Flint-40B-A2.5B

About

Recursive Flint

Overview

Flint is Recursive's open-weight mixture-of-experts (MoE) language model built for coding agents — models that investigate codebases, fix bugs, implement features, and write tests. It is developed by Recursive Superintelligence, Inc. ("Recursive") and was released on October 21, 2026 under the Apache 2.0 license. Flint has 39.8 billion total parameters with approximately 2.5 billion active per token (rounded model-size designation "40B-A2.5B"); per the card's convention the active count includes the output head and excludes the full input embedding matrix, leaving an active body of roughly 2.2B parameters. The HF repo provides the BF16 checkpoint (~80 GB download), model configuration, tokenizer, chat template, and an SGLang serving kit; Transformers modeling code is not included yet. Quantized variants (including NVFP4 for NVIDIA and runtimes for Apple Silicon) are prepared separately, and the model tree already lists 89 quantizations, 14 finetunes, and 1 adapter.

Architecture

Flint is a decoder-only mixture-of-experts Transformer (config class FlintForCausalLM, model_type: flint). Its 49 layers alternate between 2 dense layers followed by 47 MoE layers (mlp_only_layers [0,1], decoder_sparse_step 1). Each MoE layer routes among 256 routed experts, selecting 6 experts per token, plus an always-active shared expert (norm_topk_prob: true, sigmoid-based score function, moe_intermediate_size 512, shared-expert intermediate size 1024; dense layers use intermediate size 6144 with SiLU activation; hidden size 2048).

Attention is hybrid: 35 sliding-window layers and 14 full-attention layers. The sliding window is 513 tokens (current token plus 512 preceding tokens). The model uses 32 query heads and 4 key/value heads (grouped-query attention) with head dimension 128. Every attention layer applies per-head QK RMSNorm and learned per-head attention sinks (softmax_type: "learnable"). Position encoding is RoPE on sliding-window layers while the full-attention layers use no positional encoding (RoPE-free/NoPE on those layers, θ = 10000). The native context configuration is 262,144 tokens, with prompt and generated output sharing the context window. Vocabulary size is 150,144 tokens; embeddings are untied; dtype is BF16.

Capabilities and usage

Flint supports text input and output, reasoning, and multi-turn tool use (OpenAI-compatible tool calls); the chat template requires preserving returned reasoning_content in conversation history. Recommended generation settings: temperature 1.0, top-p 0.95, top-k 20, and an output budget of at least 32,768 tokens for hard problems, including reasoning. Deployment runs on one B300 or one H200 GPU per replica via the bundled SGLang kit (Docker-based serving with health checks and an example client). Flint is conversational, agentic, and local-friendly: Recursive also ships a local engine (recursive-oss/rec) plus documentation at docs.recursive.com.

Evaluation

In Recursive's release evaluations Flint scores 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified. Agentic scores use temperature 1.0, top-p 0.95, top-k 20; Terminal-Bench runs with Terminus 2, SWE-bench with mini-swe-agent, and agentic scores report average pass@1 over at least five attempts. Scores marked "†" are measured by Recursive; comparison scores are developer-reported. Flint's Terminal-Bench score counts solutions obtained through web retrieval as unsuccessful, an adjustment developer-reported comparison scores may not apply. A benchmark chart also compares the BF16 and NVFP4 checkpoints against local models running on similar hardware.

Training

Flint was trained from scratch through four stages: pretraining, mid-training, supervised fine-tuning (SFT), and reinforcement learning (RL). Training data includes text and code, instruction following, and agentic tasks; the SFT stage included safety-oriented examples. The stated knowledge cutoff is June 2026. Hyperparameters are not documented on the card.

License, limitations and reporting

The weights are released under Apache 2.0 (commercial use and open-source use permitted; hosted API terms and data practices are separate). The card documents limits worth noting: even with safety-oriented SFT examples the model can still produce incorrect or harmful content, faulty code, or inappropriate tool calls; long reasoning can exhaust the output budget; the 262,144-token context is a capacity limit rather than a guarantee of uniform accuracy across sequence lengths; and results vary with checkpoint precision and inference runtime. Contact channels: support@recursive.com, safety@recursive.com (safety/misuse), and security@recursive.com (vulnerabilities), with the checkpoint revision and runtime details included in reports.

Languages

The card does not document specific supported languages; card, docs, and reporting channels are in English, so the language list is left unset rather than guessed.

Training Data Trained from scratch through pretraining, mid-training, supervised fine-tuning, and reinforcement learning; data includes text and code, instruction following, and agentic tasks. Stated knowledge cutoff: June 2026.