The lab used Recursive’s official Q4_1 GGUF file, llama.cpp b11846, a Mac Studio, one conversation at a time, and 64K context. Its recorded run peaked at 26.4 GB of memory.
Across 15 scored tests, FreshContext reported 98.3% (59/60) on checking extractions and 100% (15/15) on choosing and chaining named tools. Flint was much less likely to call a tool without explicit instruction: it scored 2/12 when a tool held the answer but was not mentioned, and 1/12 when no tool was mentioned. Median generation speed was 74.2 tokens per second; the lab estimated that reading a fresh 32,000-token prompt took about 70 seconds.
FreshContext did not run coding benchmarks. It says Recursive’s reported 89.7 SWE-bench Pro v2 result looked implausibly high beside a 74.2 SWE-bench Verified result, but its own work neither reproduces nor disproves either coding score. The source also does not establish how Flint compares with Qwen3.6-35B-A3B on coding or whether llama.cpp matches Recursive’s own runtime results.