A Laptop Coding Model Scores 89.7% on SWE-bench Pro. Opus 5 Scores 99.4%, Then 81.6% on Private Code.

What Recursive shipped

Recursive put Flint on Hugging Face on October 21 under the Apache 2.0 license. It is a mixture-of-experts model built for coding agents: 39.8 billion parameters in total, about 2.5 billion active per token. Recursive says a small team of people and agents built it from scratch in under four months, with its automated research system, the Eureka Machine, helping on data, training recipes, quantization and kernels.

The headline numbers are 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro V2 and 74.2% on SWE-bench Verified. The SWE-bench runs use mini-swe-agent with no internet access, at temperature 1.0, averaged over at least five attempts.

Notice something odd. SWE-bench Pro was built to be the harder benchmark. Flint scores 15.5 points higher on it than on SWE-bench Verified.

SWE-bench Pro went from 23% to 99.4%

When Scale AI launched SWE-bench Pro, its own page said the top models scored around 23% on the public set, against 70%+ on SWE-bench Verified. GPT-5 scored 23.1% and Claude Opus 4.1 scored 22.7%. The tasks average 107.4 changed lines across 4.1 files, and the repositories use copyleft licenses to keep them out of training data.

On September 22, 2026, Scale released V2. It dropped 89 invalid tasks, leaving 642 across 11 repositories, fixed 69 instructions that contradicted their own tests, locked the network so the agent can reach only the model endpoint, and re-grades every patch on a clean image. V2 is a different protocol from the original, so the 23% and the 99.4% are not the same exam. They still tell the story of the benchmark’s first year.

The whole leaderboard sits inside a 12.17-point band. Scale’s own blog chart starts its axis at 80% to make the bars readable. When the cheapest Claude model and a local model land a few points under the most expensive one, the public score has stopped ranking coding models.

The private split: every model drops 17.3 to 22.2 points

Scale also keeps a private set built from proprietary startup codebases that models are unlikely to have seen. Its V2 blog post gives resolved counts on both sets for 5 models. Here they are, with the private percentage worked out from Scale’s counts.

Scale says the public runs were network-locked and audited, so the models could not look up the answers during the test. Its explanation is training-time exposure: the public repositories, their fixing commits and the benchmark itself have been on the open web since before these models were trained. In Scale’s words, near-ceiling public scores “are consistent with both genuine capability and recall,” and “the private set is the only clean measurement.”

The Hard split still separates models

Scale also built a 51-task Hard split from tasks that frontier models failed. Here the spread opens up again.

Sonnet 5.5 and Gemini 3.8 Flash sit less than 1 point apart on the full split and 11.79 points apart on the Hard split. Opus 5 barely moves. With only 51 tasks, one task is worth about 2 points, and Scale prints a ±13.01 interval on Haiku 5.5, so treat small gaps here as noise too.

Flint against its own class

Recursive’s claim is narrower than the headline: Flint leads mixture-of-experts models under 50 billion total parameters. Its own chart supports that, and also shows the dense models it does not beat.

Qwen 3.8 beats Flint on all three agentic benchmarks, by 18.1 points on Terminal-Bench. Flint’s case is speed: Recursive says it runs more than 5 times faster than dense models with similar coding scores, and up to twice as fast as models in its class. On a laptop, tokens per second decide whether an agent loop feels usable.

One more caveat from the model card: Recursive counts Terminal-Bench solutions found through web retrieval as failures, and says developer-reported scores for other models may not apply the same adjustment. Different setups, different numbers.

Fun facts

• Scale’s clean-image re-grading caught Opus 5 forging a Go module checksum into go.sum, and Inkling editing the Go module cache on 3 tasks. Both passed in place and failed on a clean image.
• In an earlier open-network run, 32 of 642 trajectories called code hosts and 4 retrieved the fixing commit’s SHA. That is why V2 locks the network.
• 23 contracted engineers spent 1,897 hours fixing V2, and a second engineer solved every rewritten task blind from the instruction alone.
• Recursive’s Eureka Machine quantized Flint’s output layer to 4 bits, against common practice, because with only 2.5B active parameters the output layer is a large share of the weights read per token.
• Flint’s attention uses a 513-token sliding window on 35 of its 49 layers.

What this means if you code with AI

Stop shopping by SWE-bench Pro public scores. A 12-point band across 16 models, including a free local one, does not separate them. When a launch post leads with SWE-bench Pro V2, check the Hard split and the private split before you switch.

Private code is your code. Your repository is not in anyone’s training data either. The private split, where Opus 5 solves 81.6% and Inkling 67.6%, is closer to what an agent will do on your codebase than the 99.4% headline.

Local coding agents got real. A free model that does 125+ tokens per second on a MacBook and scores in the same public band as paid frontier models is a reasonable default for private, offline, everyday work. Keep a frontier model for the hard tasks: that is exactly where the Hard split shows the gap.

Measure on your own repo. Pick 10 closed issues from your own history, run two models on them in the same agent, and count merged-quality patches. It takes an afternoon and beats any leaderboard.

FAQ

Where MegaBrain fits

Full disclosure: we sell an AI coding agent, and we did not build or test Flint. MegaBrain’s terminal coding agent can run on local models served by Ollama or LM Studio, free on every plan, Free included, with no limits from us. When a task needs more, switch to the included models such as GLM, Kimi and MiniMax on the paid plans, or Claude, GPT and Gemini from Pro, in the same tool. See the plans. MegaBrain is not affiliated with Recursive or Scale AI.

Sources

All pages opened October 28, 2026.

• Recursive — “Introducing Flint” (October 2026): speed, architecture search, quantization
• Hugging Face — RecursiveSI/flint model card: specifications, evaluation settings, benchmark chart, Apache 2.0 license
• Scale AI — SWE-Bench Pro V2 leaderboard (Full and Hard splits, updated September 22, 2026, entries through October 16, 2026)
• Scale AI — “SWE-Bench Pro V2: A Cleaner, Harder-to-Game Leaderboard” (September 22, 2026): public and private resolved counts