Recursive Flint: an Apache 2.0, 40B coding model with 2.5B active parameters and a self-reported 89.7% on SWE-bench Pro v2

Recursive Flint is the first model from Recursive Superintelligence, the startup founded by former Salesforce chief scientist Richard Socher. It went up on Hugging Face on October 21, 2026. Flint is a coding model for agents that run on your own machine: a mixture-of-experts (MoE) model with 39.8 billion parameters, of which about 2.5 billion work on each token, with Apache 2.0 weights. That puts it among the open models small enough to run on a high-end laptop. Its headline score, 89.7% on SWE-bench Pro v2, comes from Recursive itself. Our guide to coding-agent benchmarks explains what such scores do and don't measure.

This story is based on Recursive's Hugging Face model card, its launch post, its rec engine's GitHub README, and Scale AI's SWE-Bench Pro V2 leaderboard and methodology post. Recursive's pages were checked October 23, 2026, and Scale's October 22.

The short version

• Size: 39.8B total parameters, about 2.5B active per token (Recursive rounds this to "40B-A2.5B").
• License: Apache 2.0 for the weights.
• Context: 262,144 tokens on the model card; Recursive's launch post says "up to 512k," with 256K recommended.
• Runs on: an M4 Pro, M5 Pro or M5 Max Mac with at least 48 GB of memory via Recursive's rec engine, or one Nvidia B300 or H200 GPU with SGLang.
• Scores (vendor-reported): 89.7% SWE-bench Pro v2, 74.2% SWE-bench Verified, 54.9% Terminal-Bench 2.1.
• Built with AI: Recursive says its automated research system, the Eureka Machine, developed "nearly every part" of Flint's model-building stack. That's the company's claim.

What is Recursive Flint?

The model card calls Flint "Recursive's mixture-of-experts language model for coding agents." It has 49 layers, 256 routed experts per MoE layer (6 chosen per token) plus a shared expert, and mixes sliding-window and full attention. Recursive says it trained Flint from scratch, from pretraining through reinforcement learning, with a June 2026 knowledge cutoff.

Last verified: October 23, 2026

Can I run Flint on a laptop?

Yes, if it's a recent high-memory Mac. Recursive's rec engine lists an M4 Pro, M5 Pro or M5 Max with at least 48 GB of unified memory and about 30 GB of free disk. Recursive says the final 4-bit NVFP4 model "scores within 0.7 points of the 16-bit model across our benchmarks." It claims more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64 GB Nvidia RTX Spark at short contexts. Those speeds are vendor-reported.

The launch post says Flint works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang. Hugging Face listed 34 quantized versions of Flint on October 23. Recursive also serves the model free through OpenRouter as recursive/flint-v1.0:free, with the same 262,144-token context. For the memory math, see how to choose a small model for your laptop, and for setup, our Ollama guide. Flint is in the same "tiny active, large total" class as Liquid AI's LFM2.5-24B-A2B, but its Apache 2.0 license has no revenue cap.

What does Flint score on SWE-bench Pro v2?

Recursive reports these results on its model card. All were run by Recursive:

Last verified: October 23, 2026

The 89.7% does appear on the model card. SWE-Bench Pro V2 is real: Scale AI released it on September 22, 2026, with 642 tasks across 11 repositories after dropping 89 invalid tasks from the original 731. A score is the share of tasks where the model's patch passes hidden tests.

But Flint wasn't on Scale's public V2 leaderboard as of October 22, so no outside party has reproduced the number. Scale warns that the public set has "been on the open web since before these models were trained" and treats its private set as the clean measure. Recursive reports no private-set score. Its launch post claims Flint "leads MoE models with fewer than 50 billion total parameters on agentic coding benchmarks."

For context only: on October 22, Scale's leaderboard listed Kimi-K3 at 88.20% and Claude Sonnet 5.5 at 70.59%, both with mini-swe-agent. Scale's runs follow its own settings, including 50- or 250-turn limits, while Recursive allowed timeouts of up to five hours. A vendor's own run and a leaderboard run can't be ranked together. Our comparison of Claude Code, Codex and Cursor explains why such leaderboards score a model and harness together, not a product.

Did an AI really build Flint?

Recursive says its "Eureka Machine," an automated research system, helped develop "nearly every part of Flint's model-building stack," and that Flint was built "in under four months" by "a small team of people and agents." Its examples include searching expert configurations from a 35B/3B starting point, adding a distillation stage that "nearly halved the number of training steps," and writing a Metal megakernel for Apple's M4 and M5 chips. The post doesn't separate human and machine work in detail, and no outside review of the Eureka Machine has been published.

Recursive came out of stealth on May 13, 2026 with $650 million at a $4.65 billion valuation, according to SiliconANGLE, in a round led by Alphabet's GV and Greycroft.

What's confirmed, reported and uncertain

• Confirmed (model card, Hugging Face): 39.8B parameters, about 2.5B active, Apache 2.0, 262,144-token documented context, October 21 publication, hardware for SGLang.
• Vendor-reported: all benchmark scores, speeds, the 0.7-point quantization gap and the Eureka Machine's role.
• Reported: funding and valuation (SiliconANGLE).
• Uncertain: independent SWE-bench Pro v2 results; the context discrepancy (262K vs "up to 512k"); paid hosted API pricing.