# Recursive Flint

> Source: https://aiwiki.ai/wiki/recursive_flint
> Updated: 2026-10-22
> Categories: AI Code Generation, AI Models, Large Language Models, Mixture of Experts, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Recursive Flint." aiwiki.ai, 22 Oct 2026. https://aiwiki.ai/wiki/recursive_flint
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Recursive Flint** is a text-only [large language model](https://aiwiki.ai/wiki/large_language_model) developed by [Recursive Superintelligence](https://aiwiki.ai/wiki/recursive_superintelligence) for coding agents and announced on October 21, 2026. It uses a [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) architecture with 39.8 billion total parameters and approximately 2.5 billion active parameters per token. Its intended tasks include investigating repositories, fixing bugs, implementing features, and writing tests. Recursive released the model weights under Apache 2.0.[1][2][12]

Flint produces text, reasoning, and tool requests. A separate coding harness executes those requests and returns their results, so the model, inference runtime, and coding agent are distinct parts of a working installation. Recursive provides an optimized local engine called `rec`, alongside documentation for other local runtimes and GPU serving.[4][6]

## Architecture and specifications

Flint is a decoder-only [Transformer](https://aiwiki.ai/wiki/transformer). The released configuration combines dense feed-forward layers with sparse expert layers and alternates local and full attention.[2][3]

| Property | Released specification |
| --- | --- |
| Total parameters | 39.8 billion [2] |
| Active parameters per token | Approximately 2.5 billion, including the output head [2] |
| Hidden size | 2,048 [3] |
| Decoder layers | 49 [3] |
| Dense layers | First 2 layers [3] |
| MoE layers | Remaining 47 layers [2][3] |
| Routed experts per MoE layer | 256, with 6 selected per token [3] |
| Shared expert | One always-active shared expert per MoE layer [2] |
| Expert intermediate size | 512; shared expert intermediate size 1,024 [3] |
| Attention layout | 35 sliding-window layers and 14 full-attention layers [2][3] |
| Attention heads | 32 query heads and 4 key/value heads, with head dimension 128 [3] |
| Sliding window | 513 tokens, including the current token [3] |
| Configured context | 262,144 tokens [3] |
| Vocabulary | 150,144 tokens [3] |
| Released checkpoint precision | BF16 [3] |

The model card's active-parameter count includes the output head but excludes the full input embedding matrix. It gives approximately 2.2 billion active parameters for the model body alone and uses `40B-A2.5B` as its rounded size designation. Thus, the active count depends on which components are counted.[2]

Flint applies per-head QK RMSNorm and learned attention sinks. Its [sliding-window attention](https://aiwiki.ai/wiki/sliding_window_attention) layers use [rotary position embeddings](https://aiwiki.ai/wiki/rotary_position_embedding), while the full-attention layers have no positional encoding. The configuration specifies sigmoid expert routing and normalized top-k routing probabilities.[2][3]

## Context and generation behavior

The public checkpoint sets `max_position_embeddings` to 262,144, and the supplied SGLang launch scripts use that same context length. Recursive's hosted configuration also specifies a 262,144-token [context window](https://aiwiki.ai/wiki/context_window), with an output allowance of up to 65,536 tokens. The prompt, conversation history, tool definitions and results, reasoning, and final answer all consume context.[3][4][7]

The launch announcement describes support for up to 512k tokens but recommends remaining within 256k for coding and other tasks. This broader claim should be distinguished from the 262,144-token configuration documented for the released checkpoint and serving scripts. Local capacity also depends on the runtime, quantization, available memory, and cache settings.[3][4][5][7]

Recommended sampling settings are temperature 1.0, top-p 0.95, and top-k 20. Recursive's documentation describes reasoning as always enabled, without a supported reasoning-effort or disable-thinking switch. The model can decide whether to emit a reasoning trace on a particular turn. In SGLang, this trace is returned in `reasoning_content` and counts toward the output budget; the serving guide recommends at least 32,768 output tokens for hard problems and retaining returned reasoning in subsequent conversation history.[7][10]

## Training and automated research

Recursive says Flint was trained from scratch through pretraining, mid-training, supervised fine-tuning, and [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning). Its training covered text and code, instruction following, and agentic tasks.[2]

The company describes Flint as a practical application of its Eureka Machine automated research system. According to the launch account, a small team of people and agents developed the model in under four months. The system searched architecture configurations and contributed to data, optimization, training infrastructure, quantization, and inference kernels. These are Recursive's descriptions of its development process.[5]

For [quantization](https://aiwiki.ai/wiki/quantization), Recursive reports using a four-bit output layer and an offline [knowledge distillation](https://aiwiki.ai/wiki/knowledge_distillation) stage with the 16-bit model as teacher. The company says the resulting NVFP4 model scored within 0.7 points of the 16-bit model across its benchmark suite. This describes the particular release recipe and evaluation, rather than a guarantee for other quantizations.[5]

## Reported evaluations

Recursive reported the following release scores. They are developer measurements of Flint operating through specified agent harnesses.[2][5]

| Evaluation | Reported score | Harness and conditions |
| --- | --- | --- |
| [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) 2.1 | 54.9% | Terminus 2 with XML; internet access permitted, but solutions obtained through retrieval counted as unsuccessful [2][5] |
| [SWE-bench](https://aiwiki.ai/wiki/swe_bench) Pro v2 | 89.7% | mini-swe-agent; no internet access [2][5] |
| SWE-bench Verified | 74.2% | mini-swe-agent; no internet access [2][5] |

The agentic evaluations used temperature 1.0, top-p 0.95, and top-k 20, with timeouts of up to five hours. Scores are average pass@1 over at least five attempts, rather than a best-of-five success rate. Recursive notes that comparison results may use different evaluation setups, including different treatment of web retrieval. The figures therefore do not establish an independent ranking across models tested under other conditions.[2][5]

## Availability and inference

The public `RecursiveSI/flint` repository on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) contains the BF16 checkpoint, configuration, tokenizer, chat template, and an SGLang serving kit. The BF16 weights occupy about 79.6 GB. Quantized local artifacts have different memory requirements and must be paired with a compatible runtime.[2][8][9]

| Route | Documented requirements or behavior |
| --- | --- |
| `rec` | Optimized Mac engine for M4 Pro, M5 Pro, or M5 Max; at least 48 GB unified memory and about 30 GB free disk space; designed for a single session [6] |
| [SGLang](https://aiwiki.ai/wiki/sglang) serving kit | Validated on one B300 or H200 per replica; at least 141 GB GPU memory; 80 GB GPUs are unsupported by this kit [7] |
| [llama.cpp](https://aiwiki.ai/wiki/llama_cpp) | Requires a compatible GGUF artifact and runtime revision; context and offload settings depend on the device [9] |
| [Ollama](https://aiwiki.ai/wiki/ollama), LM Studio, and Unsloth | Separate documented local workflows with runtime-specific artifacts and requirements [9] |
| Transformers | Requires a Flint-compatible implementation; the documentation identifies an integration fork, with its Flint pull request still open [8] |
| Hosted access | Documentation lists OpenCode and OpenRouter; direct Recursive API access is limited to select partners [11] |

The `rec` engine can launch Pi or OpenCode and configure their connection automatically. It also exposes an OpenAI-compatible chat-completions API for other clients. Recursive reports 1.4-2.6 times the token throughput of llama.cpp in its M4 Pro comparison, using 48 GB memory, four-bit weights, prompts from 1K to 252K tokens, and 1,024 generated tokens. The test used five runs per context and excluded prefill, so it measures decoding throughput under those conditions.[6]

The launch blog separately reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64 GB NVIDIA RTX Spark at short contexts. These are developer-reported speed measurements for an optimized inference stack.[5]

The SGLang kit includes a model package, attention-kernel support, serve scripts, and Flint-specific tool and reasoning parsers. Its validated deployment uses independent single-GPU replicas; tensor, expert, and pipeline parallelism are listed as unvalidated. Direct loading with an unmodified Transformers installation is not the public workflow documented for this checkpoint, despite generic loading examples displayed by the Hub interface.[7][8]

## Limitations

Recursive warns that Flint can generate incorrect code, harmful content, and inappropriate tool calls. Safety-oriented fine-tuning and structured tool-call grammar do not ensure that a requested action is correct. The coding harness remains responsible for executing tools within its permissions.[2][4]

A local model's download size does not determine whether it can run at a chosen context length. Memory is also needed for the KV cache, compute buffers, the operating system, and other applications. Larger contexts and concurrent requests increase memory use, and CPU offload can change performance. Recursive recommends selecting the artifact, runtime revision, and configuration together when assessing a deployment.[9]

## References

1. Richard Socher. [Announcement of Recursive Flint](https://x.com/RichardSocher/status/2112944385415426249). X, October 21, 2026.
2. Recursive. [RecursiveSI/flint model card](https://huggingface.co/RecursiveSI/flint). Hugging Face. Accessed October 23, 2026.
3. Recursive. [Flint model configuration](https://huggingface.co/RecursiveSI/flint/blob/2a247be0c0aec895ccb0ac7ca92229ec6df37288/config.json). Hugging Face, initial release.
4. Recursive. [Using Flint locally](https://docs.recursive.com/models/flint). Accessed October 23, 2026.
5. Recursive. [Introducing Flint](https://recursive.com/updates/introducing-recursive-flint). October 2026.
6. Recursive. [rec: local inference engine](https://github.com/recursive-oss/rec). GitHub. Accessed October 23, 2026.
7. Recursive. [Serving Flint with SGLang](https://huggingface.co/RecursiveSI/flint/blob/2a247be0c0aec895ccb0ac7ca92229ec6df37288/sglang/README.md). Hugging Face, initial release.
8. Recursive. [Hugging Face Transformers](https://docs.recursive.com/gpu/transformers). Accessed October 23, 2026.
9. Recursive. [Choose your setup](https://docs.recursive.com/local/choose-your-setup). Accessed October 23, 2026.
10. Recursive. [Generation settings](https://docs.recursive.com/guides/generation-settings). Accessed October 23, 2026.
11. Recursive. [Release resources](https://docs.recursive.com/resources/releases). Accessed October 23, 2026.
12. Recursive. [Flint Apache 2.0 license](https://huggingface.co/RecursiveSI/flint/blob/2a247be0c0aec895ccb0ac7ca92229ec6df37288/LICENSE). Hugging Face, initial release.