DeepSeek-V3.2

DeepSeek-V3.2 has unusually strong public coding evidence for an open model. The supplied DeepSeek/Hugging Face materials describe a reasoning-first model aimed at tool-use and agent workflows, and SWE-bench evidence reports 70.00% resolved on SWE-bench Verified from the official leaderboard, while the DeepSeek paper reports 73.1 with robustness runs in the 72-74 range. LiveCodeBench evidence is also strong, with reported V3.2 Thinking scores in the mid-to-high 80s and Speciale near 89. The main weakness is uneven third-party coverage. DeepSWE/DataCurve was checked but does not list DeepSeek-V3.2, so no score should be inferred there. Terminal-Bench provides relevant agentic evidence, listing Terminus 2 with DeepSeek-V3.2 at 39.6%±2.8 accuracy, which is useful but not dominant. Arena evidence is mixed: broad public preference ranks are mid-pack, with Hard Prompts English rank 80 and score 1455±6 from 12,493 votes. Pricing appears highly cost-effective in Artificial Analysis at $0.28/M input and $0.42/M output tokens with 128k context, but official supplied sources did not clearly establish full pricing or license terms. Artificial Analysis lists speed as N/A, limiting availability scoring. Overall, the model looks strong for coding and agent evaluations, especially where open availability and low token cost matter, but its evidence base is split between provider-reported results, selected benchmark listings, and some missing leaderboard coverage.