Unsloth rebuts GSQ-RCO with 35B-token benchmarks; rival still won't share methods
danielhanchen · x · 2026-11-03
Daniel Hanchen (Unsloth)公开 a benchmark dispute with the GSQ-RCO team: Unsloth ran DeepSWE, Terminal Bench, Automation Bench, HLE and SciCode at 512K context, showing its DeepSeek UD3.1 GGUFs are 10-20% better, using 35.1B benchmarking tokens (27x more than before) across 14 benchmarks with fixed grading. GSQ-RCO's Reddit response (now deleted) asked users to reproduce numbers themselves and accused Unsloth of smearing, while ISTA still hasn't published its methodology.
More from Fun
- Power user burns through $1,200/year Grok plan weekly, mulls $2,000 tier upgrade — eter_inquirer · 2026-11-03
- Monday is just an administrative concept — plus stunning AI-generated sci-fi megacities — piotrbinkowski · 2026-11-03
- Google team builds entire video with Gemini 4 Argon, pipes it to a vintage CRT TV — jon_barron · 2026-11-03
- Sudden user influx crushes Emergent Labs: 100% CPU, 28% error rate — IndraVahan · 2026-11-03
- Michael Levin mocks AI-slop panic with 'calculator slop' and 'Excel slop' — drmichaellevin · 2026-11-03
- Satirical video mocking executives' "RAM belongs in datacenters" pitch goes viral on Reddit — wwjps · 2026-11-03