Flint
Recursive·
LLMsopen-weightEdge / Consumer
Overview
First model from Recursive (Recursive Superintelligence, Inc.): a 39.8B total / ~2.5B active MoE for local coding agents under Apache 2.0, trained from scratch (pretraining, mid-training, SFT and RL) rather than post-trained from another lab's weights. Recursive states that its automated research system, the Eureka Machine, carried out most of the development. 262,144-token context per the model card. Vendor-reported at launch: SWE-bench Verified 74.2 (mini-swe-agent) and Terminal-Bench 2.1 54.9. A 4-bit build sits at roughly 23 GB, right at the edge of a 24 GB consumer GPU.
Capabilities and innovations
39.8B total / ~2.5B active parameters (MoE)Agentic coding and terminal tasks262K-token contextGGUF and MLX buildsApache 2.0 licenseFirst release from RecursiveDevelopment largely automated by Recursive's Eureka Machine system (vendor statement)
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| SWE-bench Verified | 74.2% | vendor SWE-bench Verified 74.2 via mini-swe-agent (temperature 1.0, top_p 0.95, top_k 20). Era-1 benchmark, recorded only because Recursive reports it. Card not fetchable from this environment; value taken from consistent secondary transcriptions (MarkTechPost, LM Studio, Wired coverage). |
| Terminal-Bench 2.x | 54.9% | vendor Terminal-Bench 2.1 54.9 (temperature 1.0, top_p 0.95, top_k 20). 2.x generation, so it belongs in this field. Card not fetchable from this environment; value taken from consistent secondary transcriptions (MarkTechPost, LM Studio, Wired coverage). |
- SWE-bench Pro: not reported by the vendor — The model card reports 89.7 on "SWE-bench Pro v2", a different generation from the SWE-bench Pro tracked in this field (field values sit around 50 to 66). Not comparable across that boundary, so the value is kept here only.
- GPQA Diamond: not reported by the vendor — Coding-only launch table.
- Humanity’s Last Exam (no tools): not reported by the vendor — Coding-only launch table.
- MMLU-Pro: not reported by the vendor — Coding-only launch table.
Architecture and hardware
- Parameters
- 39.8B total, 2.5B active per token (MoE)
- Estimated VRAM at Q4
- ~23 GB, Edge / Consumer class
- Quantization formats
- GGUF, MLX, MLX-4bit, BF16
- Recommended runtime
- LM Studio
- License
- Apache 2.0
Links
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.