flint MoE Language Model for Coding Agents: 16GB+ VRAM, GGUF Builds

At a Glance
| Item | Value |
|---|---|
| Repository | RecursiveSI/flint |
| Family guide | flint guide (2 articles) |
| Published | 2026-10-21 |
| License | apache-2.0 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Recursive has released “Flint", a Mixture-of-Experts (MoE) language model designed for coding agents. It has a total parameter count of 39.8 billion (39.8B) with approximately 2.5 billion (2.5B) active parameters per token.
This model has been developed with the objective of conducting codebase investigations, bug fixing, implementing new features, and writing test code. It supports text input and output, executes reasoning, and supports multi-turn tool usage.
Specifications
- Total parameters: 39.8 billion (39.8B)
- Active parameters: Approximately 2.5 billion per token (approx. 2.5B, including the output head; approximately 2.2B for the active body alone)
- Architecture: Decoder-only mixture-of-experts Transformer
- Number of layers: 49 layers (2 dense layers, 47 MoE layers)
- Attention: 35 sliding window layers, 14 full attention layers
- Sliding window: 513 tokens (current token and 512 preceding tokens)
- Attention heads: 32 query heads, 4 key/value heads (head dimension 128)
- Attention normalization: Per-head QK RMSNorm, and learned per-head attention sinks
- Positional encoding: RoPE adopted for sliding window layers, no positional encoding for full attention layers
- Expert configuration: 256 routed experts per MoE layer (6 selected per token), and 1 constantly active shared expert
- Context length: 262,144 tokens (sum of prompt and generated output)
- Vocabulary size: 150,144 tokens
Performance
According to evaluation measurement results published by Recursive, the following scores are reported across various benchmarks concerning coding and agent tasks:
- Terminal-Bench 2.1: 54.9%
- SWE-bench Pro v2: 89.7%
- SWE-bench Verified: 74.2%
In these evaluation measurements, Flint uses sampling settings of temperature 1.0, top-p 0.95, and top-k 20. Additionally, Terminus 2 is used for executing Terminal-Bench, and mini-swe-agent is used for SWE-bench, with agent scores calculated as the average pass@1 across at least 5 trials.
Terminal-Bench is a metric that measures whether commands can be actually executed in a terminal to carry out given tasks to completion, where higher Pass@1 (%) figures indicate superior performance. Furthermore, SWE-bench is a metric that measures whether repositories can be edited and fixed targeting actual GitHub issues, where higher figures in the percentage of resolved issues (%) indicate superior performance.
Although evaluation environments and settings vary by source, an adjustment is applied to Flint’s Terminal-Bench score to treat solutions obtained through web searches as failing. These measurement results conducted by the publishers themselves indicate that the model possesses high response capabilities and task-solving abilities in code correction tasks for actual development repositories and agent-like task execution accompanied by terminal operations.
Strengths and Use Cases
Flint is designed with a primary focus on utilization as a coding agent in software development. Specific intended use cases include investigating existing codebases, identifying and fixing bugs, implementing new features, and creating test code starting with unit tests.
This model goes beyond mere text generation, possessing the capability to handle multi-turn tool usage while executing complex thinking processes (reasoning). In the model’s training process, in addition to pre-training from scratch, mid-training, instruction tuning (SFT), and reinforcement learning (RL) have been conducted, with adjustments specialized not only for text and code generation but also for instruction-following performance and agent-like behavior.
It also supports OpenAI-compatible tool calling features. When using chat templates, it is recommended to retain reasoning_content returned by the model within the conversation history. When tackling high-difficulty problems, setting up an output budget of at least 32,768 tokens including the reasoning process is suggested as a recommended configuration.
As a precaution for use, since code and tool calls generated by the model are not guaranteed to be always accurate and safe, conducting tests and code reviews prior to application in actual environments is required. Even if structured tool call grammars are used, they cannot prevent inappropriate actions, so clear execution privileges and execution boundaries must be set on the tool-using application side. Furthermore, attention must be paid to the point that longer inference times increase latency and may exhaust the output budget before responses are completed.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 39.8B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 16GB (RTX 5060 Ti 16GB / 4060 Ti 16GB, etc.) | Q2_K_XL | 12.6GB | 15.2GB |
| 24GB (RTX 4090 / 3090, etc.) | Q3_K_XL | 17.2GB | 20.7GB |
| 32GB (RTX 5090, etc.) | Q4_K_XL | 23.3GB | 28.0GB |
| 48GB (RTX 6000 Ada / A6000, etc.) | Q8_0 | 39.4GB | 47.3GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build unsloth/flint-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp via a converted build.
The publisher ships safetensors, but bartowski/RecursiveSI_flint-GGUF provides a GGUF build you can use.
License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.
Compression: the Q8_0 build measures 8.51 bits per weight — about 53% the size of the original 16-bit weights, calculated by this site from the actual file sizes.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Recent Models in the Same Size Class
Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| LiquidAI/LFM2.5-24B-A2B-Base | 24.0B | 80GB | lfm1.0 | LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-FP8 | 24.0B | 32GB | lfm1.0 | LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-MLX-bf16 | 24.0B | 80GB | lfm1.0 | LFM2.5-24B-A2B-MLX-bf16 Reasoning Model for Apple Silicon: 80GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-MLX-8bit | 24.0B | 32GB | lfm1.0 | LFM2.5-24B-A2B-MLX-8bit Text Generation Model: 32GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B | 24.0B | 12GB | lfm1.0 | LFM2.5-24B-A2B Agent-Optimized MoE Reasoning Model: 12GB+ VRAM (2026-11-17) |
How to Get It
BF16 checkpoints for Flint are published in safetensors format on the Hugging Face repository (RecursiveSI/flint). The Apache License 2.0 applies, and agreeing to terms of use (gating) is not required to obtain the model. Note that quantized variants for NVIDIA and Apple Silicon are described as being separately prepared.
At this time, Hugging Face Transformers modeling code is not directly included, and the execution environment takes the form of utilizing the SGLang serving kit provided within the repository (sglang/ directory). An environment using 1 B300 or H200 GPU per replica is assumed.
Model acquisition and the construction of the SGLang serving environment are performed by executing the following commands:
pip install -U huggingface_hub openai
hf download RecursiveSI/flint --local-dir flint && cd flint
After acquisition, build the SGLang image using Docker and start the container (use serve_b300.sh for B300 GPUs, or serve_h200.sh for H200 GPUs).
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.sh
After confirming server startup, you can access it by specifying the http://localhost:30000/v1 endpoint from the included sample client or OpenAI-compatible SDKs.
python3 sglang/example_client.py --base-url http://localhost:30000/v1
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-10-24 | bartowski | GGUF (imatrix) | bartowski/RecursiveSI_flint-GGUF | IQ2_S 15.0GB (fits in 16GB VRAM) |
| 2026-11-17 | unsloth | GGUF (imatrix) | unsloth/flint-GGUF | Q2_K_XL 15.2GB (fits in 16GB VRAM) |
| 2026-11-17 | lmstudio-community | GGUF | lmstudio-community/flint-GGUF | Q4_K_M 27.0GB (fits in 32GB VRAM) |
| 2026-11-17 | RecursiveSI | MLX | RecursiveSI/flint-int4 | – |
| 2026-11-17 | mlx-community | MLX | mlx-community/flint-OptiQ-4bit | MLX 4bit 30.7GB (fits in 32GB VRAM) |
File sizes of each build:
- Available builds in bartowski/RecursiveSI_flint-GGUF: IQ2_XXS 11.2GB / IQ2_XS 11.9GB / IQ2_S 12.5GB / IQ2_M 13.9GB / Q2_K 14.4GB / IQ3_XXS 16.3GB / IQ3_XS 17.1GB / Q3_K_S 17.2GB / Q3_K_M 18.2GB / Q3_K_L 19.2GB / IQ3_M 19.6GB / IQ4_XS 20.9GB / Q4_0 21.3GB / Q4_K_S 22.2GB / Q4_1 23.5GB / IQ4_NL 23.6GB / Q4_K_M 23.6GB / Q4_K_L 25.4GB / Q5_K_S 26.6GB / Q5_K_M 28.7GB / Q6_K_S 31.3GB / Q6_K 32.8GB / Q6_K_L 34.4GB / Q8_0 39.4GB / BF16 74.2GB
- Available builds in unsloth/flint-GGUF: IQ1_M 10.6GB / IQ2_XXS 11.0GB / Q2_K_XL 12.6GB / IQ3_XXS 13.5GB / Q3_K_XL 17.2GB / Q4_K_XL 23.3GB / Q5_K_XL 27.9GB / Q6_K_XL 33.8GB / Q8_0 39.4GB / BF16 74.1GB
- Available builds in lmstudio-community/flint-GGUF: Q4_K_M 22.5GB / Q6_K 30.5GB / Q8_0 39.4GB
- Available builds in mlx-community/flint-OptiQ-4bit: MLX 4bit 25.6GB
In addition, 39 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
What to Read Next
- Find models by VRAM (This model runs from the 16GB tier) → Other models that run on a 16GB GPU
- Explore the same model family → flint family overview (2 articles, 5 converted builds)
- Engines that run this model → llama.cpp / Ollama / vLLM
- What Q8_0 mean and where to get this model → GGUF format guide and models / MLX format guide and models / Quantization and model-format glossary
- Other models for the same task → Other text generation models
Sources
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: unsloth/flint-GGUF, lmstudio-community/flint-GGUF, RecursiveSI/flint-int4, mlx-community/flint-OptiQ-4bit
- 2026-11-17: Updated the hardware requirements table with the actual file sizes of unsloth/flint-GGUF.
- 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).

