On Oct 21, Recursive, the startup founded by Richard Socher, released its first model, Flint: a 39.8B-total MoE coding model with about 2.5B active parameters per token, with weights on Hugging Face under Apache 2.0. Recursive says it leads sub-50B MoE models on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified, runs at 125+ tok/s on an M4 Pro MacBook Pro, and was built from scratch in under four months with its automated research system, the Eureka Machine, contributing to nearly every part of the stack.
Key Takeaways
- ✓Scale: 39.8B total, ~2.5B active incl. output head; 256 routed experts pick 6 plus 1 shared expert; Apache 2.0 (model card)
- ✓Benchmarks (vendor-reported): Terminal-Bench 2.1 54.9%, SWE-bench Verified 74.2%; the card lists SWE-bench Pro v2 at 89.7%
- ✓Speed (vendor-reported): 125+ tok/s on M4 Pro MacBook Pro, 150+ tok/s on 64GB RTX Spark at short contexts, up to 2x comparable models
- ✓Context: 262,144 tokens in the model card config; the blog says up to 512K, best within 256K
- ✓Ecosystem: day-one llama.cpp, Ollama, LM Studio, vLLM, SGLang, plus the open-source rec local engine; free to try on OpenCode and Pi
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the pain point
Local coding models have been stuck between small models that are fast but weak and 25-40B dense models that are capable but slow on a laptop, while many teams cannot send code to the cloud. Recursive, founded by former Salesforce chief scientist Richard Socher, has raised $650M at a $4.65B valuation according to WIRED. On Oct 21 it released its first model, Flint, pitched as fast, local and unmetered.
Architecture and mechanism
Per the model card, Flint is a 49-layer MoE (2 dense plus 47 MoE layers) with 256 routed experts picking 6 plus one shared expert, 35 sliding-window attention layers (513 tokens) and 14 full-attention layers, 39.8B total and about 2.5B active parameters, with a June 2026 knowledge cutoff. Recursive says the Eureka Machine contributed to architecture search, data, optimizers, training recipes and inference kernels; it also quantized the output layer to 4 bits with self-distillation, leaving the NVFP4 build within 0.7 points of 16-bit.
Benchmarks
All vendor-reported: 54.9% on Terminal-Bench 2.1 (Terminus 2, retrieval-obtained solutions counted as failures) and 74.2% on SWE-bench Verified (mini-swe-agent), average pass@1 over at least five attempts. The card lists 89.7% on SWE-bench Pro v2, higher than its Verified score, so check the official chart and independent re-runs. Recursive says Flint beats sub-50B MoE models and rivals 25-40B dense models while running over 5x faster.
Getting started
Try it with ollama run flint (about 23GB); on M4 Pro and M5 Pro/Max Macs, Recursive's open-source rec engine is optimized for it. For servers, the model card ships an SGLang kit: one B300 or H200 per replica, about 80GB of BF16 weights. Recommended sampling is temperature 1.0, top-p 0.95, top-k 20, with at least 32,768 output tokens for hard problems and reasoning_content preserved across turns. It can also be tried free on OpenCode or Pi.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.