Recursive Debuts Flint, a Compact MoE Coding Model for Local Agent Workloads
Recursive Superintelligence has released Flint, its first model and a notable entry in the fast-growing open-weight coding model space. Flint is a mixture-of-experts (MoE) model with 40B total parameters and about 2.5B active parameters per token, designed specifically for local agentic coding. The company says its automated research system, the Eureka Machine, helped build nearly every part of the stack, from architecture through training. The model weights are available on Hugging Face under the permissive Apache 2.0 license.
The release fits a clear 2026 trend: developers increasingly want capable coding assistants that run on their own hardware, avoiding API costs and keeping proprietary code in-house. Flint's relatively small active parameter count makes this practical without the multi-GPU setups that larger dense models typically demand.
At a Glance
- Architecture: 39.8B total parameters with roughly 2.5B active per token.
- Context window: 262,144 tokens documented on the model card, with the announcement blog citing support for up to 512k tokens.
- Hardware requirements: Runs on 48 GB Apple Silicon Macs (M4 Pro, M5 Pro, M5 Max) or 48 GB NVIDIA GPUs. Quantized 4-bit weights occupy about 23 GB.
- Positioning: Tops Recursive's internal leaderboard among MoE models under 50B for agentic coding, though it trails the strongest dense models overall.
- Strongest results: IFEval at 95.1% and SWE-bench Verified at 74.2%.
- Weakest results: Terminal-Bench 2.1 at 54.9%, well behind the 73.0% posted by dense Qwen 3.8.
Why a Small Active Footprint Matters
The appeal of an MoE design at this scale is efficiency. By activating only a fraction of its parameters per token, Flint keeps inference costs low enough to run on a single high-memory consumer device. That makes it a viable option for developers who want always-available coding agents without cloud dependencies, a priority that has intensified as agentic workflows become more common in 2026. Local deployment also sidesteps latency and data-residency concerns that come with remote inference.
Performance Tradeoffs
Recursive is candid about where Flint excels and where it falls short. Its SWE-bench Verified score of 74.2% places it competitively for its size class, and the IFEval result of 95.1% suggests strong instruction-following behavior, which matters for agent orchestration. However, the Terminal-Bench 2.1 gap against dense models indicates that Flint may struggle with longer-horizon, tool-heavy tasks that stress sustained reasoning. Developers evaluating it for complex agent pipelines should weigh that weakness carefully.
The Eureka Machine Factor
Beyond Flint itself, the release offers a signal about Recursive's broader approach. The Eureka Machine, an automated research system, reportedly contributed to nearly every component of the stack. If automated research can meaningfully accelerate model development, that has implications for how quickly open-weight models could close the gap with proprietary frontier systems. Flint is an early test case for that thesis.
Getting Started
Flint is available now on Hugging Face, with documentation covering context limits, quantization, and hardware requirements on the model card. For teams already running 48 GB workstations, it represents a low-friction way to add a capable local coding agent to their development environment.
via MarkTechPost