Recursive Flint benchmarks: matches Qwen3.8-27B on agentic tasks, weak on long context
ChengleiSi · x · 2026-10-23
Blogger nettermina published a detailed hands-on benchmark of Recursive's new coding model Flint, calling it "the closest thing to a hypothetical Qwen 3.8 35B."
- Thinking mode: no traditional switch — thinking is always on with no level control, so how your harness handles its reasoning is critical to results
- Agentic capability: matched a 27B model on both tier2/tier3 agentic tests (96.3% vs Qwen's 100% on tier2; 67.6% vs 76.5% on tier3), rarely seen at this tier
- Speed: 40B params (2.5B active), 24.2GB weights at Q4; 261 tok/s decode on RTX 5090, 147 tok/s on RTX 3090
- Weakness: long context — payload fidelity drops to 80.5% vs Qwen's 98.3%, Ledger walk falls to 25.9%, and format drift after long thinking is common
Verdict: a genuinely competitive same-tier model with a proper harness, but be careful in long-context workloads.
More from Models
- Abliterated GLM-5.3 sells for $10 as "world-class penetration tester", fueling vulnerability surge — deedydas · 2026-10-23
- DeepSeek harness slashes tokens, time and cost vs Codex, but tanks on FrontierCode — teortaxesTex · 2026-10-23
- 345K trajectories reveal LLM agent personalities: Claude most planful, GPT-5.5 least energetic — mdredze · 2026-10-23
- DeepSeek V4.1 Coding Benchmark Labels Look Wrong: 47 vs 46 Score Confusion — teortaxesTex · 2026-10-23
- DeepSeek 4.1 Flash Hits ~80 tok/s Locally on DGX Station via LM Studio — MatthewBerman · 2026-10-23
- MiniMax M3.1 Flash launches day-zero on Modular Cloud, tops open models in cowork automation — MiniMax_AI · 2026-10-23