Recursive Flint: a small, fast coding model from a US lab

Recursive, Richard Socher's US lab, released Flint on 21 Oct under Apache-2.0: 39.8B parameters in total, about 2.5B active, 25.1 GB at 4-bit. Its benchmark scores are vendor-reported only so far.

What happened

Recursive, the US lab led by Richard Socher, released Flint on 21 October 2026, with weights published under Apache-2.0. TechCrunch reports the company raised $650M at a $4.65B valuation.

Key facts

• Architecture: mixture of expertsMixture of experts (active parameters): A design where only part of the model (the active parameters) works on each piece of text. Glossary, 39.8B total parametersParameters: The learned numbers inside a model. More parameters usually means a more capable but larger model. Glossary, about 2.5B active, 49 layers, text only
• Context: 262,144 tokensTokens: The small chunks of text a model reads and writes, often a word or part of a word. Glossary documented; the vendor recommends staying within 256k
• Licence: Apache-2.0
• Vendor-reported scores: Terminal-Bench 2.1 54.9%, SWE-bench Verified 74.2%
• Independent scores: none yet on Artificial Analysis or LMArena
• Download sizes (unsloth/flint-GGUF): 25.1 GB at 4-bit (UD-Q4_K_XL), 42.3 GB at 8-bit (Q8_0)
• Runtimes (vendor): llama.cpp, Ollama, LM Studio, vLLM, SGLang

How it compares

Against Qwen3.8-27B on benchmarks both report, Flint scores lower on Terminal-Bench 2.1 (54.9 vs 73.0) and LiveCodeBench v6 (80.3 vs 90.3), according to OrcaRouter's comparison. Its edge is speed and memory: about 2.5B active parameters against Qwen's 27B dense.

Our view

Flint (Recursive, US) is a useful option at the small end, and its size means it runs on workstation-class hardware rather than servers. Treat its coding quality as unproven until independent scores appear.

Why it matters: A small Apache-2.0 model that runs on a single workstation, though its coding quality is vendor-reported only so far.

Source: Recursive, “Introducing Flint”, 21 Oct 2026