Original title: RecursiveSI/flint
Recursive's Mixture-of-Experts (MoE) language model for coding agents: 39.8B total parameters, approximately 2.5B activated per token (including the output head; the rounded name is 40B-A2.5B); 49 layers (2 dense layers + 47 MoE layers, where 35 layers use 513-token sliding window attention and 14 layers use global attention), with 6 of 256 routed experts selected per layer, plus 1 persistent shared experts; 262,144 token context window, with a self-reported knowledge cutoff of 2026 6. Trained from scratch through pre-training, mid-training, SFT, and reinforcement learning, supporting reasoning and multi-turn tool calling. This repository contains the BF16 checkpoint (approximately 80 GB), chat template, and SGLang deployment package; quantized versions for NVIDIA and Apple chips are being prepared separately. The repository was created on 10 21, and was not included in previous issues.
Why it matters Small number of activated parameters allows serving on a single high-VRAM GPU, making it suitable for teams wanting to run coding agents on their own machines
Rising score 84 · Likes 1353 · Downloads 38357
Reasons
- The authors report scores of 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified (agent tasks averaged over at least 5 pass@1 runs, using Terminus 2 for Terminal-Bench and mini-swe-agent for SWE-bench); they also note that Flint's Terminal-Bench score counts solutions obtained via web search as failures, whereas other self-reported scores may not handle this similarly.
- Deployment instructions are actionable: build the SGLang Docker image and serve a single replica on one GPU (B300 or H200), using an OpenAI-compatible interface with direct tool calling support, along with example clients for reasoning and tool calling (including nested parameters and parallel calls).
- Apache-2.0 license, with separate terms for weights and hosted API; Hugging Face likes 1353 (indicating attention only).
Usage
- hf download RecursiveSI/flint --local-dir flint (approximately 80 GB; authors recommend downloading a fixed revision and verifying with hf cache verify), then cd flint && docker build -t flint-sglang sglang, followed by 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 (use serve_h200.sh for H200); once localhost:30000/health_generate returns, test reasoning and a single tool call with python3 sglang/example_client.py --base-url http://localhost:30000/v1.
- Sampling parameters: temperature 1.0, top_p 0.95, top_k 20 (automatically applied by the server if not specified in the request); for difficult problems, allocate an output budget of at least 32768 tokens (including reasoning). For multi-turn conversations, use the chat template included with the checkpoint and retain the returned reasoning_content in the conversation history.
Risks
- Strict VRAM requirements: each replica requires a GPU with ≥141 GB VRAM (B300 / B200 / H200); 80 GB cards are not supported; Transformers modeling code is not yet available, so only SGLang can be used currently; building the image requires access to Docker Hub, PyPI, flashinfer.ai, and GitHub.
- All scores are self-reported by the authors, and evaluation setups vary across providers; the authors caution that long reasoning increases latency and may exhaust the output budget before completion, and that tool-calling applications must implement their own permission and execution boundaries.
- The quantized versions are noted only as "being prepared separately" in the card, with no links provided.