fyi
sample on this subset of term-bench2.0 tasks was already enough to me feel free to bench more if you want, tested with pi-agent
βββββββββββββββββββββββββββββ¬ββββββββββββββ¬βββββββββββββ¬ββββββββ¬ββββββββ¬ββββββββ¬ββββββββ
β Task β Qwen3.6 β Darwin β Q dur β D dur β Q out β D out β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β fix-git β 3/3 β 2/3 β 41s β 31s β 2.3K β 1.6K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β prove-plus-comm β 2/3 β 2/3 β 377s β 36s β 11K β 1.9K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β cobol-modernization β 1/3 β 2/3 β 439s β 215s β 26K β 13K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β overfull-hbox β 0/3 β 0/3 β 484s β 103s β 29K β 5.9K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β break-filter-js-from-html β 0/3 β 0/3 β 297s β 275s β 18K β 16K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β filter-js-from-html β 0/3 β 0/3 β 80s β 671s β 4.6K β 33K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β kv-store-grpc β 2/3 β 0/3 β 34s β 42s β 1.3K β 1.8K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β multi-source-data-merger β 3/3 β 1/3 β 64s β 98s β 3.5K β 5.8K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β regex-log β 1/3 β 0/3 β 461s β 580s β 28K β 34K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β git-leak-recovery β 2/3 β 1/3 β 35s β 39s β 1.8K β 1.9K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β pypi-server β 0/3 β 0/3 β 23s β 46s β 0.9K β 2.5K β
βββββββββββββββββββββββββββββΌββββββββββββββΌβββββββββββββΌββββββββΌββββββββΌββββββββΌββββββββ€
β TOTAL β 14/33 (42%) β 8/33 (24%) β β β β β
βββββββββββββββββββββββββββββ΄ββββββββββββββ΄βββββββββββββ΄ββββββββ΄ββββββββ΄ββββββββ΄ββββββββ
Thanks for running the benchmark and sharing the numbers.Quick note on positioning: Darwin-36B-Opus is published as a reasoning-focused evolutionary merge (GPQA Diamond 88.4%, tying Qwen3.5-397B-A17B), not as an agentic coder. The Darwin Opus line is bred for graduate-level scientific reasoning β physics, chemistry, biology Q&A in the GPQA style β and is not tuned for terminal/agent workflows. For agent and coding tasks we'd recommend the Qwen Coder line.Two observations on your runs that may explain part of the gap:
System prompt: Darwin needs enable_thinking=true via the Qwen chat template, and the agent harness needs to leave room for the ... block before tool calls. If pi-agent strips or truncates the thinking trace, Darwin loses most of its reasoning lift. You can confirm in the output β if you don't see a block, the harness is filtering it.
Output token compactness is by design: Darwin Opus inherits a Father with 75% Gated-DeltaNet + 25% Gated-Attention. Post-thinking responses are deliberately compressed (FFN Ξ± asymmetry from the merge genome), which is the opposite of what agent benchmarks reward β they reward verbose step-by-step tool chains. That's a known trade-off for this checkpoint, not a regression.
We'd be very interested to see your numbers on the same subset with (a) enable_thinking=true set in the request, and (b) the agent template that preserves the thinking trace. Happy to help if there's a specific task where you'd like to dig in.For full context: the Darwin Family methodology is currently under peer review at ARR May 2026 (training-free reasoning scaling) β coding/agent performance is explicitly out of scope of that submission.