Perplexity 和 NVIDIA 搞了个叫 JOLT 的训练方法,一个模型当老师又当学生,在 MATH 上准度追平 GRPO,训练 token 少了 6.5 倍。
Perplexity
AI answer engine with cited sources
What they do
Founded in San Francisco in 2022, Perplexity runs an AI answer engine that cites its sources, and has launched the Comet browser.
Latest coverage
Stories in the feed that mention this companyPerplexity 开源了 0.6B 和 9B 两款多模态嵌入模型,0.6B 就能打大模型,还支持 9B 建库、0.6B 查询的省资源玩法,做检索的可以试试。
云算力公司 Modal 第一次用上英伟达的 Vera CPU,测试显示智能体吞吐量翻倍,CPU 市场有新玩家了。
一篇把 ChatGPT、Gemini、DeepSeek V3 等 10 个模型拉到同一起跑线测幻觉的论文,结论很反直觉:答得全的不一定准,Falcon 零幻觉但漏掉大半答案。
做 LLM 部署的可以看看,EAGLE-4 开源了,推测解码能自适应硬件和负载,Qwen3.8-27B 实测跑到 961.6 tokens/s。
Perplexity 的桌面应用现在能完全离线跑了,本地模型用的是微调过的 Qwen 3.8,复杂任务还能切回云端,36 GB 内存起步。
Perplexity Computer 能在 Mac 上离线跑模型了,用 Qwen 3.6 端侧推理,不过你得有 36GB 内存的 Apple 芯片机器。
Perplexity 把开源的 SPACE 沙箱交给 Crusoe 云运营,跑 Agent 又多了一个带隔离环境的算力选择。
The benchmark was run with a Stirrup agent harness and GPT-5.6 Luna (medium), so the gains reflect real agent workflows rather than single-shot queries. Teams building search-backed agents can now choose between auto mode for quality and a lite tier costing 33% less per task, both positioned on the cost-versus-capability Pareto frontier. The results put Exa level with Perplexity's Search API, giving agent developers a directly comparable alternative at published per-task prices.
Perplexity 出了个 Flash 决策模型,16GB 的 M4 Mac 就能跑,准确率跟 Wald 打平还快近一倍,适合邮件补全这类本地场景。
Isolating each agent task in a Firecracker microVM with externally managed credentials is a concrete, reusable pattern for agent safety that goes beyond model-level guardrails. Developers building autonomous agents can adopt SPACE directly or borrow its architecture to limit blast radius from malicious prompts. The Apache 2.0 license allows both commercial and open-source integration without restrictions.
DeepSeek 开源了推理更强的 V4.1-Flash-1017,Perplexity 也开源了两款决策模型,程序员和折腾党今天能下载的好东西不少。