Did Spark release a 1.7B X2.6 model that nearly matches its 4B sibling on SWE Verified and τ²-bench?
UnconfirmedLast checked
SparkFirst seen
Verdict
Reddit post links an HF card with self-reported benches; no independent vendor primary checked in this run.
Evidence timeline · 1 source · newest first
Unconfirmed
A LocalLLaMA post says Spark-X2.6-1.7B, with under half the parameters of a 4B sibling, scores 59.8 vs 62.8 on SWE Verified and 85.3 vs 86.9 on τ²-bench at high thinking, and links a Hugging Face model card.