Recursive Releases Flint, a 40B Open-Weight MoE Coding Model for Local Agents

Flint is designed to run locally. Recursive says it runs on 48 GB Apple Silicon Macs, including M4 Pro, M5 Pro and M5 Max systems, or on 48 GB Nvidia GPUs; 4-bit weights take about 23 GB. The model card documents a 262,144-token context, while Recursive's blog cites up to 512,000 tokens. The company reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64 GB Nvidia RTX Spark, both short-context figures and up to twice as fast as comparable models.

Flint is a decoder-only MoE Transformer. It has 49 layers, two dense and 47 MoE. Each MoE layer has 256 routed experts, and a router selects six per token plus one always-active shared expert. Thirty-five layers use a 513-token sliding window, while 14 use full attention with no positional encoding. The model has 32 query heads and four key/value heads. The 2.5 billion active figure includes the output head; the active body is about 2.2 billion parameters.

Recursive says its Eureka Machine, an automated research system, helped build nearly every part of the stack. The system proposes ideas, implements them, tests them and chooses what to try next, according to the company. It started from a 35-billion-total, 3-billion-active baseline and searched expert configurations and widths to arrive at 40 billion total and 2.5 billion active parameters. It also quantized the output layer, or lm_head, to 4 bits, against common practice. An offline distillation stage with the 16-bit model as teacher recovered accuracy and nearly halved training steps versus a quantization-aware training baseline, Recursive says. The NVFP4 model scores within 0.7 points of the 16-bit model. Eureka also wrote specialized kernels for low-bit projections, MoE operations, attention and prompt processing; on M4 and M5 chips, a Metal megakernel fuses the entire decode forward pass.

All Flint benchmark scores cited by the company are self-reported. In Recursive's results, the BF16 model scores 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified; the NVFP4 version scores 53.7% and 74.1%, respectively. The BF16 model reports 89.7% on SWE-bench Pro v2, 80.3% on LiveCodeBench v6, 89.2% on IFBench and 95.1% on IFEval. The NVFP4 version reports 88.5%, 80.8%, 88.1% and 94.6% on the same tests.

Terminal-Bench used the Terminus 2 harness, and solutions found through web retrieval counted as failures. SWE-bench runs used mini-swe-agent with no internet access. Agentic scores average pass@1 over at least five attempts, according to Recursive. The company's comparison chart places Flint at the top among MoE models under 50 billion parameters on agentic coding but behind the strongest dense models. In that chart, dense Qwen 3.8 scores 73.0% on Terminal-Bench 2.1, above Flint's 54.9%, while Qwen3.6 scores 73.4% on SWE-bench Verified, slightly below Flint's 74.2%.

The report says competitor figures come from each developer's own model card, not from Recursive's chart. Qwen3.6-35B-A3B lists 35 billion total and 3 billion active parameters, a 262,144-token native context expandable to 1,010,000 tokens, text and image input, an Apache 2.0 license and 73.4% on SWE-bench Verified. Poolside's Laguna XS 2.1 lists 33 billion total and 3 billion active parameters, a 262,144-token context, text-only input, an OpenMDW-1.1 license, 70.9% on SWE-bench Verified and a 36 GB Mac requirement. Google DeepMind's Gemma 4 31B lists 30.7 billion dense parameters, a 256K context, text and image input, an Apache 2.0 license and 80.0% on LiveCodeBench v6.

Editor's Summary Recursive Superintelligence has released Flint, a 39.8-billion-parameter open-weight MoE coding model with about 2.5 billion active parameters and Apache 2.0 weights for local deployment on 48 GB hardware. Recursive says its Eureka Machine automated much of the design and reports 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1, though those results are self-reported. The company's chart puts Flint ahead of sub-50B MoE models on agentic coding but behind dense Qwen 3.8 on Terminal-Bench, while competitor figures elsewhere come from developers' model cards.