Primary use stated or inferred cautiously from official documentation.
R
RECURSIVE / RecursiveSI/flint
Recursive Flint
Recursive model released on October 21, 2026 to investigate repositories, fix bugs, develop features and write tests with local agents. It is a mixture-of-experts Transformer with 39.8 billion total parameters and approximately 2.5 billion active per token, including the output head. It accepts and generates text; it does not process images or audio. This profile groups the BF16 checkpoint and official INT4 and NVFP4 distributions without counting them as independent models.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | RecursiveSI/flint |
|---|---|
| Model type | Code · Text and reasoning |
| Access model | API and open weights |
| License | Apache 2.0 |
| Deployment | API or self-hosted deployment |
| Release | 21 OCT 2026 |
| Knowledge cutoff | June 2026 |
| Entrance | Text |
| Exit | Text |
| Context window | 262,144 tokens shared between input and output |
| maximum output | Configurable output; OpenRouter lists up to 235,929 tokens, including reasoning |
| Reasoning | Multi-turn reasoning and tool use |
| Published tools | Features · Code agents |
| Structured Outings | OpenRouter: JSON without JSON Schema enforcement |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · OpenRouter · SGLang · vLLM · llama.cpp · Rec |
Assisted coding and tool-using agents working on private repositories when code needs to remain on your own infrastructure. The BF16 kit uses SGLang on an H200 or B300 GPU; quantized exports have different requirements. For complex tasks, break down the work, retain the reasoning history and verify changes with tests before applying them.
The announcement mentions up to 512k tokens, but the published configuration and serving examples specify 262,144 tokens across prompt and generation; we use this verifiable limit, not a guaranteed extension. The recommended minimum of 32,768 tokens for difficult problems is a budget, not the maximum output. Transformers modeling code is not yet included in the checkpoint. Tools require permissions and isolation: the call format does not guarantee safe actions.
03 / CAPABILITIES
What it can do
Orientation
Assisted coding and tool-using agents working on private repositories when code needs to remain on your own infrastructure. The BF16 kit uses SGLang on an H200 or B300 GPU; quantized exports have different requirements. For complex tasks, break down the work, retain the reasoning history and verify changes with tests before applying them.
Context
262,144 tokens shared between input and output · Configurable output; OpenRouter lists up to 235,929 tokens, including reasoning
Tools and integration
Functions · Coding agents
Access
Hugging Face · OpenRouter: recursive/flint-v1.0:free
Input modalities
Text
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | USD 0 / 1M tokens | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | USD 0 / 1M tokens | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
OpenRouter lists recursive/flint-v1.0:free at USD 0 for input and output, subject to usage limits; we do not guarantee that it will remain free or that EU in-region routing is available. Self-hosting does incur costs. The files occupy approximately 80 GB for BF16, 22.6 GB for INT4 MLX, 26.0 GB for INT4 Q4_1 GGUF and 22.9 GB for NVFP4 GGUF. These are not total memory requirements: add the KV cache and runtime; the GGUF guides recommend at least 48 GB.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. These are not Inferama tests. It uses temperature 1.0, top-p 0.95 and top-k 20; agent results are mean pass@1 over at least five attempts, with up to five hours per task. Terminal-Bench uses Terminus 2 and counts solutions obtained through search as unsuccessful; SWE-bench uses mini-swe-agent without Internet. Precisions and runtimes must be recorded separately; we do not assume equivalence between BF16, NVFP4 and INT4.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| Terminal-Bench 2.1 | 54.9% | Higher is better | View source ↗ |
| SWE-bench Pro v2 | 89.7% | Higher is better | View source ↗ |
| SWE-bench Verified | 74.2% | Higher is better | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
RecursiveSI/flint · BF16 / INT4 / NVFP4
Release confirmed by Recursive, NVIDIA and vLLM's day-zero integration. One family profile; exports do not duplicate the catalog.
07 / SOURCES
Traceability
TECHNICAL SHEETRecursive Flint · ficha fijadahuggingface.co↗
LAUNCHAnuncio oficial de Recursiverecursive.com↗
CONFIGURATIONConfiguración BF16 y contextohuggingface.co↗
IMPLEMENTATIONSoporte de lanzamiento y metodología vLLMvllm-project.github.io↗
API AND PRICINGOpenRouter · ruta gratuita, límites y condicionesopenrouter.ai↗
TECHNICAL SHEETFlint NVFP4 GGUFhuggingface.co↗
TECHNICAL SHEETFlint INT4 MLXhuggingface.co↗
TECHNICAL SHEETFlint INT4 Q4_1 GGUFhuggingface.co↗