Prime Intellect

The next wave of AI will not be won by better prompts. It will be won by systems that learn from experience. Today, Prime Intellect Lab is out of beta, open for you to start training your own models. The era of self-improving agents is here.

1/n For browser agents, a major bottleneck in evaluation is truthful scoring on the live web. A task is only as good as your ability to confirm the agent actually did it, on a real site whose state keeps moving and that the agent can potentially misreport. So we took matters into our own hands. Today, we're releasing Ecom Bench on @PrimeIntellect: 40 shopping tasks on real Shopify storefronts, each run in a live @browserbase browser and graded by a deterministic verifier. vibrantlabs.com/research/ecom-…

@VibrantLabsAI 🔥

this is a good one

prime-rl is fully open source, and we're hiring systems engineers to take it further. Read the full prime-rl performance deep dive: primeintellect.ai/blog/rl-at-1t-…

Huge thanks to the @vllm_project team, and @robertshaw21 in particular, for all the help along the way. Also to the llm-d and Dynamo teams for the collaboration on routing and inference.

The trainer is 3D-parallel (FSDP2 + CP + EP), built on TorchTitan. FSDP2 shards params, grads & optimizer state. EP keeps experts sharded and routes tokens with all2all instead of all-gathering ~80GB per layer. CP handles the 131k context and GLM-5's DSA attention.

Over a long run the trainer and inference policies slowly drift apart, and that mismatch can kill your training. R3 (router replay) captures the routing decisions from the inference engine, replays them on the trainer - KL mismatch drops ~10x.

One Mooncake store pools KV cache across all nodes, so any worker can reuse any prefix. The router picks workers by a score over load, queue depth, KV usage and prefix overlap. You get cross-replica cache hits with balanced routing across the whole deployment.

We disaggregate prefill and decode onto separate workers. A long prefill used to stall decode for everyone. Now it doesn't.

In RL, inference is the bottleneck — we optimize for throughput, not latency. High concurrency, FP8 precision, and wide expert parallelism over 32+ GPUs. Every GPU holds its own slice of experts and acts as its own endpoint.

Today we're releasing prime-rl v0.6.0 — enabling RL at trillion-parameter MoE scale on agentic workloads at the highest efficiency. We've relentlessly optimized our RL infra. The result: GLM-5 on agentic SWE tasks at 131k context and sub-5-minute step time.

awesome post by @kimbochen covering RL systems end-to-end, including a SWE training run on GLM-5 using our prime-rl framework.

nice blog by @kimbochen about the current RL ecosystem, goes into detail about the different settings and tradeoffs to consider when RLing open models

Excited to support this epic Inference-time compute hackathon with Prime Intellect credits for post-training + compute 24hr hack on > Agents: Multi-step systems that take a goal and execute. Tool use, planning, long horizons. > Real-Time and Interactive: Sub-second loops, live multimodal. > ​RL + Applied AI: Systems that judge capability of a person or a model. Autograders, preference ranking, rubrics, verified-skill signals, human-in-the-loop. luma.com/hncudfxb

@_xjdr @garybasin @nisten 🙏

@garybasin @nisten @PrimeIntellect for most of them now. they are the absolute best!

Ok i did it . this is Noumena's version of your favorite, ai code TUI . I rebuilt the REPL engine to be event based added a new rust markdown rendering stack that considerably increased performance and stability i added a rust OpenAI compat websocket client i added a handful of tools and exposed a handfull of others that were hidden by default this system is optimized for our internal inference system and models but i have recently tested it with a vanilla kimi k2.7 and it works quite well. over the next few days we will make the Noumena Inference Engine available and serve at least the k2.7 model . this project will support native noumena code (ncode) login and keys when its available. either way, i hope y'all enjoy and find this useful. this is the first of several things i hope to launch this week github.com/Noumena-Networ…

We are so back! Future looking bright to post-train, serve, and continuously improve your own model on top of models like GLM-5.2 using primeintellect.ai 🫡

@SemiAnalysis_ 👀

RL Systems Mind the Gap: Matching Trainer and Generator Throughput RL Training Infrastructure, GRPO, PipelineRL, Async RL, Policy Staleness, RL Sandbox Infra, CPU Requirements, TCO Analysis, Thinking Machines Tinker newsletter.semianalysis.com/p/rl-systems-m…

Great RL systems deep dive by @SemiAnalysis_ Scaling RL is as much of an infra problem as an algorithm one SemiAnalysis ran experiments on our stack: Prime RL + Sandboxes. System efficiency is ultimately queue health to match generator and trainer throughput

been beating this drum since early 2025, seems like people are starting to see why it's so important :) RL works -> "train or get trained on" -> open models + post-training infra are the path to institutional flywheels + democratization of AI progress

Satya is perfectly describing the why and what behind @primeintellect since 2023 🫡 > AI needs to be open & sovereign > Let every company create its own self-improving agents: and own their loop to make them better > A rich open ai ecosystem creates far more abundance than a future locked down by a few closed labs > Every company is becoming an ai company: so every company needs to own its own product <> model improvement loop @primeintellect enables this today: > Your own evals + rl envs for the outcomes you care about > models self-improving in production from your real traces > don't cede your moat to a handful of labs. This self-improvement loop is the IP and it compounds Open self improving agents for everyone 🫡