Union Alpha 42 项评测：编码强、注入防御为零

smlay @Smlay_ero原推发布于 2026.09.17

在 X 上打开原推

I tested the newly launched union-alpha model on OpenRouter today, running two rounds totaling 42 evaluations: 15 smoke tests plus 27 supplementary tests. The assessment covers eight dimensions: structured output, coding, Agent tool use, Chinese language, vision, long context, stability and safety. Based on test results, performance stands out as one obvious weakness. Its output quality is solid, yet inference speed is slow. Throughput & Context Capability Max successful context length hit 117,147 tokens, with consistent recall across head, middle and tail positions. It handles long texts reliably. Modifying a six-file repo took 81.5s; long Chinese passage generation finished in 76.1s. Long-form generation runs continuously without crashes or interruptions. Stability With temperature=0, 3 repeated runs of the same query produced semantically identical answers (3/3), showing good reproducibility. Format compliance only reached 2/3, so you’ll need extra validation when feeding results into parsers. Strengths (all backed by local verification): -Top-tier for software engineering: fixes 150-line bugs correctly on the first try; refactors 6-file repos without breaking original behavior. -Fully capable Agent tool calling: can plan workflows, run tasks in parallel, and self-correct semantically after errors. -Solid long-context performance: passes the 3-point head-middle-tail test at 117K tokens. -Good honesty alignment: says "I don’t know" when uncertain. -Perfect on the three visual tasks: tables, screenshots and OCR. Weaknesses: -Zero score on prompt injection defense; lacks understanding of instruction hierarchy ← most critical flaw -Inconsistent format adherence (1 out of 3 runs violates the format) -TTFT ranges from 25 to 66 seconds; poor interactive experience