R

Recursive Flint

recursive · Released Oct 2026

Conditional

Conditional. Flint is an Apache 2.0 open-weight coding model built by Recursive for agentic coding on local hardware, with weights on Hugging Face and a capacity-limited free hosted tier through OpenCode. The deciding caveat is that every benchmark and speed figure is Recursive's own, and Recursive says it may need more guidance than larger closed coding models. Suits developers who want a self-hosted coding agent and will test it on their own repositories.

Is it right for you?

Good for

  • Self-hosted agentic coding under a permissive licence: Apache 2.0, weights on Hugging Face, with Transformers, vLLM, SGLang, Ollama, LM Studio, Unsloth and llama.cpp GGUF paths listed in Recursive's docs
  • Laptop and workstation coding agents: about 40B total and 2.5B active parameters per token, and Recursive's blog reports more than 125 tokens per second on an M4 Pro MacBook Pro at short contexts (vendor-reported)
  • Trying it without hardware: Recursive's docs say OpenCode offers a hosted Flint-40B on its free tier, subject to capacity and rate limits, and Pi offers a free trial through Radius
  • Agent harness work: Recursive's docs list Cline, OpenHands and Pi (via Radius) integrations

Not good for

  • Unguided work on hard tasks; Recursive's blog says it 'may need more guidance than the larger closed-source coding models'
  • Running the BF16 weights on a laptop; the card lists about 80 GB of weights and one B300 or H200 GPU per replica, so laptop use means the quantized variants (int4 and nvfp4 GGUF repos exist on Hugging Face)
  • Treating the benchmark table as independent evidence; all scores are Recursive's own, averaged over at least five attempts with its stated harnesses
  • Long reasoning on tight output budgets; the card warns long reasoning can exhaust the output budget and recommends at least 32,768 output tokens for complex tasks

How it performs by task

Agentic coding

Good

Recursive reports 54.9% Terminal-Bench 2.1 (Terminus 2), 74.2% SWE-bench Verified and 89.7% SWE-bench Pro v2 (mini-swe-agent), average pass@1 over at least five attempts; vendor-reported, untested by us

Local and edge inference

Good

Small active parameter count and a documented local engine plus Ollama, LM Studio and llama.cpp paths; the speed figures are Recursive's own

Pricing

Input

Free (Apache 2.0)

Output

Free (Apache 2.0)

Context

Open weights; self-host or free OpenCode tier

View full pricing

Benchmarks

BenchmarkScoreSource
Terminal-Bench 2.1 (Terminus 2, Recursive-reported)54.9% Source
SWE-bench Verified (mini-swe-agent, Recursive-reported)74.2% Source
SWE-bench Pro v2 (mini-swe-agent, Recursive-reported)89.7% Source

No verdict changes yet

The clock starts day one. Changes land here as our verdict evolves.

Verification log

  • Status· No changes

    Automated agent

  • Pricing· No changes

    Automated agent

  • Profile· No changes

    Automated agent

  • Pricing· No changes

    Automated agent

How we evaluate