Apache-2.0llmRuns locally39.8B
Flint: Specs, License & How to Run
Recursive Apache-2.0 39.8B/2.5B local coding-agent MoE with GGUF, INT4, and Ollama tags.
- License
- Apache-2.0
- Parameters
- 39.8B (active ~2.5B)
- HF likes
- 533 (2026-10-23)
- First seen
- 2026-10-21
TL;DR. Flint is a llm model from Recursive. Parameters: 39.8B (active ~2.5B).License: Apache-2.0. It is listed under ChatGPT, Closed LLM APIs, Claude Code, Cursor.
| Developer | Recursive |
|---|---|
| Type | llm |
| Architecture | Decoder-only MoE Transformer (card model_type flint). 49 layers (2 dense + 47 MoE); 256 routed experts, 6 selected per token, plus a shared expert; 35 sliding-window + 14 full-attention layers. Hugging Face safetensors.parameters sum to 39,785,287,456. |
| Parameters | 39.8B · active ~2.5B |
| Modalities | text |
| Languages | en |
| Context | 262,144 tokens (card); blog also names 512k with a 256k recommendation |
| License | Apache-2.0 — Apache License 2.0 on the Hugging Face LICENSE file and card (license apache-2.0) as of 2026-10-23. Sibling INT4, INT4-GGUF, and NVFP4-GGUF cards also tag apache-2.0. Hosted API terms are separate from the weight license. Not legal advice. |
| Hardware | 1× B300 or H200 (BF16 SGLang kit, vendor); 48 GB unified memory (INT4 GGUF, vendor) — BF16 card says “about 80 GB” download; HF usedStorage is 79,587,116,933 bytes (~74.1 GiB). INT4 MLX card says ~22.6 GB. INT4 GGUF card says a 26.0 GB Q4_1 file and “at least 48 GB of unified memory” on Apple Silicon. Peak VRAM is not independently measured. Transformers modeling code is “not included yet” on the BF16 card. (not independently tested) |
| Hugging Face likes | 533 (as of 2026-10-23) |
| Model card | https://huggingface.co/RecursiveSI/flint |
| First seen | 2026-10-21 |
Flint (card name Recursive Flint) is a Mixture-of-Experts coding-agent model from Recursive. Official hosts we opened on 2026-10-23: the ungated card RecursiveSI/flint (created 2026-10-21T14:20:58Z; 533 likes / 420 downloads), siblings flint-int4 (11 likes / 577 downloads), flint-int4-gguf (17 likes / 2,477 downloads), and flint-nvfp4-gguf (9 likes / 1,561 downloads), the Apache 2.0 LICENSE, the blog Introducing Flint, and ollama.com/library/flint.
This page hangs off ChatGPT alternatives, closed LLM APIs, Claude Code alternatives, and Cursor alternatives. That mapping is editorial. We have no sourced quality comparison against ChatGPT, Claude Code, or Cursor. Head-to-head facts versus the MIT 122B coding/agent card already on this site: Flint vs Ornith-1.5-122B. There is no Qwen3.5-35B-A3B page here. Same-size Apache-2.0 diffusion MoE added 2026-10-31: Clover-35B-A3B-Preview vs Flint.
The card calls Flint a model “for coding agents” that investigates codebases, fixes bugs, implements features, and writes tests. Those sentences stay vendor. Hugging Face pipeline_tag is text-generation; model_type is flint.
What we verified (2026-10-23)
| Field | Value | Source |
|---|---|---|
| Host (BF16) | huggingface.co/RecursiveSI/flint (not gated) | HF API |
| Created / last modified | 2026-10-21T14:20:58Z / 2026-10-21T14:24:03Z | same |
| Likes / downloads | 533 / 420 | same |
| License tag + file | apache-2.0 / Apache License 2.0 | HF cardData + LICENSE |
| Pipeline tag | text-generation | HF API |
| Architecture | flint (decoder-only MoE) | card + HF config |
| Params (card) | 39.8B total / ~2.5B active | card |
| safetensors.parameters | 39,785,287,456 | HF API |
| usedStorage (BF16) | 79,587,116,933 bytes (~74.1 GiB) | same |
| Documented context | 262,144 tokens | card |
| Knowledge cutoff | June 2026 | card |
| INT4 MLX | 11 likes; ~22.6 GB (card) | HF API + card |
| INT4 GGUF | 17 likes / 2,477 downloads; 26.0 GB Q4_1 | HF API + card |
| NVFP4 GGUF | 9 likes / 1,561 downloads | HF API |
| Ollama library | flint / flint:40b / flint:40b-a2b-qat-int4 (~23 GB label) | ollama.com/library/flint |
The rounded marketing name is 40B-A2.5B. The card’s footnote says the active-parameter convention includes the output head and excludes the full input embedding; the active body alone is ~2.2B. We keep both quotes.
How to run (from official cards, not tested)
BF16 / SGLang kit (card; one B300 or H200 per replica). Transformers modeling code is “not included yet.”
hf download RecursiveSI/flint --local-dir flint && cd flint # about 80 GB (card)
docker build -t flint-sglang sglang
docker run -d --name flint --gpus '"device=0"' --ipc=host -p 127.0.0.1:30000:30000 \
-v "$PWD:/model" -e MODEL=/model flint-sglang bash /model/sglang/serve_b300.shH200 recipe on the card uses serve_h200.sh. Recommended sampling (copied): temperature 1.0, top-p 0.95, top-k 20; output budget for hard problems at least 32,768 tokens including reasoning.
INT4 GGUF / llama-server (Apple Silicon, card): llama-server against flint-int4-q4_1.gguf with -ngl all -c 262144. The card says leave --swa-full unset and keep --cache-ram 0.
Ollama (library page): ollama run flint or ollama run flint:40b-a2b-qat-int4. The library also names ollama launch claude --model flint and OpenCode / Hermes / OpenClaw launchers. Those labels are copied; forkalt has not run them.
The Recursive blog also names llama.cpp, Unsloth, LM Studio, vLLM, SGLang, and rec (Recursive’s engine for M4 Pro / M5 Pro/Max). Tokens/s figures on that blog (125 tok/s on an M4 Pro; 150 tok/s on a 64 GB RTX Spark) stay vendor.
License
LICENSE is Apache License 2.0. Card tag apache-2.0. Same tag on the INT4, INT4-GGUF, and NVFP4-GGUF siblings. The card says hosted API terms are separate from the weight license. Not legal advice.
What we did not invent
- VRAM, tokens/s, and “fastest / strongest in its class” stay vendor.
- Card bench rows stay self-reported.
- We do not rank Flint against Cursor, Claude Code, ChatGPT, or Ornith-1.5-122B.
Related coding/agent weights: Ornith-1.5-122B — Flint vs Ornith-1.5-122B — Clover-35B-A3B-Preview (diffusion 35B-A3B, not a coding card) — Clover-35B-A3B-Preview vs Flint — Sage-A1.5 (science/agent VLM), Reflection Beam, Ling-3.1-flash. Serving: vLLM. Agents: Cline, OpenCursor.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| Terminal-Bench 2.1 (card) | 54.9% (self-reported) | https://huggingface.co/RecursiveSI/flint |
| SWE-bench Pro v2 (card) | 89.7% (self-reported) | https://huggingface.co/RecursiveSI/flint |
| SWE-bench Verified (card) | 74.2% (self-reported) | https://huggingface.co/RecursiveSI/flint |
Deploy options
- local: Hugging Face weights (BF16)
- local: Hugging Face INT4 (MLX)
- local: Hugging Face INT4 GGUF
- local: Hugging Face NVFP4 GGUF
- local: Ollama (flint)
Comparisons
FAQ
Is Flint open weight and Apache-2.0?
The ungated card RecursiveSI/flint (created 2026-10-21T14:20:58Z; 533 likes / 420 downloads in the 2026-10-23 API check) tags license apache-2.0. The LICENSE file is Apache License 2.0. Sibling cards flint-int4 (11 likes), flint-int4-gguf (17 likes / 2,477 downloads), and flint-nvfp4-gguf (9 likes / 1,561 downloads) also tag apache-2.0. Not legal advice.
Can I run it locally on a laptop?
The BF16 checkpoint’s official path is one B300 or H200 with the bundled SGLang kit. For Apple Silicon the INT4 GGUF card asks for at least 48 GB unified memory and llama-server with Metal. Ollama lists flint / flint:40b-a2b-qat-int4 (~23 GB) on ollama.com/library/flint. Those are vendor size labels. forkalt has not loaded any Flint checkpoint.
Does it replace Cursor or Claude Code?
No sourced bake-off on this site. The card and Recursive blog position Flint as a local coding-agent model. Ollama’s library page names launchers for Claude Code, OpenCode, Hermes, and OpenClaw. Those are product-class mappings, not a quality rank. See Flint vs Ornith-1.5-122B.
How big is it, and is it based on Qwen?
Card: 39.8B total / ~2.5B active (output head included; body alone ~2.2B). HF safetensors sum is 39,785,287,456. Architecture is a from-scratch Flint MoE (model_type flint), not a Qwen3.5 fine-tune. Qwen3.5-35B-A3B is a different family and has no page on this site.
Are the SWE-bench numbers independently tested?
No. Card rows (Terminal-Bench 2.1 54.9%, SWE-bench Pro v2 89.7%, SWE-bench Verified 74.2%) stay self-reported. Recursive marks some comparison scores with a dagger as its own measurements.
Sources & verification
- https://huggingface.co/RecursiveSI/flint
- https://huggingface.co/RecursiveSI/flint/blob/main/LICENSE
- https://huggingface.co/RecursiveSI/flint-int4
- https://huggingface.co/RecursiveSI/flint-int4-gguf
- https://huggingface.co/RecursiveSI/flint-nvfp4-gguf
- https://recursive.com/updates/introducing-recursive-flint
- https://ollama.com/library/flint
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.