Recursive Flint: Open-Weight MoE Coding Model, 2.5B Active

Introduction

On October 21, 2026, Recursive released Flint, an open-weight mixture-of-experts (MoE) model for agentic coding that is built to run on local hardware. Flint has 40 billion total parameters (39.8 billion on the model card) and activates roughly 2.5 billion per token. The weights are on Hugging Face under the Apache 2.0 license and are not gated. Recursive says a small team of people and agents built the model from scratch in under four months, and that its automated research system, the Eureka Machine, helped develop nearly every part of the model-building stack. TechCrunch describes Flint as Recursive's first model and an early public test of recursive self-improvement.

A note on evidence: the benchmark and speed figures below are reported by Recursive unless stated otherwise. We have not reproduced them.

Feature Overview

Sparse architecture. The model card describes a decoder-only MoE Transformer with 49 layers: 2 dense layers followed by 47 MoE layers. Each MoE layer has 256 routed experts, 6 of which are selected per token, plus an always-active shared expert. Of the 49 layers, 35 use sliding-window attention (513 tokens) and 14 use full attention. The "approximately 2.5B active" figure includes the output head; the card says the active body alone is about 2.2B. Recursive says the Eureka Machine started from a baseline of 35 billion total and 3 billion active parameters, explored expert configurations and widths, and arrived at a shape it projected to give the best speed and coding trade-off among the configurations it tested.

Quantization. After training, the Eureka Machine searched for quantization recipes. Against common practice, it quantized the output layer (lm_head) to 4 bits. Recursive's reasoning is that with only 2.5 billion active parameters, the output layer is a large share of the weights read for each token. To recover accuracy it added an offline distillation stage in which the quantized model trains on its own pre-sampled responses from a self-discovered coreset, with the 16-bit model as teacher, which Recursive says nearly halved the training steps compared with its quantization-aware training baseline. The final NVFP4 model scores within 0.7 points of the 16-bit model across Recursive's benchmarks. TechCrunch, citing the launch video, reports that low-bit training shrank Flint by 71% with almost no loss in quality.

Kernels and inference. Recursive says the Eureka Machine produced specialized kernels for low-bit projections, MoE operations, attention and prompt processing, including a Metal megakernel that fuses the entire decode forward pass on M4 and M5 chips. Recursive also released rec, a lightweight inference engine for Flint on M4 Pro and M5 Pro/Max Macs. Its README claims 1.4 to 2.6 times the token throughput of llama.cpp across context lengths from 1K to 252K tokens on an M4 Pro with 48 GB and 4-bit weights. Recursive also lists llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang as compatible.

Speed claims. Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64GB NVIDIA RTX Spark, "at short contexts" and "up to twice as fast as comparable models." TechCrunch reports about 131 tokens per second on the M4 Pro in Recursive's tests, against 58 to 69 for similar models including Gemma 4 and Nemotron 3.5 Lightning. Recursive also says it runs more than five times faster than dense 25 to 40 billion parameter models with similar coding scores. The caption on Recursive's first chart describes average tokens per second across a 128k-token session with prefill excluded, which is a different measurement from the short-context headline numbers.

Context window. Recursive's post says Flint supports up to 512k tokens and recommends staying within 256k. The model card documents a 262,144-token context configuration. The two documents do not match, so 256k is the safer planning figure.

Benchmark Reading

Recursive says Flint leads MoE models with fewer than 50 billion total parameters on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified, and rivals dense 25 to 40 billion parameter models. That is a claim about MoE models under 50B total parameters, not about all models of that size. The model card's chart gives the detail (BF16 unless noted; read from the chart image, where some comparison scores are marked as measured by Recursive and others as developer-reported):

Against the sparse MoE group in the chart, Flint's BF16 model is ahead on five of six panels and effectively tied on LiveCodeBench v6 (80.3 versus 80.4 for Qwen 3.6). Against the dense group it is mixed: the Qwen 3.8 dense model scores higher on the four coding panels, by the widest margin on Terminal-Bench 2.1, and a model labelled Muse Glimmer scores 76.0 on SWE-bench Verified against Flint's 74.2. TechCrunch's report that only Alibaba's Qwen3.8 27B beat Flint on Recursive's six-test average fits that picture. The chart also shows BF16-to-NVFP4 gaps of up to 1.2 points on individual benchmarks, so the 0.7-point figure reads as an aggregate rather than a worst case.

Evaluation setup matters. Flint results use temperature 1.0, top-p 0.95 and top-k 20, with agentic scores averaged over at least five attempts. Terminal-Bench counts solutions obtained through web retrieval as unsuccessful, and the card notes that developer-reported scores for other models may not apply the same adjustment.

Usability Analysis

"Laptop" needs qualifying. The rec README lists an M4 Pro, M5 Pro or M5 Max Mac with at least 48 GB of unified memory and about 30 GB of free disk. The full BF16 checkpoint on Hugging Face is roughly 80 GB, and the card's SGLang serving kit targets one B300 or H200 GPU per replica. The card says quantized variants are prepared separately for NVIDIA and Apple Silicon runtimes. The card also says Transformers modeling code is not included yet.

Recursive recommends the sampling settings above and an output budget of at least 32,768 tokens for hard problems, and asks users to preserve returned reasoning content in conversation history. It is candid about capability: Flint "may need more guidance than the larger closed-source coding models," so Recursive suggests breaking complex work into smaller subtasks with explicit instructions.

For independent impressions there is one early report. According to TechCrunch's summary, Wired's Will Knight ran Flint on a Mac with 48GB of memory to build several small programs, including a to-do list app and a tool that scours the web for research, and found it handled his requests every time, where smaller coding models often hit errors or ran slowly. That is a few tasks, not a benchmark.

Openness and Licensing

Flint's weights are Apache 2.0, and the rec engine's GitHub metadata also shows Apache 2.0. "Open-weight" is the accurate description, not "fully open." The model card describes training only in general terms (pretraining, mid-training, supervised fine-tuning and reinforcement learning on text, code, instruction-following and agentic tasks). The repository lists weights, configuration, tokenizer, chat template and serving scripts, and no training data or training code. The card also says hosted API terms and data practices are separate from the weight license.

Pros and Cons

Strengths: permissive licensing for local use, a sparse layout aimed at fast decoding, a quantized variant reported close to full precision, and published evaluation caveats.

Limitations: vendor-reported results, hardware needs above a typical laptop, a stated need for extra guidance, and no released training data or code.

Outlook

Recursive frames Flint as a starting point. Its stated goal is recursively self-improving superintelligence, beginning with model development, and it says more capable models should make the Eureka Machine a better researcher. TechCrunch notes Recursive did not say how much time the system saved its researchers, and reports the company emerged from stealth in May with $650 million at a $4.65 billion valuation. The near-term test is outside validation: independent benchmark runs, community quantizations for more modest hardware, and the context-length discrepancy.

Conclusion

Flint is a notable open-weight option for developers who want an agentic coding model on their own machine and have a high-memory Apple or NVIDIA setup. Its numbers are strong for its active size but remain vendor-reported, and the dense Qwen 3.8 model scores higher on several coding panels in the model's own chart. It suits developers who value local control and permissive licensing, and who are comfortable giving a smaller model explicit, decomposed instructions. We rate it 4 out of 5, pending independent verification.

Editor's Verdict

Recursive Flint: Open-Weight MoE Coding Model, 2.5B Active earns a solid recommendation within the Other LLM space.

The strongest case for paying attention: the Apache 2.0 license on downloadable weights allows the model to run offline with code and data kept on the user's own machine. That alone raises the bar for what readers should expect in this space. Reinforcing that, a sparse expert layout with about 2.5B active parameters per token gives vendor-reported decoding above 125 tokens per second on an M4 Pro MacBook Pro — practical value rather than just headline appeal. The broader signal worth registering is straightforward: a system that proposes, implements and tests its own ideas is the notable claim here, although Recursive has not said how much researcher time it saved. On the other side of the ledger, one constraint is real rather than a marketing footnote: the headline benchmark and speed numbers come from Recursive itself, and the dense Qwen 3.8 model scores higher on all four coding panels in the model card's chart. It should factor into any serious decision. Layered on top of that, a 48 GB unified-memory floor for the rec engine and an 80 GB BF16 download make running on a laptop a high-memory-machine proposition in practice — which narrows the set of teams for whom this is an obvious yes.

For developers with a high-memory Apple Silicon or NVIDIA machine who want a permissively licensed coding model that runs locally, this is a serious evaluation candidate, not just a curiosity to bookmark. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.