Recursive Releases Flint: Apache 2.0 Open-Weight 40B MoE Coding Model With 2.5B Active Parameters
According to an October 21, 2026 MarkTechPost report, Recursive released its first model, Flint, a 40B-total-parameter MoE coding model with roughly 2.5B active parameters, published under Apache 2.0 weights at about 23GB in 4-bit, runnable on a 48GB Mac or GPU.
What changed
According to an October 21, 2026 MarkTechPost report, Recursive Superintelligence released its first model, Flint, a decoder-only MoE model aimed at coding agents. The report states 39.8B total parameters with about 2.5B active parameters per token, a model card documenting 262,144 tokens of context, and a blog claiming up to 512k. Weights are published on Hugging Face under the Apache 2.0 license. Architecturally, it has 49 layers: the first 2 are dense, followed by 47 MoE layers, each with 256 routed experts selecting 6 per token plus 1 shared expert; attention mixes 35 layers of 513-token sliding window with 14 layers of global attention without positional encoding, with 32 query heads and 4 key-value heads. The report says Recursive's automated research system Eureka Machine contributed to architecture search, quantized the output layer to 4-bit with offline distillation to recover accuracy, and wrote specialized kernels; on M4/M5 chips a Metal megakernel fuses the entire decode forward pass. The report states 4-bit weights are about 23GB and can run on a 48GB Apple Silicon Mac (M4 Pro, M5 Pro, M5 Max) or a 48GB NVIDIA GPU, citing short-context throughput above 125 tokens/sec on M4 Pro and above 150 tokens/sec on RTX Spark. All benchmark scores are self-reported by Recursive, including 74.2% on SWE-bench Verified, 54.9% on Terminal-Bench 2.1, and 95.1% on IFEval, with the NVFP4 version within 0.7 points of BF16.
What it means for developers
For teams building coding agents and local deployments, Flint's reported value lies in combining a high SWE-bench score with the ability to run on a single machine with 48GB of memory. Per the report, 4-bit weights are about 23GB, meaning existing M4 Pro/M5 Pro or 48GB GPU machines may host it without upgrades, changing the prior default assumption that the strongest coding models require a cloud API. Second, the reported MoE active parameters are only about 2.5B (including the output head), and the vendor claims short-context throughput up to 2x faster than comparable models and over 5x faster than dense 25-40B models with similar scores; if this holds in long context, the inference cost structure for high-frequency small tasks such as local batch test fixing and codebase inspection would differ markedly. Third, Apache 2.0 is an explicit open-source license, unlike merely open weights, and the report says the weights are published under it, though commercial use still requires checking the repository's license file and third-party component terms. Note that the report also shows Flint at 54.9% on Terminal-Bench 2.1, clearly below the 73.0% reported for dense Qwen 3.8, so it looks more like a fast, good-enough local agent than the strongest choice for complex terminal-level tasks.
Integration and evaluation
Treat Flint as a candidate rather than a direct replacement for your existing coding model. The first decision is local versus cloud: the report's premise is that your machine has at least 48GB of memory (Apple Silicon or NVIDIA), with 4-bit weights at about 23GB; if you are on a 16-32GB development machine, that premise does not hold, and the alternative is to keep using an API or switch to a smaller model. A verifiable metric is to run one pass@1 evaluation on your own codebase task set against your current baseline, and record whether throughput or quality degrades at long context (for example near 100k tokens), because the reported 125/150 tokens/sec figures are short-context data. The second decision is quantization format: the report says INT4 MLX weights are about 22.6GB on Apple Silicon, and about 22.9GB NVFP4 GGUF on NVIDIA; if you take the NVIDIA path, confirm your inference stack's level of NVFP4 support. A verifiable metric is to compare NVFP4 versus BF16 results on the same task batch; the report claims a gap within 0.7 points, but that figure is vendor self-reported and needs to be reproduced on your own workload. The third decision is harness and tool calling: the report says Flint supports OpenAI-compatible tool calling, that SWE-bench testing used mini-swe-agent without internet access, and that Terminal-Bench used Terminus 2. If your agent relies on web retrieval or a specific tool protocol, first validate tool-calling compatibility and failure modes in an isolated environment before deciding to put it into production. All of the above are recommendations based on the report, not verified conclusions.
Limits and open questions
First, this version of Flint is not currently in KimiSeek's live catalog; the local and online availability mentioned in the report refers to Recursive's own channels and is unrelated to this site. Second, all benchmark scores are self-reported by Recursive, and the report provides no third-party reproduction; the report also specifically notes that its SWE-bench Pro v2 score was measured by Recursive itself and is not comparable to the public SWE-bench Pro. Third, the comparison tables in the report mix vendor self-reported data with measurements by others, so direct cross-comparison requires caution. Fourth, the 125/150 tokens/sec figures are short-context data; long-context behavior, peak memory usage, and stability are unknown. Fifth, the report says the blog mentions context up to 512k, but the model card records 262,144 tokens, so which figure to use needs to be confirmed. Sixth, the specific proportion of Eureka Machine's contribution to the build cannot be independently verified from the report. Eighth, this model is text-only and does not support image input.