# Flint ![Organization](https://img.shields.io/badge/Org-Recursive-555) ![Context](https://img.shields.io/badge/Context-512K%20tokens-1f6feb) ![Pricing](https://img.shields.io/badge/Price-Open%20weights-2da44e) ![License](https://img.shields.io/badge/License-Apache%202.0-d29922) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) ![Released](https://img.shields.io/badge/Released-Oct%202026-555) > **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 | Field | Value | |---|---| | Organization | Recursive (Recursive Superintelligence, Inc.) | | Parameters | 39.8B total, ~2.5B active per token (MoE) | | Context window | 512,000 tokens (recommended ≤256K) | | Architecture | Decoder-only MoE, 49 layers (2 dense + 47 MoE), 256 routed experts (top-6 + 1 shared) | | License | Apache 2.0 | | Release date | 2026-10-21 | | Knowledge cutoff | 2026-06 | | Modality | Text | ## Pricing | | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | Open weights | Free (self-hosted) | Free (self-hosted) | 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 | Benchmark | Score | Source | Date | |---|---|---|---| | [SWE-bench Verified](/benchmarks/princeton-nlp-openai/swe-bench-verified) | 74.2% | [Recursive launch blog](https://www.recursive.com/updates/introducing-recursive-flint) | 2026-10 | | [SWE-bench Pro](/benchmarks/swe-bench-team/swe-bench-pro) (v2) | 89.7% | [Recursive launch blog](https://www.recursive.com/updates/introducing-recursive-flint) | 2026-10 | | [TerminalBench](/benchmarks/tbench/terminalbench-21) (2.1) | 54.9% | [Recursive launch blog](https://www.recursive.com/updates/introducing-recursive-flint) | 2026-10 | | [LiveCodeBench v6](/benchmarks/livecodebench-v6) | 80.3% | [Recursive launch blog](https://www.recursive.com/updates/introducing-recursive-flint) | 2026-10 | | [IFBench](/benchmarks/ifbench) | 89.2% | [Recursive launch blog](https://www.recursive.com/updates/introducing-recursive-flint) | 2026-10 | | [IFEval](/benchmarks/ifeval) | 95.1% | [Recursive launch blog](https://www.recursive.com/updates/introducing-recursive-flint) | 2026-10 | *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 ```python # 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](/benchmarks/princeton-nlp-openai/swe-bench-verified) · [SWE-bench Pro](/benchmarks/swe-bench-team/swe-bench-pro) · [TerminalBench](/benchmarks/tbench/terminalbench-21) · [LiveCodeBench v6](/benchmarks/livecodebench-v6) · [IFBench](/benchmarks/ifbench) · [IFEval](/benchmarks/ifeval) ## Frequently Asked Questions

What is Flint?

Flint is Recursive's open-weight 40B-A2.5B mixture-of-experts coding model, built for fast local/offline agentic coding on consumer hardware.

What is Flint's context window?

Flint supports up to 512,000 tokens (≈1,000 pages), though Recursive recommends staying within 256K tokens for best results.

How much does Flint cost?

Flint is open weights — free to download and run locally; only your own hosting/compute cost applies.

Is Flint open source?

Yes, Flint is released under the Apache 2.0 license.

What is Flint's knowledge cutoff?

Recursive states a knowledge cutoff of approximately June 2026.

--- *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.*

Flint — Evaluations

Evaluations