Flint

Flint

Quick answer: Flint is Recursive's open-weight 40B-A2.5B mixture-of-experts model, purpose-built for fast agentic coding on local hardware (Apple Silicon, NVIDIA RTX Spark). It scores 74.2% on SWE-bench Verified and 89.7% on SWE-bench Pro v2, with a 512K-token context window, under Apache 2.0.

At a Glance

Where Flint leads

• Best-performing MoE model under 50B total parameters on agentic coding benchmarks (SWE-bench Pro v2, Terminal-Bench 2.1, SWE-bench Verified)
• Runs 125+ tok/s on an M4 Pro MacBook Pro and 150+ tok/s on a 64GB NVIDIA RTX Spark — over 5x faster than comparable dense 25-40B models at similar coding scores
• Fully local/offline: no API fees, data stays on-device

Where it lags

• Smaller 40B-A2.5B footprint means it needs more explicit task decomposition than larger closed-source coding models on complex multi-step work
• No long-context retrieval or multimodal support — text-only

Best for: developers who want a fast, private, locally-run coding agent model rather than an API-hosted frontier model.

What Flint Is

Flint is Recursive Superintelligence's first public model release, built by a small team augmented with the company's "Eureka Machine" automated-research system, which proposed architecture, quantization, and kernel optimizations across the model-building stack. The architecture was chosen specifically to maximize per-token throughput on consumer/prosumer local hardware by minimizing active parameters relative to total capacity (40B total, ~2.5B active).

Unlike most frontier coding models that are served via API, Flint is designed to be the "daily coding model" running entirely on a developer's own machine — trading some raw capability against the largest closed models for speed, privacy, and zero marginal cost. Recursive recommends keeping tasks within roughly 256K tokens of its 512K context window and breaking complex work into smaller, explicitly-scoped subtasks.

Specifications

Pricing

Flint is distributed as open weights on Hugging Face (Apache 2.0) — no API pricing; only local hosting/compute cost applies.

Context Window

Flint has a 512,000-token context window — roughly ~1,000 pages of text in a single request. Recursive recommends staying within 256K tokens for best coding results. This supports large single-file or multi-file local coding sessions without external API calls.

Public Benchmark Scores

Scores above are BF16-checkpoint results reported by Recursive and shown for context. They are not Benchgen measurements. Terminal-Bench 2.1 used the Terminus 2 scaffold with internet access (retrieval-obtained solutions counted unsuccessful); SWE-bench Pro v2 and SWE-bench Verified used mini-swe-agent with no internet access; all scores average pass@1 over 5+ attempts. The NVFP4 quantized checkpoint scores within ~0.7 points of BF16 on these same benchmarks.

Note (bucket-C exclusion): Recursive's launch blog also reports a composite "mean score" chart against comparator models anonymized as "Laguna XS 2.1," "Muse Glimmer," and a "Sparse MoE" / "Dense" class average — these are not named, individually-sourced models and are excluded from leaderboard mapping as non-reproducible, vendor-internal comparisons.

How Flint Performs on Real Agent Tasks

Flint's benchmark profile (SWE-bench Pro/Verified, Terminal-Bench 2.1) signals strong agentic coding and tool-use ability relative to its size class, consistent with Recursive's positioning of it as a local daily-driver for coding agents rather than a general-purpose frontier assistant. Its small active-parameter footprint (2.5B) is the headline trade-off: Recursive itself notes the model may need more explicit task decomposition on complex multi-step work compared to larger closed-source coding models — a real agent-reliability signal, not just a raw-capability one.

Use Flint Locally

# Flint is open-weight; serve locally via vLLM, SGLang, llama.cpp, or Recursive's own "rec" engine.
from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="unused")
response = client.chat.completions.create(
    model="flint",
    messages=[{"role": "user", "content": "Write a Python function that merges overlapping intervals."}],
    max_tokens=32768,
)
print(response.choices[0].message.content)

Related Models & Benchmarks

• Benchmarks: SWE-bench Verified · SWE-bench Pro · TerminalBench · LiveCodeBench v6 · IFBench · IFEval

Frequently Asked Questions

Specs and scores sourced from Recursive's official launch blog and Hugging Face model card; third-party benchmark scores attributed inline. Last updated 2026-10-22.