Evaluation

Flint results use temperature 1.0, top-p 0.95, and top-k 20. Terminal-Bench uses Terminus 2; SWE-bench uses mini-swe-agent. Agentic scores report average pass@1 over at least five attempts. † marks scores measured by Recursive; other comparison scores are developer-reported.
Evaluation setups vary across sources. Flint’s Terminal-Bench score counts solutions obtained through web retrieval as unsuccessful; developer-reported scores may not apply the same adjustment.
Model specifications
Table with columns: Property, Flint| Property | Flint |
|---|
| Developer | Recursive |
| Architecture | Decoder-only mixture-of-experts Transformer |
| Total parameters | 39.8B |
| Active parameters per token | Approximately 2.5B, including the output head¹ |
| Layers | 49: 2 dense layers followed by 47 MoE layers |
| Attention | 35 sliding-window layers and 14 full-attention layers |
| Sliding window | 513 tokens: the current token and 512 preceding tokens |
| Attention heads | 32 query heads and 4 key/value heads; head dimension 128 |
| Attention normalization | Per-head QK RMSNorm and learned per-head attention sinks |
| Position encoding | RoPE on sliding-window layers; no positional encoding on full-attention layers |
| Experts per MoE layer | 256 routed experts, with 6 selected per token, plus an always-active shared expert |
| Documented context configuration | 262,144 tokens, including prompt and generated output |
| Vocabulary | 150,144 tokens |
| Input / output | Text / text |
| Stated knowledge cutoff | June 2026 |
¹ The active-parameter convention includes the output head and excludes the full input embedding matrix. The active body alone is approximately 2.2B parameters. “40B-A2.5B” is the rounded model-size designation.
Getting started
Flint runs on one B300 or H200 GPU per replica with the SGLang kit in sglang/. Transformers modeling code is not included yet.
pip install -U huggingface_hub openaihf download RecursiveSI/flint --local-dir flint && cd flint # about 80 GBdocker build -t flint-sglang sglangdocker 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 # H200: serve_h200.shuntil curl -sf localhost:30000/health_generate >/dev/null; do sleep 15; done # progress: docker logs -f flintpython3 sglang/example_client.py --base-url http://localhost:30000/v1 # reasoning and a tool call
from openai import OpenAI client = OpenAI(base_url="http://localhost:30000/v1", api_key="unused")response = client.chat.completions.create( model="flint", messages=[{"role": "user", "content": "Write a Python function that merges overlapping intervals."}], max_tokens=32768,)print(response.choices[0].message.content)
sglang/README.md covers requirements, serve-script options, multi-GPU deployments, and API details.
Recommended generation settings
Table with columns: Setting, Value| Setting | Value |
|---|
| Temperature | 1.0 |
| Top-p / top-k | 0.95 / 20 |
| Output budget for hard problems | At least 32768 tokens, including reasoning |
The server applies these sampling settings when a request omits them. Flint supports OpenAI-compatible tool calls. Use the checkpoint’s chat template and preserve returned reasoning_content in conversation history. Prompt and output share the context window.
Training
Flint was trained from scratch through pretraining, mid-training, supervised fine-tuning, and reinforcement learning. Training includes text and code, instruction following, and agentic tasks.
Limitations and responsible use
Flint’s supervised fine-tuning included safety-oriented examples, but the model can still produce incorrect or harmful content, faulty code, or inappropriate tool calls. Test generated code and review consequential changes before applying them. Tool-using applications should enforce their own permissions and execution boundaries; structured tool-call grammar does not prevent the model from choosing an unsafe or incorrect action.
Long reasoning can increase latency and exhaust the output budget before an answer is complete. The context configuration is a capacity limit, not a guarantee of uniform accuracy across all sequence lengths or tasks. Results can vary with checkpoint precision and inference runtime; evaluate the specific checkpoint and runtime you intend to deploy.
License and reporting
Flint is released under the Apache 2.0 license. Hosted API terms and data practices are separate from the weight license.
For model and serving issues, contact support@recursive.com and include the checkpoint revision, runtime version, hardware, launch command, and a minimal reproduction. Report safety concerns or suspected misuse to safety@recursive.com, and security vulnerabilities to security@recursive.com.