OpenHands LM is a focused, text-only 33B coding-agent model based on Qwen Coder 2.5 Instruct 32B. Its official launch material and model card report 37.2% SWE-bench Verified resolve rate and a 128K/131K context limit. That is meaningful task-specific evidence, but materially below the coding and broad-capability profiles represented by Claude Opus 4.8, GPT-5.5, GPT-5.6 Sol, and Kimi K2.5. It is also below DeepSeek-V3.2 and GLM-4.6 on the calibrated coding score because their published evaluations support stronger or broader current performance. MIT-licensed downloadable weights are its main practical advantage. Local deployment can make it cost-effective for teams with suitable hardware, and OpenHands provides a mature open workflow, documentation, runtime material, and repository integration. This supports a stronger developer-experience score than GLM-4.6 and DeepSeek-V3.2 in the supplied anchors. The supplied material does not establish image, audio, structured tool-I/O, throughput, latency, or a concrete hosted price, so multimodal, speed, and pricing scores remain conservative. Evidence quality is moderate rather than high: the SWE-bench figure is provider-reported, though repeated in the official model card. No DeepSWE/DataCurve result, LiveCodeBench score, Terminal-Bench/Aider result, or LMArena/Arena-Hard preference result is published in the supplied evidence. Artificial Analysis was checked but supplied without extractable benchmark, price, or speed figures. Adoption is supported by the OpenHands ecosystem, not model-specific usage data.
