast coding agent model that runs locally on a Mac or a single GPU.
Alternatives
Flint: common questions
- Is Flint worth it?
- A coding model genuinely sized for a 48 GB laptop, from a well-funded lab, with every headline score still Recursive's own. We rate it 3.9 out of 5.
- How much does Flint cost?
- Free open weights under Apache 2.0; free, rate-limited endpoint on OpenRouter. Paid Flint API for select partners, price not published.
- What are the downsides of Flint?
- All benchmarks self-reported; sWE-bench Pro v2 score not comparable with other labs' Pro; needs 48 GB of memory for the full context; paid API partner-only, no published price.
- Who is Flint best for?
- Developers who want a fast coding agent model that runs locally on a Mac or a single GPU.
- How much does Flint cost for a team or in production?
- Weights free. Locally, a Mac or GPU with 48 GB of memory for the 4-bit builds; on servers, one H200 or B300 per BF16 replica (about 80 GB). Hosted: free and rate-limited on OpenRouter; no paid per-token price published.
- Is Flint safe for company data?
- Run locally, nothing leaves the machine. The OpenRouter endpoint is served by Recursive; the card says hosted terms and data practices are separate from the weight licence, and we found no published retention terms.
- What are the best Flint alternatives?
- The closest alternatives we have reviewed are Qwen3.8-27B, Devstral Small 2 and gpt-oss-20b.
Answers checked Oct 23, 2026.