Web search in Hermes Agent is now fast and free.
@perplexity_ai built Fast Search for agents, and it is free for all Nous Portal tiers.
We are omnivorous at Shopify and constantly measure all the Search APIs, Perplexity has became the main one on merit. I much prefer C++, of course, but will take Rust, too :-) (compiled, no GC). Interestingly Rust seems to be more in-distribution for LLMs…
Show more
cool, only took $3k, more iterations are running hopefully can surpass
updated Decision Index 0.1 → 0.2 🎯
better formula, +29 jev-like models, +21 benchmarks
AutoJev-27B by @perplexity_ai CTO @denisyarats took the open lead, trailing jev by 0.8 points 🏆
come find the best model at every size, speed and use-case
Show more
we perform post-training via multiple stages: SFT, RL, RFT, and on-policy self-distillation (OPSD). we use RFT and OPSD as the last stage of training to bridge the sim2real gap. RL environments, while very realistic, still lack the full diversity and complexity of the real-world use cases we receive from production.
in this research we describe how OPSD can be quite effective at learning from user feedback in follow-up conversations, as well as from tool call errors. we design a training method that combines RFT (forward KL/CE) and OPSD (reverse KL) to learn from both successful and problematic trajectories. for problematic trajectories, we construct hints from tool call errors or negative user feedback and use self-distillation to alleviate those errors.
in this specific work we post-trained GLM-5.2 with this approach and observed a substantial 21% reduction in tool call errors, as well as positive improvements in user satisfaction (not yet statistically significant, which we are currently addressing). it also made the trajectories more cost-efficient.
another reason i'm excited about this direction is that it seems to be a potential way of doing online/continual learning.
for more details, take a look at the blog:
Show more
New research: We post-trained a Computer model to learn from its own errors using hint-guided self-distillation.
In a live A/B test, a later trained checkpoint reduced tool-call failures by 21.2% relative to an earlier checkpoint.
Show more
fun weekend project: AutoJev.
i was curious to see if i could train a competitive Jev-like model completely autonomously with a swarm of agents using our internal system. turns out you can get quite far!
some details:
- gave the swarm a devbox with an h200 gpu
- the swarm is a mix of astra and fable; used sol and luna for synthetic data and filtering; ran for 20 hours
- spent $3.1k in total ($1.9k on agents, $1.2k on data)
- used qwen3.8-27b as the backbone. it underperforms Jev initially, but after training it is competitive/better, and it also supports 260k context and multimodal input
- the agents tried RL, but it hasn't worked out so far, so they fell back to SFT on high-quality synthetic data
- both accuracy and calibration are better
- jev-compatible api, playground, and inference. on an h200 it gets around 120ms p95. that will be higher under contention, but still quite fast for an unoptimized version
code:
model:
going to let it run longer, and will also bring up inference so folks can try it out.
Show more
consider joining @bo_wangbo in Berlin, he is an exceptional researcher and a great person to work with!
We started to hire RE/RS for AI team now in Berlin for the first time at Perplexity, come and help us mine intelligence from our massive search index:
A new Perplexity Search SDK cookbook is live.
The recipe fans out focused searches, filters results to official docs, extracts relevant passages, and writes a source-linked brief for your coding agent.
Get started:
Show more
we are sharing how we built CobbleDB, our replacement for AWS DynamoDB that powers key-value storage for our web-scale search engine.
we optimized CobbleDB for our exact read and write patterns. result: 5x lower batch read latency and at least 20% lower serving cost (in practice much higher).
what i'm most excited about is how it was built and deployed: 2 human engineers plus our internal system that runs a swarm of hundreds of persistent AI agents in an autonomous loop. the agents reviewed code and infra changes, caught blockers that were easy to miss, prepared fixes, tests, and monitoring, ran the migrations, and did head-to-head comparisons on real traffic to hunt down bugs and inefficiencies. they tracked CI and rollout gates and kept project readouts up to date. production actions stayed explicitly human-owned though.
more details and some insights about our search engine in the blog:
if you want to work on problems like this, DM me or apply at
Show more
We’re publishing research on how we built CobbleDB, our key-value database that serves web content for Perplexity search.
Two engineers and a team of hundreds of proactive, always-on AI agents built the core infrastructure in two months.
Show more
Portable Computer is now available on Windows PCs with @NVIDIA RTX GPUs.
Run the harness, agents, and models locally on your PC.
Work with local files and connected apps without sending tasks to the cloud. Use frontier cloud models when needed.
Show more
we are releasing a new benchmark and a public leaderboard for embedding models that measure the performance of large-scale agentic web search. more info can be found here.
paper:
blog:
leaderboard:
Show more
We're introducing Q2D-Web (Query2Doc-Web), a benchmark and public leaderboard for evaluating retrieval in agentic RAG systems.
Q2D-Web tests how embedding models perform on large-scale web search using agent-reformulated search queries.
Read more:
Show more
it's been great working with @NousResearch to integrate our search api into Hermes Agent. we are making a lot of cool improvements to our search stack to make it the most accurate and cost-efficient.
Perplexity Search API is now available in Hermes Agent.
Search API gives Hermes access to an index of more than 400 billion URLs. It returns real-time results and ranks snippets by relevance.
Show more
What does it take to deliver high-quality results in seconds at massive scale?
@perplexity_ai serves 50M+ queries daily across 15 models, including NVIDIA Nemotron 3 Ultra. That takes high-performance infrastructure and smarter routing without compromising quality, speed, or cost.
That’s why Perplexity has built on NVIDIA generation after generation. With NVIDIA Blackwell, the larger NVLink domain enables models to scale across multiple nodes without sacrificing latency.
Learn more:
Show more
check out our new blog post on how we serve embeddings and rerankers for our SOTA pplx-embed models over an exabyte-scale search index.
embedding models have a similar serving profile to LLMs: batch indexing is compute-bound prefill, online serving is memory-bound decode. this lets us reuse the same optimized kernels we built for large LLMs and get great efficiency for free. on top of that, we optimized the runtime: whole-model CUDA graphs captured lazily as the engine serves, and a LazyTensor in Rust that overlaps CPU scheduling with GPU execution.
result: up to 3x lower p50 and 4.8x lower p99 latency than vLLM on BGE-M3 at 128 tokens, single H200!
if you are interested in working on problems like this, DM me or apply at
Show more
Every answer in Perplexity starts with embedding and ranking models picking the most relevant results for the query.
Today we published research on how we built SoTA serving infrastructure behind those models.
Read the research:
Show more
Perplexity will be at RustConf 2026 next week.
@zbraniecki will present “Inside SPACE, the Rust Sandbox Platform Behind Perplexity Computer” on September 10. Then join us for happy hour from 5:30–7:30 PM.
Register at:
Show more
New internship slot opened at Perplexity Berlin:
we are open-sourcing Lily, lightweight local inference engine for Apple Silicon, optimized directly for the Qwen MoE architecture and implemented on top of custom Metal kernels and a compact Rust runtime that manages both the session state and the generation loop.
we carefully optimized the engine around Qwen's exact architecture and dimensions: matrix/vector kernels for prefill/decode, minimal data movement across expert, attention, and recurrent paths, and tiles and layouts tuned to the live workload.
with these optimizations we achieve 1.23x faster prefill and 1.35x faster decode than MLX-LM on Qwen3.6-35B-A3B Q4!
for more details, check out our blog and code:
code:
blog:
Show more
Today we’re open-sourcing Lily, the local inference engine we built for hybrid compute in Perplexity Computer.
Lily is specialized for Qwen3.6-35B-A3B on Apple silicon, built so on-device compute doesn’t bottleneck Computer tasks.
Read more:
Show more
Introducing hybrid compute in Perplexity Computer.
Computer can start a task in the cloud, then move to a local model that runs on your Mac. Use it for steps involving private files or sensitive data.
Available today in the Perplexity Mac app.
Show more
thanks to the @ArtificialAnlys team for including us in their new search benchmark and running an independent study comparing many search engines!
even more excited to see us on top across all settings and with a decent margin, and this is not even the latest tech, lots of updates are coming soon.
our team has been prioritizing longer-term bets, focusing on substance, and letting the evals speak for themselves rather than putting the focus on marketing and hype. we believe this is what will win long term.
Show more
Perplexity Search debuts on the Artificial Analysis Search Index, with all three context size variants taking top positions on the leaderboard
The @perplexity_ai Search API comes with three context settings (low, medium, and high) that control how much extracted content each search result carries. We tested all three variants using our standardized methodology: the same model (GPT-5.6 Luna at medium reasoning), running inside Stirrup, our open-source agent harness, with tools for searching and fetching pages from the web. Only the provider behind the search tool changes.
Key results:
➤ Perplexity Search (medium) scores 80 on the Artificial Analysis Search Index, ahead of the previous leaders, Parallel (advanced) and Brave Search (LLM context), at 75. The high and low variants score 79 and 77 respectively. Its lead is concentrated in BrowseComp results, with AA-Omniscience and DeepSearchQA scoring comparably to other leading providers
➤ Efficient search payloads: smaller overall search results mean the model reads less per task, so Perplexity has the lowest model inference cost per task of providers we’ve tested so far, ranging from $0.028 to $0.034 across the three variants vs $0.036 for the next lowest provider
➤ Total cost per task is ~$0.091 for the medium and high context variants, at mid-pack latency. For comparison, Parallel (advanced) costs $0.084 per task and Brave (LLM context) costs $0.13 per task
Show more
great working with @DecagonAI team, now their agents are powered by our state of the art search technology!
We’re partnering with @Perplexity_AI to bring live web search to Decagon agents.
Customer questions often depend on information that changes by the hour.
Agents can now search the live web mid conversation, pull in current information, and respond with cited sources.
Show more
Connectors are now available in Agent API.
Connect your agents to GitHub, Slack, Google Drive, and Datadog without passing a server URL or token on every request.
API Group admins can connect each service once for all API Group members.
Show more
