> For the complete documentation index, see [llms.txt](https://docs.nvcl.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.nvcl.ai/benchmarks/agents.md). # Agents Coding and terminal agents on Qwen3.8-27B, on NVCL AI and on full-precision vLLM. Coding and terminal agents on NVCL AI's default Standard tier show no significant difference from full-precision vLLM in the share of tasks solved (within noise). They also work the same way on both servers: about as many agent steps per task, about the same peak context, and the same share of reasoning in their output. | Benchmark | NVCL AI | vLLM BF16 | p | | ------------------------------------------------- | ------------------------- | ---------- | ---: | | Terminal-Bench 2.0: trials solved | 107 of 174 | 104 of 174 | 0.68 | | Terminal-Bench 2.0: agent timeouts, both attempts | 18 | 28 | | | DeepSWE v1.1: tasks solved over four attempts | no significant difference | | 0.29 | **p** is the chance of seeing a gap at least this big if the two setups were really equal; above 0.05 means not significant.
Conditions * Qwen3.8-27B, one H200 NVL, mini-swe-agent, thinking at effort `medium`, 262K-token window, 4 agents at once, prefix cache on for both servers. * NVCL AI: Standard tier, with drafts and its own cache, builds of 14 and 16 October 2026, served through its OpenAI-compatible API. * vLLM 0.31.0 with the BF16 model and depth-3 drafts. * Terminal-Bench 2.0: 89 tasks, 87 analysed (2 failed to build on both servers), two attempts per server; NVCL AI build of 16 October 2026; measured 2026-10-16 to 2026-10-17. Each task has its own time limit for the agent; an agent timeout is a trial stopped by that limit. Noise: each server alone changes the outcome of 10-11 tasks between its own attempts. * DeepSWE v1.1: 111 buildable tasks (110 with results in every attempt), 250-step limit per agent; measured 2026-10-15 to 2026-10-19.