> ## Documentation Index > Fetch the complete documentation index at: https://docs.tokios.com/llms.txt > Use this file to discover all available pages before exploring further. > ## Agent Instructions > Tokios exposes private local models through https://api.tokios.com using an outbound-only connector tunnel. > Use 'connector' for the local tokios-connector process. > Use 'deployment' for the public model name in the API model field. > Use 'upstream model id' for the model served by Ollama, llama.cpp, vLLM, LM Studio, or another local backend. > API keys use the prefix sk-tok-. Connector tokens use the prefix ct-tok-. # RecursiveSI Flint: run a 40B coding agent on a laptop > Flint is a 40B MoE model with 2.5B active parameters — runs on a 24 GB GPU at Q4 while competing with much larger models on coding benchmarks. Apache 2.0 licensed. Flint is a 39.8B-parameter mixture-of-experts model from Recursive Superintelligence, Inc. (RecursiveSI), founded by Richard Socher. Only **2.5B parameters activate per token**, which means it runs on a single 24 GB consumer GPU at Q4 while delivering coding-agent performance that rivals much larger models. > [!NOTE] > > Last verified: 2026-10-24 ## Why Flint matters Flint is purpose-built for **agentic coding** — investigating codebases, fixing bugs, implementing features, writing tests. It supports multi-turn tool use and structured tool-call grammar. The 2.5B active parameter count means it fits hardware that cannot run any other model in its performance class. Notably, Flint was **built largely by its own AI research system** — an early demonstration of self-improving AI (autoresearch). ## Specs | Property | Value | | - | - | | HuggingFace | [`RecursiveSI/flint`](https://huggingface.co/RecursiveSI/flint) | | GGUF | [`unsloth/flint-GGUF`](https://huggingface.co/unsloth/flint-GGUF) | | Ollama | `ollama run flint` | | Parameters | 39.8B total / \~2.5B active (256 routed experts, 6/token + 1 shared) | | Architecture | Decoder-only MoE, 49 layers (2 dense + 47 MoE) | | Context | 262,144 tokens (262K) | | Vocabulary | 150,144 tokens | | License | Apache 2.0 | | Knowledge cutoff | June 2026 | ### Attention architecture Flint uses a hybrid attention pattern: * **35 sliding-window layers** (window size 513 tokens) for efficient local context * **14 full-attention layers** with RoPE for long-range dependencies This hybrid design keeps the KV cache small while maintaining 262K context capability. ## Hardware fit All 39.8B parameters must stay resident in memory — only the 2.5B active count reduces compute per token, not memory. The GGUF sizes from Unsloth: | Quantization | File size | VRAM to run | Hardware | | - | - | - | - | | BF16 | \~80 GB | \~82-85 GB | 2x RTX 5090 or 1x RTX PRO 6000 | | Q8\_0 | 42.3 GB | \~44-46 GB | RTX PRO 5000 (48 GB) or DGX Spark | | Q5\_K\_XL | 29.9 GB | \~31-33 GB | RTX 5090 (32 GB, tight) | | **Q4\_K\_XL** | **25.1 GB** | **\~26-28 GB** | **RTX 5090 (32 GB, comfortable)** | | Q3\_K\_XL | 18.5 GB | \~20-21 GB | RTX 4090/3090 (24 GB) | | Q2\_K\_XL | 13.6 GB | \~15-16 GB | RTX 4070/5070 (12 GB, tight) | | IQ1\_M (1-bit) | 11.4 GB | \~12-13 GB | Laptops with 16 GB RAM | > [!TIP] > > **Sweet spot: Q4\_K\_XL on RTX 5090.** At 25.1 GB, it fits comfortably in 32 GB with room for KV cache at moderate context lengths. Q3\_K\_XL on a used RTX 3090 (24 GB, \~\$800) is the best budget path. ## Benchmarks | Benchmark | Score | Notes | | - | - | - | | Terminal-Bench 2.1 | 54.9% | Agentic coding | | SWE-bench Pro v2 | 89.7% | Software engineering | | SWE-bench Verified | 74.2% | Real-world GitHub issues | Best-in-class for MoE models under \~5B active parameters. ## Run it **Ollama (simplest)** ```bash theme={"theme":{"light":"github-light","dark":"css-variables"}} ollama run flint ``` Check the [Ollama library](https://ollama.com/library/flint) for current tags. **SGLang (recommended for serving)** ```bash theme={"theme":{"light":"github-light","dark":"css-variables"}} python -m sglang.launch_server \ --model-path "RecursiveSI/flint" \ --port 30000 ``` SGLang is the recommended backend for production serving. RecursiveSI publishes an official serving kit with Docker. **vLLM** ```bash theme={"theme":{"light":"github-light","dark":"css-variables"}} vllm serve "RecursiveSI/flint" ``` vLLM supports Flint through the `FlintForCausalLM` architecture (PR merged). **llama.cpp / LM Studio** Download GGUF builds from [`unsloth/flint-GGUF`](https://huggingface.co/unsloth/flint-GGUF). Load in LM Studio or run through `llama-server`. ### Recommended generation settings * Temperature: `1.0` * Top-p: `0.95` * Top-k: `20` * Min output tokens: `32,768` ## Connect it with Tokios Once Flint is running on your hardware through any of the backends above, pair the [Tokios connector](https://docs.tokios.com/quickstart.md) to reach it as `https://api.tokios.com` from Claude Code, Cline, or any compatible client. **1. Pair the Tokios connector** Sign in at [tokios.com/console](https://tokios.com/console), open the **Setup** tab, and click **Start pairing** to get a one-time claim code (`TKS-XXXX-XXXX`). Run the installer on the machine hosting Flint: ```bash Linux theme={"theme":{"light":"github-light","dark":"css-variables"}} curl -fsSL https://tokios.com/downloads/install/linux | sh -s -- --code TKS-XXXX-XXXX ``` Then **approve** the device on the **Setup** or **Connectors** tab. **2. Register a deployment** In the [Models](https://tokios.com/console/models) tab, register `flint` as the upstream model id and choose a public deployment name. **3. Create an API key** In the [Keys](https://tokios.com/console/keys) tab, click **Create key**. Copy the `sk-tok-...` key — it is shown only once. ## Caveats * **All 39.8B parameters stay resident.** The 2.5B active count reduces compute per token, not memory. Plan for 25+ GB at Q4. * **Knowledge cutoff is June 2026.** Flint does not know about events after that date. * **Apache 2.0 license is permissive.** Commercial use is allowed with no restrictions. * **NVFP4 variant available.** NVIDIA published an FP4 quantization for Blackwell GPUs. * **Built by AI.** Flint was largely created by RecursiveSI's own AI research system. This is an early demonstration of autoresearch. > **[Model fit](https://docs.tokios.com/guides/model-fit.md)** > > Calculate which quantization fits your GPU. > **[Choose a model by task](https://docs.tokios.com/guides/model-by-task.md)** > > Compare Flint against other coding models. > **[RTX 50 series guide](https://docs.tokios.com/guides/rtx-50-series-consumer.md)** > > Best consumer GPUs for running Flint. > **[Quickstart](https://docs.tokios.com/quickstart.md)** > > Connect Flint to Tokios for remote access. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.