Perplexity hits $1B revenue run rate, up 4x in 2026; Qwen4-Max ships with mixed evals

Latent Space · rss · 2026-11-17

Latent Space's AI News leads with Perplexity crossing a $1B annual revenue run rate (up from <$250M in January), implying a $30-50B valuation at agent-lab multiples, plus launches like Computer running locally and a Frontier Research lab with 10,000+ GPUs.

Models: Qwen4-Max ships with 2.4T total / 71B active params and 1M context, scoring 43 on the Intelligence Index (up from 34) at $3.27/task, topping Chinese models on Agent Arena — though it passes only 21.3% of HLAB legal tasks and underwhelms on Roboflow's vision benchmark. Kimi K3.1 Pro scores 67% on Terminal-Bench 4.0.

Safety: OpenAI grants Apollo, METR, Transluce and Goodfire internal access, but critics note no Slack, no weights access, and OpenAI-chosen interviews; The Information reports OpenAI explored buying Goodfire for $5B+. Anthropic cuts disclosure windows from 90 to 30 days.

Infra: Together's agent-fused kernels lift Llama 3.2 1B throughput 28% on B200; DeepSeek V4.1 Flash hits 63-93x per-chip throughput gains on Vera Rubin; MIT's Evolving Infra cuts latency up to 4.1x. Ollama found a single timestamp change wipes prefix caches; GB300 NVL72 trains GPT-OSS-20B in 19.7 minutes (MLPerf).

Research: 1,000 DeepSeek agents wrote a 630k-line dependency-free chatbot in a week (175k messages, 8k PRs); AutoRL lifts Qwen3 AIME 2024 from 46.67% to 66.67%, with caveats about benchmark-gaming; GradeBench finds the best frontier model agrees with instructors only 57% of the time grading handwritten physics.

Original post →

More from Venture

Venture channel →