Recursive Flint: 40B-param coding model with 2.5B active claims to dethrone Qwen3.6-35B
returnity · reddit · 2026-10-22
Reddit users are buzzing about Flint, a new coding-agent model from Recursive now on Hugging Face (Unsloth GGUFs available).
- Specs: 40B total params with only 2.5B active, a 3:1 SWA-to-full-attention ratio, and 262k context — aimed at GPU-poor users running coding agents.
- Benchmarks: It decisively beats Qwen3.6-35B-A3B on TerminalBench 2.1 and SWE-bench Pro v2, and ties or beats it on the other benchmarks shown, potentially ending Qwen's reign among small coding models.
- WIRED called it "scary good"; a llama.cpp support PR is already up. The poster is running local tests and promises benchmarks in the coming days.
More from coding & agent
- "My job is telling Claude to think bigger"—devs joke about AI ambition gap — paulnovosad · 2026-10-22
- Superset MCP Server Lets LLMs Query Apache Superset via REST API — modelcontextprotocol · 2026-10-22
- Crates MCP Connector Wraps crates.io REST API, Free and Auth-Free — modelcontextprotocol · 2026-10-22
- Ropedia's S-Agent 2.0 Tops Spatial Reasoning Benchmarks With Zero Training, Scoring 66.5 on MMSI — CyberRobooo · 2026-10-22
- Eden Deploys Vercel's Eve Agent to Cloudflare With One Command, Fully Open Source — ritakozlov · 2026-10-22
- Ben Bajarin: rambling research notes into AI for organized workshop analysis works well — BenBajarin · 2026-10-22