Fireworks AI 发布了 Tenet，一个与 Harvey 密切合作开发的新模型，专门针对长时法律工作进行了训练。Tenet 是在 Kimi K3 基础模型上进行后训练的，在不增加成本的情况下显著提高了性能，每个 LAB 任务完成的任务量几乎是其基础模型的两倍。该模型在包括法律知识和代理人性能在内的各种基准测试中显示出普遍的改进，在 LAB Contracts 上取得了最先进的成果。 AI
影响 为专业的法律人工智能模型树立了新的标杆，有可能加速其在法律科技领域的采用。
排序理由 来自知名人工智能实验室 (Fireworks AI) 的新模型发布，并附有具体的性能声明。
在 X — Fireworks (inference infra) 阅读 →
AI 生成摘要 · Google Gemini · 来自 5 个来源。 我们如何撰写摘要 →
完整方法见我们的编辑标准。
Tenet came out of a close collaboration with @PereyraJulio @nikogrupen @calvincongelado @gabepereyra: many base models, recipes, and data approaches; through many runs, rollbacks, and harness revisions to reach this checkpoint. We’re so grateful for the partnership!
The performance came without a cost penalty: Tenet runs at $5.92 per LAB task vs $5.62 for base Kimi K3, effectively flat, while completing nearly twice as many tasks. That comes from open-weight per-token pricing and reward shaping for token efficiency.
The gains generalize. On benchmarks Tenet never trained on, it improved on @mercor's Apex Agents and @crosbylegal's Redline Bench, the latter in a different harness entirely. And it showed no meaningful regression on legal knowledge benchmarks like LegalBench, CUAD and MAUD.
On LAB, Harvey's Legal Agent Benchmark, agents produce finished legal deliverables graded against dozens of criteria. All-pass counts a task only if it clears them all. Tenet lifts all-pass from 10.8% to 19.7% over the Kimi K3 base, reaching state-of-the-art on LAB Contracts. h…
Recently @Harvey introduced Tenet, its first model, trained for long horizon legal work. Harvey post-trained it from a Kimi K3 base in collaboration with Fireworks using async RL on our Training API. Promising initial results for both performance and cost-efficiency:🧵 https://t…