Recursive Flint
by Recursive (Recursive Superintelligence, Inc.)
Apache 2.0 open-weight coding model (40B MoE, 2.5B active) built to run coding agents locally on a laptop
Recursive Flint is an open-weight coding model from Richard Socher's AI lab Recursive, released October 21, 2026 under Apache 2.0. It is a 40B-parameter mixture-of-experts model with about 2.5B active per token, aimed at engineering teams that want an agentic coding model running on a laptop or their own GPUs, so source code never leaves the machine and there are no per-token API fees.
Recursive Flint is the first model from Recursive, an AI lab founded in 2025 and led by CEO Richard Socher, formerly chief scientist at Salesforce. Recursive came out of stealth in May 2026 with $650 million at a $4.65 billion valuation, with GV, Nvidia and AMD Ventures among its backers, according to TechCrunch. Flint shipped on October 21, 2026. It is a text-only, decoder-only mixture-of-experts Transformer with 39.8B total parameters and about 2.5B active per token: 49 layers (2 dense, 47 MoE), 256 routed experts with 6 selected per token plus a shared expert, and a mix of 35 sliding-window and 14 full-attention layers. The Hugging Face model card documents a 262,144-token context; Recursive's release post says windows up to 512k work but recommends staying within 256k. The model was trained from scratch for coding agents, with reasoning and OpenAI-compatible multi-turn tool calls for investigating codebases, fixing bugs, implementing features and writing tests. Recursive says its automated research system, the Eureka Machine, contributed to the architecture search, a 4-bit output layer with distillation from the 16-bit model, and custom kernels including a Metal megakernel for Apple M4 and M5 chips. Self-reported scores are 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1, with the NVFP4 checkpoint within about a point of BF16. Recursive claims more than 125 tokens per second on an M4 Pro MacBook Pro. Weights are on Hugging Face, quantized builds run in llama.cpp, Ollama, LM Studio, vLLM and SGLang, and OpenRouter lists a free hosted route.
Platform or developer-productivity leads at companies that cannot send source code to a hosted coding assistant, because Flint runs entirely on developer laptops or internal GPUs under a permissive licence.
An agentic coding model with published SWE-bench scores that runs offline on a 24GB-plus laptop, with no per-token fees and no code leaving the device.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Free
- Target Market
- CTOs, Enterprise Developers, Platform Engineering Teams, AI Engineers
- Deployment
- Open-source, Self-hosted, Edge-first
- Founded
- 2025
Key Features
- ✓Sparse 40B MoE with 2.5B active
Only about 2.5B parameters fire per token, which is why the model reaches laptop-class speeds while scoring near dense 25-40B coders.
- ✓Agentic tool use
Multi-turn, OpenAI-compatible tool calls let it drive coding agents that read repositories, run commands, fix bugs and write tests.
- ✓Long context
The model card documents a 262,144-token window shared by prompt and output, enough to load large parts of a codebase in one session.
- ✓Apache 2.0 open weights
Weights on Hugging Face can be used commercially, modified, fine-tuned and redistributed without a separate licence negotiation with Recursive.
- ✓NVFP4 low-bit checkpoint
Recursive reports the 4-bit NVFP4 build scoring within about 0.7 points of BF16, which cuts memory needs for local and GPU deployment.
- ✓Broad runtime support
Runs in llama.cpp, Ollama, LM Studio, vLLM, SGLang and Recursive's own Rec engine, so teams can keep their existing local tooling.
Capabilities
Use Cases
- •Private local coding assistant
Developers run Flint on their own MacBooks through LM Studio or Ollama, so proprietary code stays on the device and no API key is needed.
- •Self-hosted agent backend
A platform team serves Flint with SGLang on one H200 or B300 GPU and points internal coding agents at its OpenAI-compatible endpoint.
- •Fine-tuned in-house coder
An engineering org fine-tunes the Apache 2.0 weights on its own repositories and conventions, then ships the result internally without licence restrictions.
- •Offline field and secure-site work
Engineers in regulated or disconnected environments keep an agentic coding model available with no network connection and no outbound data flow.
Ideal For
Best For
- ✓Offline or air-gapped development environments where code cannot reach a third-party API
- ✓Teams that want a local default coding model on Apple Silicon MacBooks with 24GB or more of memory
- ✓Building internal coding agents on an Apache 2.0 base model that can be fine-tuned and redistributed
- ✓Cutting coding-assistant API spend on routine bug fixes, tests and small features
- ✓Self-hosting a coding model on a single H200 or B300 GPU behind an OpenAI-compatible endpoint
Not Ideal For
- ✗Teams that hand large, loosely specified tasks to an agent: Recursive says Flint needs more explicit guidance than large closed coding models and works best when work is split into small steps
- ✗Buyers who need a vendor SLA, indemnity or enterprise support contract, since Recursive publishes no enterprise offering for Flint
- ✗Developers on machines under 24GB of memory, which LM Studio lists as the minimum for the smallest build
- ✗Workloads that need image or multimodal input, because Flint is text-only
Deployment
Market Analysis
Pros
- ✓Permissive Apache 2.0 licence allows commercial use, fine-tuning and redistribution
- ✓Fast local inference: Recursive measured about 131 tokens per second on an M4 Pro MacBook Pro
- ✓Code stays on the device, with no API fees and no network dependency
- ✓Fits common local tooling (LM Studio, Ollama, llama.cpp) from launch day
Cons
- ✗All benchmark and speed figures are self-reported; RuntimeWire noted they had not been independently verified at launch
- ✗Hacker News commenters said it still trails Qwen 3.8 and questioned whether the benchmark numbers hold up in real-world testing
- ✗Needs more explicit guidance than large closed coding models, so big tasks must be broken into smaller steps
- ✗No Hugging Face inference provider hosts it yet, and Transformers modeling code was not included at release
Pricing
Open weights (Hugging Face)
$0
- ✓Apache 2.0 licence
- ✓BF16 and quantized builds
- ✓Self-host on laptop or GPU
OpenRouter hosted route
$0
- ✓recursive/flint-v1.0:free endpoint
- ✓No local hardware required
- ✓Not available through EU in-region routing
The weights are free under Apache 2.0, and the only hosted access Recursive links to is a free OpenRouter route with unpublished rate limits. Your real cost is hardware: at least 24GB of memory locally, or one H200 or B300 GPU per replica for the official SGLang setup. Recursive publishes no paid tier, SLA or enterprise support plan.
Security & Compliance
Connect
Sources
This page was written from 8 sources, 6 on domains other than recursive.com.
- 1.recursive.com — recursive.comvendor
- 2.recursive.com — introducing recursive flintvendor
- 3.huggingface.co — flint
- 4.lmstudio.ai — recursive flint
- 5.techcrunch.com — recursive launches flint a laptop coding model it says was l
- 6.runtimewire.com — recursive flint local coding model eureka machine
- 7.hn.algolia.com — 50183678
- 8.openrouter.ai — flint v1.0:free
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Arena AI Evaluations
Paid model evaluations built on the Arena leaderboard's crowdsourced human votes, for enterprises choosing an LLM
Kimi K3
2.8T-parameter open-weight MoE model with a 1M-token context, priced at $3/$15 per million tokens
TypeSafe Jev
A 'System One' decision model that returns typed choices, scores and calibrated probabilities instead of text, for agent routing and classification
Xiaomi MiMo-V2.6
MIT-licensed, 1M-context multimodal MoE models that took the top open-weights score on Artificial Analysis at well under $1 per million tokens