Jev 对比 GPT-OSS/DeepSeek/GLM 的基准实测
开发者通宵跑了 LLMBar、JudgeBench、RewardBench，称 Jev 质量接近 GPT-OSS 但便宜 5-6 倍、快约 15 倍。
Early results from my all-nighter experiments with the new (non-LLM, "System One" class) Jev model from @typesafeai I ran LLMBar, JudgeBench, and RewardBench comparing Jev (hosted by Typesafe) to LLMs GPT-OSS, Deepseek flash, and GLM flash (hosted by @FireworksAI_HQ) as you can see, Jev's nearest competitor in these evals (which are designed to be hard, and may be a little too tough for both) is GPT-OSS. GPT-OSS has higher quality on most tests across the board (but lower on the easiest one, LLMBar), at "only" 5-6x the price of Jev. it's also ~15x slower. GLM and Qwen are 50-100x more expensive for similar quality scores on the easiest one. I think Jev is going to find its way into a _LOT_ of production workloads beyond evals... but evals alone would give it a pretty freaking large market. I'm salivating and I can't wait for it to be enterprise ready!