Prime Intellect debuts INTELLECT-3, an RL-trained 106B parameter open source MOE model it claims outperforms larger models across math, code, science, reasoning
Today, we release INTELLECT-3, a 100B+ parameter Mixture-of-Experts model trained on our RL stack, achieving state …
Prime Intellect
Context & Ripple Effects
INTELLECT-3 extends Prime Intellect’s position from tooling and compute for companies building agents to releasing a model layer of its own, a direction reflected in its later funding for its agent-building compute and tools platform.
It joins a growing set of open-weight reasoning and MoE releases, including Arcee’s commercially customizable reasoning model and Thinking Machines Lab’s broad open-weight MoE model. The differentiator here is the combination of a 106B-parameter MoE architecture, reinforcement-learning training, and claimed results against larger models.
First-order effects
Developers can download and evaluate a new open-source MoE model aimed at math, code, science, and reasoning workloads rather than relying solely on larger proprietary or open models.
Prime Intellect gains a model artifact to pair with its compute and agent-building offering; its performance claims now become a key test for users and independent evaluators.
Second-order effects
Open availability gives teams another candidate for customization and self-hosting, increasing pressure on competing open-model providers to substantiate reasoning performance and licensing flexibility.
If the reported results hold in outside testing, buyers may compare models more on task performance and active inference requirements than on total parameter counts alone.
Third-order effects
The release adds to an open-model market in which reinforcement-learning post-training and MoE design are becoming practical routes to challenge larger dense or broadly trained models.
As credible alternatives proliferate, model buyers could gain leverage over providers, though that depends on reproducible evaluations, deployment costs, and the quality of available tooling.
The trend: Open-weight AI competition is shifting from raw model size toward deployable reasoning performance, efficient architectures, and the ability to customize models for specific workloads.
Very nice: Prime Intellect unveiled INTELLECT-3, a 106B-parameter Mixture-of-Experts model trained with large-scale asynchronous RL on top of GLM-4.5-Air - and they open-sourced the whole recipe: weights, RL framework, environments, and evals. It posts state-of-the-art results [i…
there's something special about releasing a SOTA model trained on a fully open framework every detail, environment, dataset and algorithm for us to see, replicate & play with is insane!!
Expected better from PI After all the hype, shipping a finetune of an existing base model where eval results are all over the place & you still barely eek them out Compare 🇨🇳 open source labs putting out imo gold/frontier models (with much less fan fare)
Cool model, but... >looks inside >GLM 4.5 Air Base Continued Chinese victory. To PI's credit, they disclose it quickly and don't imply they trained it from scratch. So this is not a dunk, just an observation.
Benching is hard Their GPT-OSS scores are off by up to 10 pts (GPQA-D), their GPQA-D impl affected scores by 2-3 pts vs reference impl They did not use R1 for the benches but 3.1 (or 3.2) depending on the timing. MMLU-Pro is weird (-10 vs AA), but they used DS API, so idk
INTELLECT-3 is live We hope it shows that you don't need to be from the big labs to compete in the big leagues. We believe in a future where everyone can optimize their own models. Where every company can be an AI company. We're building the stack to enable this future.
INTELLECT-3 is our first model I can use daily It's build using our open source stack by scaling RL of MoE over 512 H200 and pushing the sota at its size Incredible proud of leading such a team of talented, dedicated and hard working individual collaborating together to push
I've always found Prime Intellect impressive, so it is especially notable to me that this uses a Chinese LLM as the base model (GLM 4.5). I mean no alarmism or jingoism. But America could, and should, do much better in open-source than it currently does.
working on this model, with this team, with all of the infrastructure we built to get here, is the most rewarding thing i've ever been a part of still can't believe i get to wake up every day and work on this stuff with these people and put all the code for free on the internet […
RL at 100B+ scale isn't just for hyperscalers anymore. Prime Intellect's INTELLECT-3 is a 106B MoE model trained end-to-end with large-scale reinforcement learning on a 512×H200 cluster, and they're open-sourcing the entire post-training stack that made it possible. For
State of the art model, fully open source, and trained using Prime's unique asynchronous RL framework PRIME-RL. Respect. Prime might be the only serious crypto x AI project in history (and no I do not consider Bittensor serious). Huge win for both sides of the aisle.
That's cool - naming your company after an AI which basically destroys humanity. For /r peterexplainsthejoke It comes from a cult sci-fi novel “The Metamorphosis of Prime Intellect” where “Prime Intellect” is an ASI that saves humanity by destroying everything that makes being
we've scaled RL for a 100B+ MoE model achieving SOTA benchmark results for its size more important than the final model checkpoint is making the frontier infra required to train models like this accessible to everyone details on the full training recipe, our open source [image]
Claim from the abstract: “106B-parameter MoE (12B active) trained with large-scale reinforcement learning on our end-to-end RL infrastructure stack.” I expected all RL from scratch. Reality: Already existing base model + SFT + RL 😿
We are releasing INTELLECT-3: Scaling async RL to 100B+ MoE on our end to end stack SOTA for its size in math, code, reasoning, science - outpacing bigger models Fully open source: weights, data, frameworks, envs, evals Trained on the same stack we're opening up to everyone