Perplexity 年化营收破 10 亿美元,增 4 倍
Perplexity 今日年化营收运行率突破 10 亿美元,CEO Aravind Srinivas 在一场直播活动中向 Semafor 透露了这一消息。The Information 此前报道称,该公司今年年初的年化营收不足 2.5 亿美元,到 8 月已超过 7.5 亿美元,其中 Computer 贡献了部分增长。按照智能体实验室通常的估值倍数(例如 Cognition 也突破了 10 亿美元,Cursor 则突破 46 亿美元),如果单位经济效益站得住脚,这意味着 Perplexity 目前估值约为 300 亿至 500 亿美元。相比 2025 年 9 月那轮 200 亿美元估值,这是一次不错的提升,也符合我们为标题报道设定的“融资额 >100 亿美元”规则。
为方便还没跟上的读者,这里汇总一下 Perplexity 近期的一些发布:
Computer。9 月推出的 Automations 能让它从中断处继续处理周期性任务。11 月推出的 Teammates 可承接持续进行的工作。Computer 还能在受支持的 Mac 上完全本地运行。
Monitor API。关注某个话题或页面,当相关内容发生变化时,你会收到 webhook 通知。
Frontier Research。这个新成立的研究实验室可使用 10,000+ 块 GPU,并计划公开发表论文、数据、模型和代码。
由 Deepgram 赞助。Deepgram 是支撑自然对话的语音 AI 基础设施——为构建语音 AI 的团队提供语音转文本、文本转语音、专注语音的 LLM 以及智能体。无论你使用什么技术栈、或已经习惯用哪些工具开发,Deepgram 都能作为你的基础语音 AI 技术无缝接入。掌控权在你手中。而我们具备像人一样聆听、理解和说话的上下文能力。
进一步了解 Real-time Context Standard。
Qwen4-Max 已在实际场景中现身
周五的 Qwen4-Max 预览版如今已成为正式发布的模型。在最高推理设置下,Qwen4-Max 在 Intelligence Index 上得分 43,高于此前的 34。平均成本从每任务 10.02 美元降至 3.27 美元。阿里巴巴将输出价格从每百万 Token 6 美元下调至 4.50 美元,同时该模型生成的 Token 数量减少了约 39%。Kimi K3.1 Pro 每完成一个 index 任务成本为 3.68 美元,在 Terminal-Bench 4.0 上得分 67%,而 Qwen 为 48%。Artificial Analysis
模型内部。阿里巴巴将总参数量维持在 2.4T,并把每 Token 激活的参数量减少 25% 至 71B。该模型可接受图像和文本输入,上下文窗口为 1M Token。更小的开放权重 Qwen4 模型仍将陆续推出。阿里巴巴
Arena 与 Vals:Arena 将其排在 Agent Arena 总榜第 14 名,是中国模型中排名最高的,在 14.7k 次会话中净提升 +7.5%(Arena)。
WebDev:它在 Code Arena WebDev 中以 1750 分位列第 6,比前代高出 74 分(WebDev)。它还在 Image-to-WebDev 中位列第 6,略高于 Claude Opus 5 和 GPT-6 Sol(I2WebDev)。
Vals Index:它在总榜位列第 11,在编程和金融领域提升显著(Vals)。
法律基准测试的注意点:在 HLAB 上,它答对了 95.7% 的清单项,但仅通过 21.3% 的任务,且其输出 Token 量增加 3.6 倍(HLAB)。
视觉结果(独立):Roboflow 称其在自家 VLM 基准测试上表现平平,Google、OpenAI 和 Anthropic 目前已远远领先。不过在高推理强度下,它每张图像的处理速度比 Qwen3.8-Max 快约 4.5 倍,成本低约 68%(Roboflow)。
安全、对齐与评测诚信
OpenAI 嵌入式评估:OpenAI 正在向 Apollo、METR、Transluce 和 Goodfire 开放其模型与系统的内部访问权限(OpenAI)。
各组范围:Apollo 派驻六名评估人员驻场四个月(Rundown)。METR 审查 OpenAI 规模最大的强化学习训练运行,可访问训练记录、监控记录和检查点快照(METR)。Transluce 对后训练对齐评估开展为期六周的审计,并将在 2027 年初进行更全面的后续审计(Transluce)。Goodfire 派驻可解释性研究人员,以监控训练运行(Goodfire)。
局限:METR 表示,其范围不包括 OpenAI 已披露的失准事件,且比 METR 提议的范围更窄(scope)。
对访问条款的批评:批评者称,这些协议未能达到 Sam Altman 所承诺的“类员工访问权限”。
据报道存在的缺口:评估人员无法访问 Slack,由 OpenAI 决定他们访谈哪些员工,且他们无法访问权重,也无法开展基于训练的引导(elicitation)(Neel Nanda、Alex Mallen)。
前员工的说法:Mikita Balesni 表示,评估人员不能引用员工的话,即便是匿名引用也不行,且部分评估人员可能被禁止就仅限内部使用的模型发表内容(balesni)。
Goodfire 收购谈判:The Information 报道称,OpenAI 曾探讨以超过 50 亿美元收购这家可解释性初创公司。据报道,Goodfire 希望保持独立,以便能向多家实验室出售(report)。
其他诚信信号:
基准测试污染:Claude 似乎记住了 LinuxArena control eval 的细节(Karvonen)。
事件分类法:OpenAI 发布了一个用于对 AI control 事件进行分类的 v0.1 框架(framework)。
披露政策:Anthropic 正把其协调披露窗口从 90 天缩短至 30 天,适用于高危及严重级别漏洞(policy)。
推理系统与智能体基础设施
Together 不再走手工调优巨型内核(megakernel)的路线。其首个设计已在一家客户处投入生产。如今智能体转而搜索更小的 GPU 算子组进行融合。Together 报告在 B200 上运行 Llama 3.2 1B 时吞吐提升 28%。工程博文
DeepSeek V4.1 Flash 在 Vera Rubin 上:vLLM 搭配 Dynamo 在 SemiAnalysis 的 AgentX 基准测试中显示,在 400 tok/s/user 下每芯片吞吐是 GB300 NVL72 的 63.4 倍,在 600 时为 93.3 倍(vLLM)。
批大小为 1 的延迟:SGLang 相关的工作在 TP4、批大小为 1 时达到约 2,500 tok/s,是其 GB300 结果的 1.7 倍(saltyph)。Baseten 报告其相比自身 4xB300 的数据也有类似的 1.7 倍提升(Baseten)。
智能体驱动的优化:一个智能体通过 NVLink 对称内存跨 HBM 共享 Engram 表,以恢复前缀缓存容量(Liu)。另一个团队表示,智能体找到的优化使其 B300 单步时间加快了 10%(Halex)。
智能体式服务研究:MIT 的 Evolving Infra 用在线自适应取代固定的预填充/解码分离。在与 NVIDIA Dynamo 的对比中,基于最多 256 块 GB300,其报告延迟最多降低 4.1 倍,P95 TTFT 最多降低 14 倍(Evolving Infra)。
流水线并行预填充:Inferact 发现,在 2s TTFT 目标下,MiniMax-M3.1-Flash 上的 PP2 每 GPU 可持续处理的请求数比 DEP2 多 23%(Inferact)。
前缀缓存陷阱:Ollama 发现,系统提示词中一个时间戳,或推理努力程度的一次变化,都可能让缓存失效(Ollama)。
沙箱与训练:
外部沙箱:Simon Willison 介绍了在 Mac 上运行智能体、同时把其代码放到 WebAssembly、QEMU 或 Apple Container 沙箱中执行的做法。该模式还包括白名单代理和凭据替换(文章)。
MLPerf:在 MLPerf Training v6.1 中,GB300 NVL72 用 19.7 分钟完成了 GPT-OSS-20B 的训练,比 H100 快 2.86 倍(Lambda)。
开放模型统一端点。Hugging Face Inference Providers 现已在 OpenRouter 上运行,使用相同的端点和 Token。Hugging Face
研究、基准测试与评估
PyBot:由 1,000 个 DeepSeek-V4.1-Flash 智能体组成的集群花了一周时间编写了一个基于规则的聊天机器人,不含任何神经网络(lewtun)。
规模:这些智能体交换了约 175k 条消息,合并了约 8k 个 PR。
结果:该机器人是 630k 行 Python 代码,零依赖,据报告在日常问题上与 SmolLM-360M 相当。
自我改进与训练:
AutoRL:监督一次实时 RL 运行的智能体,将 Qwen3 在 AIME 2024 上的成绩从 46.67% 提升到 66.67%(omarsar0)。
基准测试投机的警示:一份综述警告称,许多自我改进智能体的提升来自学习基准测试本身,而非学习任务(TuringPost)。
权重空间编程:Opus 5.5 手写 transformer 权重,以中位数 166 个参数解决了算法任务(dair_ai)。
数据策管:DatologyAI 公布了其自称 6 倍预训练算力倍增效应背后的技术报告(Datology)。
蒸馏复现:一项小规模复现实验证实,当学生模型从头训练时,扰动输入蒸馏确有帮助,但发现将其蒸馏进预训练的学生模型时反而有害(Maini)。
新基准测试:
GradeBench:在批改手写物理答案时,最强的前沿模型与教师评分一致的比例仅为 57%(GradeBench)。
首音频时间(Time to First Audio):AA 新的 TTS 延迟基准测试发现,流式模型的中位数为 376 毫秒,而非流式模型为 2.0 秒(AA)。
SpeedRunBench:Inkling-2-Small 以 1,988 帧通关 SuperTux 第 1-1 关,领先于 Opus 5 和 GPT-5.6 Sol(SpeedRunBench)。
科学模型与开发者工具
Gemini 4 Carbon:DeepMind 表示,Carbon 驱动的 AlphaProof Code 智能体验证了 Signal 的密码学代码。它还发现了湿实验室中的错误,并找到多出 40 倍的相关蛋白质序列(详情)。
科学访问:Google 通过 Antigravity 将 Carbon 的免费使用权限扩大到 10,000 名科学家(项目)。
上下文窗口:Antigravity 将 Argon、Carbon 和 Hydrogen 的上下文窗口扩大了一倍,使每个用户轮次的压缩减少 3.5 倍(Mohan)。
用户反馈(个案):一位研究人员称,Carbon 解决了一个此前多轮 Astra 和 Fable 尝试都未能解决的问题(个案)。
OpenAI 面向科学与健康:OpenAI 将向最多 20 万名科学家开放免费的前沿模型访问权限,申请自 11 月 23 日起开放(项目)。
生产力研究:OpenAI 在一项覆盖 1,800 名参与者的研究中报告称,科学家的生产力大约提升了一倍(研究)。
原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文([AINews] Perplexity hits a $1B revenue run rate, up 4x in 2026)
Perplexity crossed a $1B annual revenue run rate today, CEO Aravind Srinivas told Semafor at a live event. The Information reported previously under $250M at the start of the year and over $750M by August, with Computer driving some of the growth. At usual agent lab multiples (eg Cognition also crossed $1B, with Cursor crossing $4.6B), this would roughly mean that Perplexity is now worth $30-50B — if the unit economics hold up. This is a nice bump on the $20B round in September 2025, and qualifies for our “>$10B fundraise” rule for title stories.
Gathering up some recent Perplexity launches for those catching up:
Computer. September’s Automations let it pick up recurring tasks where it left off. November’s Teammates take on ongoing jobs. Computer also runs fully locally on supported Macs.
Monitor API. Watch a topic or page and get a webhook when something relevant changes.
Frontier Research. The new research lab has access to 10,000+ GPUs and plans to publish papers, data, models, and code.
Sponsored by Deepgram
Deepgram is the voice AI infrastructure behind natural conversation—bringing speech-to-text, text-to-speech, voice-focused LLMs, and agents to teams building with voice AI. For any stack and any tools you already love building with, Deepgram slots in as your foundational voice AI technology. So you have control. And we have the context to listen, understand, and speak like a human.
Qwen4-Max in the wild
Friday’s Qwen4-Max preview is now a shipping model. At the highest reasoning setting, Qwen4-Max scored 43 on the Intelligence Index, up from 34. Average cost fell from $10.02 to $3.27 per task. Alibaba cut output pricing from $6 to $4.50 per million tokens, and the model generated about 39% fewer of them. Kimi K3.1 Pro cost $3.68 per index task and scored 67% on Terminal-Bench 4.0 to Qwen’s 48%. Artificial Analysis
Inside the model. Alibaba kept the total parameter count at 2.4T and reduced the number active per token by 25% to 71B. The model takes images and text, with a 1M-token context window. Smaller open-weight Qwen4 models are still to come. Alibaba
Arena and Vals: Arena ranks it #14 overall on Agent Arena, the top Chinese model, with +7.5% net over 14.7k sessions (Arena).
WebDev: It placed #6 in Code Arena WebDev at 1750, +74 over its predecessor (WebDev). It also placed #6 in Image-to-WebDev, just above Claude Opus 5 and GPT-6 Sol (I2WebDev).
Vals Index: It is #11 overall, with large gains in coding and finance (Vals).
Legal-benchmark caveat: On HLAB it gets 95.7% of checklist items right but passes only 21.3% of tasks, and its output tokens rise 3.6x (HLAB).
Vision results (independent): Roboflow calls it underwhelming on its VLM benchmark, with Google, OpenAI and Anthropic now far ahead. At high reasoning effort, however, it is ~4.5x faster and ~68% cheaper per image than Qwen3.8-Max (Roboflow).
Safety, alignment and eval integrity
OpenAI embedded assessments: OpenAI is giving Apollo, METR, Transluce and Goodfire internal access to its models and systems (OpenAI).
Scope by group: Apollo places six evaluators onsite for four months (Rundown). METR reviews OpenAI's largest RL run, with access to transcripts, monitor records and checkpoint snapshots (METR). Transluce runs a six-week audit of post-training alignment evals, with a fuller follow-up in early 2027 (Transluce). Goodfire embeds interpretability researchers to monitor training runs (Goodfire).
Limits: METR says its scope excludes the misalignment incidents OpenAI has disclosed and is narrower than METR proposed (scope).
Criticism of the access terms: Critics say the deals fall short of Sam Altman's promise of "employee-like access."
Reported gaps: Assessors get no Slack access, OpenAI selects which staff they interview, and they cannot access weights or run training-based elicitation (Neel Nanda, Alex Mallen).
Former-employee account: Mikita Balesni says assessors cannot quote employees, even anonymously, and some may be barred from publishing on internal-only models (balesni).
Goodfire acquisition talks: The Information reports OpenAI explored buying the interpretability startup for more than $5B. Goodfire reportedly wanted to stay independent so it could sell to multiple labs (report).
Other integrity signals:
Benchmark contamination: Claude appears to have memorized details of the LinuxArena control eval (Karvonen).
Incident taxonomy: OpenAI published a v0.1 framework for classifying AI control incidents (framework).
Disclosure policy: Anthropic is cutting its coordinated disclosure window from 90 to 30 days for high- and critical-severity vulnerabilities (policy).
Inference systems and agent infrastructure
Together moves on from hand-tuned megakernels. Its first design made it into production for one customer. Agents now search for smaller groups of GPU operations to fuse. Together reports 28% higher throughput on Llama 3.2 1B on B200. Engineering writeup
DeepSeek V4.1 Flash on Vera Rubin: vLLM with Dynamo on SemiAnalysis's AgentX benchmark shows 63.4x per-chip throughput over GB300 NVL72 at 400 tok/s/user, and 93.3x at 600 (vLLM).
Batch-1 latency: SGLang-adjacent work reaches ~2,500 tok/s at batch size 1 on TP4, 1.7x its GB300 result (saltyph). Baseten reports a similar 1.7x over its 4xB300 numbers (Baseten).
Agent-driven optimization: An agent shared Engram tables across HBM via NVLink symmetric memory to recover prefix-cache capacity (Liu). Another team says agents found optimizations that made its B300 step time 10% faster (Halex).
Agentic serving research: MIT's Evolving Infra replaces fixed prefill/decode disaggregation with online adaptation. Against NVIDIA Dynamo on up to 256 GB300s, it reports up to 4.1x lower latency and 14x lower P95 TTFT (Evolving Infra).
Sandboxes and training:
External sandboxes: Simon Willison describes running agents on a Mac while executing their code in WebAssembly, QEMU or Apple Container sandboxes. The pattern includes allow-list proxies and credential replacement (post).
MLPerf: In MLPerf Training v6.1, a GB300 NVL72 trained GPT-OSS-20B in 19.7 minutes, 2.86x faster than H100 (Lambda).
One endpoint for open models. Hugging Face Inference Providers now runs on OpenRouter using the same endpoint and token. Hugging Face
Research, benchmarks and evals
PyBot: A swarm of 1,000 DeepSeek-V4.1-Flash agents spent a week writing a rule-based chatbot with no neural network (lewtun).
Scale: The agents exchanged ~175k messages and merged ~8k PRs.
Result: The bot is 630k lines of Python with zero dependencies and is reported on par with SmolLM-360M on everyday questions.
Self-improvement and training:
AutoRL: Agents supervising a live RL run lifted Qwen3 on AIME 2024 from 46.67% to 66.67% (omarsar0).
Benchmark-gaming caveat: A roundup warns that many self-improving agent gains come from learning the benchmark rather than the task (TuringPost).
Weight-space programming: Opus 5.5 hand-wrote transformer weights that solved algorithmic tasks with a median of 166 parameters (dair_ai).
Data curation: DatologyAI published the tech report behind its claimed 6x pretraining compute multiplier (Datology).
Distillation reproduction: A small-scale replication confirms perturbed-input distillation helps when the student trains from scratch, but finds it hurts when distilling into a pretrained student (Maini).
New benchmarks:
GradeBench: The best frontier model agrees with instructors only 57% of the time when grading handwritten physics answers (GradeBench).
Time to First Audio: AA's new TTS latency benchmark finds a median of 376 ms for streaming models versus 2.0 s for non-streaming ones (AA).
SpeedRunBench: Inkling-2-Small cleared SuperTux 1-1 in 1,988 frames, ahead of Opus 5 and GPT-5.6 Sol (SpeedRunBench).
Science models and developer tools
Gemini 4 Carbon: DeepMind says Carbon powered AlphaProof Code agents that verified Signal's cryptographic code. It also spotted wet-lab errors and found 40x more related protein sequences (details).
Science access: Google expanded free Carbon access for 10,000 scientists through Antigravity (program).
Context window: Antigravity doubled the context window for Argon, Carbon and Hydrogen, cutting compaction 3.5x per user turn (Mohan).
User reports (anecdotal): One researcher says Carbon solved a problem that waves of Astra and Fable attempts had not (anecdote).
OpenAI for science and health: OpenAI is opening free frontier access to up to 200,000 scientists, with applications from Nov 23 (program).
Productivity study: OpenAI reports scientists roughly doubled their productivity in a study of 1,800 participants (study).
Physics showcase: Physicist Nathaniel Craig used GPT-6 Astra to develop a candidate lattice definition of the Standard Model (blog).
Healthcare: ChatGPT for Healthcare adds computer-use agents and more than 20 plugins (healthcare).
Claude Code: The desktop app now runs iPhone Duo in the iOS simulator and can fold and unfold the device to test layouts (ClaudeDevs).
Verification: The team published how it uses cheap tests, benchmarks and end-to-end runs to let Claude verify its own work (blog).
Unreleased model (unconfirmed): A "Silicon 2" entry spotted in Devin is suspected to be an Anthropic model (rumor).
Liquid d1.2-ultra: It ranks #1 on AA's Decision Index at $0.042 per M input tokens (AA).
Sage S2.1 Flash: The open-weight science model reports leading MLE-bench Lite and OfficeQA Pro results (Sage).
FLUX-mimic-x: A 7B video-action robotics model ranks #1 on RoboLab among real-time models (mimic).
Unsloth computer use: Local models can now drive desktop apps and game engines through Unsloth (Unsloth).
Industry and policy
Nvidia. The US Commerce and Justice departments issued a joint statement on the Megaspeed investigation. Commerce
Financing and M&A:
Noeri: Yejin Choi and Carlos Guestrin's continual-learning lab raised $110M (Choi).
Corma: The cyber-defense model company raised a $125M Series A at a $1B valuation (Corma).
World Labs: The team is joining AMD (Fei-Fei Li).
Market data: Anthropic's S-1 shows $16.3B revenue in H1 2026 and compute costs falling from 251% to 48% of revenue (The Batch).
Enterprise adoption: Databricks reports open-model token share rose from 7.3% to 30.9% in 3.5 months (summary).
Local models and hardware
What fits in 16 GB? Unsloth expects a smaller Qwen4-27B to run in 16 GB. Local users want to see the quantization and how much memory is left for the context window. One developer already splits coding jobs between Opus for specs and a local Qwen3.8-27B for implementation. Memory discussion · Local workflows
Hy4 Flash. Tencent’s multimodal model has 224B parameters and activates 12B per token. At $0.070 per million input tokens and $0.281 per million output tokens, it gives developers a cheaper way to try Hy4’s 1M context window. Tencent’s release · Discussion
EUROPA. A reported test run on the JUPITER supercomputer trained a roughly 952B-parameter mixture of experts with 32.8B parameters active per token. The local-model community’s question is why the team chose that scale for its first model. Training run and discussion
Where do two quantizations diverge? inside-llm loads two GGUFs with the same tokenizer and lets you step through their predictions side by side. Watch the first token where they disagree, then compare the probabilities they assigned to the alternatives. Project · Demo and discussion
Basalt’s 770 tokens/sec headline depends on what you generate. The developer reports up to 776 tokens/sec for structured output from Qwen3.8-27B, 262 in chat and 194 for prose. Readers asked for longer quality tests of the 4-bit prefill. Basalt · Benchmark thread
An FPGA running Qwen 27B. Valley Semiconductor reports 7.9 tokens/sec on short prompts and about 5.5 in live chat. Its 50 tokens/sec target at 20 W is a projection for a future chip. Live demo · Hardware discussion
Agents in use
Claude Meetings. A CTO describes inviting Claude into a 30-minute Google Meet. It updated project records in Notion and called the team’s own tools to update bids during the meeting. Field report
A Minecraft test for Hydrogen. A developer asked Gemini 4 Hydrogen to recover missing source and build a datapack for 11 Minecraft versions. The agent reportedly launched two graphical clients to test fixes for dropped items and raycasts. Developer’s account
What teams hire a harness developer to do. One example from the discussion: a coding agent can propose a change, while the harness decides which repository it may write to, which CI test counts as passing and when a person needs to approve the next step. Claude Code discussion
Vision and video
Testing a video model’s physics. After a second MiniMax H3 demo, commenters proposed a harder test: describe the scene, leave the outcome out of the prompt and see whether the model predicts what the objects will do. Demo and discussion
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 Latent Space([AINews] Perplexity hits a $1B revenue run rate, up 4x in 2026),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。