Recursive Flint on BenchGecko · open weights Apache 2.0 coding MoE with 262,144 context
Recursive released Flint on 2026-10-21 under Apache 2.0. BenchGecko lists the 39.8B MoE with about 2.5B active parameters, a 262,144 token context and a free OpenRouter endpoint.
What launched
Recursive released Flint on 2026-10-21. It is an open weight mixture of experts model built for coding agents: investigating codebases, fixing bugs, implementing features and writing tests. The BF16 checkpoint is on Hugging Face as RecursiveSI/flint under Apache 2.0. Quantized int4 and GGUF builds sit in the same org.
BenchGecko added the model and the new Recursive provider today. OpenRouter lists a free endpoint as recursive/flint-v1.0:free.
Specs
From the Hugging Face model card and OpenRouter listing:
- Parameters · 39.8B total, about 2.5B active per token including the output head, about 2.2B active body
- Architecture · decoder only MoE, 49 layers with 2 dense and 47 MoE, 256 routed experts with 6 selected plus 1 shared
- Context · 262,144 tokens for prompt and output, stored as 262,144 on BenchGecko
- Max completion on OpenRouter · 235,929 tokens
- Inputs and output · text
- Reasoning · mandatory on OpenRouter
- Tools · multi turn tool use
- Knowledge cutoff · June 2026
- Weights · open, Apache 2.0, BF16 about 80 GB
The Recursive blog says context up to 512k is supported and recommends staying within 256k. BenchGecko stores the documented 262,144 figure.
Pricing on BenchGecko
No paid first party API price was found. The live OpenRouter route is free. BenchGecko leaves the USD price fields empty rather than storing 0 as a list price.
Vendor-reported benchmarks
Recursive reports these release evaluation scores on the model card. BenchGecko records them as vendor claims. No independent scores or Gecko Tests are attached yet.
- Terminal-Bench 2.1 · 54.9%
- SWE-bench Pro v2 · 89.7%
- SWE-bench Verified · 74.2%
The blog also claims 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.
Sources
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.