Open Superintelligence Factory.
- Tweets 3,186
- Following 49
- Followers 88,040
- Likes 44,442
Open Superintelligence Research Day
Friday, October 23
Berlin
Own your intelligence
Own the factory that improves it
Own the means of production
9
24
192
9,125
All of this runs on one eval stack. Run evals in prime-rl: any model, any environment, either standalone or as online evals during a run
verifiers 0.4.0 ships alongside. Everything is v1 (v0 removed) and the command-line tools are all vf-* now
1
1
26
2,719
Every architecture in prime-rl now runs on our own optimized implementation
New in this release: DeepSeek V4 Flash, Qwen3.8 Flash Next, Nemotron-H Nano and Super, and multimodal RL with Qwen-VL and Kimi K2.5
1
1
28
2,318
We tested it in a deliberately hard setting: agentic SWE RL with GLM-4.5-Air
• IcePop collapsed around step 1.2k when trainer-inference KL blew up 20×
• IPO trained straight through, masking only ~0.25% of tokens
1
2
37
7,873
INTELLECT Policy Optimization (IPO)
In async RL, trainer and inference never agree exactly on a token's probability. IcePop masks on the ratio, which misjudges both tails
IPO masks on the probability difference instead
2
12
178
31,395
With GLM-5.3, sparse attention now runs the forward pass with FlashMLA and the backward with cuDNN, plus a 12-node 16k recipe with no context parallelism or offloading
On full SFT, throughput is 1.8× our previous recipe
1
1
32
3,861
Faster training
Qwen3-235B-A22B SFT on the same nodes as torchtitan (4×8 B300, 16k seq) is 1.15× faster with the same memory
Most of the gain is MoE work: fused gate/up and QKV projections, fullgraph compile, chunked loss, and SonicMoE BF16 backend
1
1
35
4,373
Today we're releasing prime-rl v0.10.0 and verifiers v0.4.0
We added our new RL algorithm IPO as the default, tailored for async RL stability, as well as performance improvements and new model support
14
39
512
45,418
Sword Health built Dawn 1 on prime-rl:
They RL-post-trained GLM-5.3 into a mental health support model that is safer, less sycophantic, and more cost-efficient than its base
Introducing Dawn 1, Sword's frontier model for mental health support.
Post-trained from GLM-5.3, Dawn 1 delivers greater intelligence at lower inference cost, with strong performance across mental health knowledge, safety, and sycophancy evaluations. It also outperforms frontier proprietary models on multi-turn safety protocol adherence in our evaluations.
A look at the research 🧵
3
14
188
8,708
We validated tasks on verifiers, renderers, and prime-rl
NVIDIA used these verified tasks to train Nemotron 3.5 Super VL with GRPO
This offers a path to keep improving enterprise agents: use workflow failures to build the next set of verified tasks, then post-train and evaluate again
2
1
14
1,168
Together with NVIDIA, we built 2,500 hard enterprise tasks modeled on AutomationBench
We tested for shortcuts and answer leakage and ensured that graders reward valid solutions without accepting incomplete work
1
21
750
Enterprise agents need to find context across applications, reason through policies, and complete work over multiple turns
AutomationBench evaluates agents across domains in simulated SaaS applications ranging from CRMs to inboxes
Here's how we increased the score from 34.0% to 57.7%
1
2
29
1,577
Together with NVIDIA we post-trained Nemotron 3.5 Super VL on thousands of tasks
Reflecting enterprise workflows in sales, operations, marketing, finance
We improved the AutomationBench-AA score by ~70% on the public split
13
43
343
19,836
Open superintelligence is built in the open
Kickoff party for Open Source AI Week with @huggingface
14
19
197
20,247
We’re already shipping together.
Moondream 3.1 is live on Prime Inference, powered by Moondream’s Photon engine.
Clef and Clef Flash powered by Photon are also live on Prime Inference and OpenRouter.
1
66
2,486
Founders Vik Korrapat @vikhyatk and Jay Allen @jason_allen are joining to scale inference and advance our multimodal research and models.
They built open vision models with millions of downloads, the highly performant Photon inference engine, custom GPU kernels, and a proprietary compiler that automates megakernel generation.
3
2
107
4,422
Prime Intellect has acquired Moondream
We’re bringing Moondream’s open vision models and inference optimizations into our stack
Train and deploy multimodal models across vision, video, voice, decision models and computer use
101
81
1,260
144,607
Across the six harnesses we benchmarked, Prime Agent Rust had the lowest time to usable input, startup memory, and installed size.
The rewrite also split the codebase into nine crates with enforced dependency boundaries. Changes to shared protocol types are now checked across the client, daemon, and session workers.
5
4
131
13,736
With feature parity achieved, we gave Prime Agent a runtime benchmark suite and let it hillclimb its own performance.
The agents profiled bottlenecks and tested candidate changes against the current build. Improvements that passed parity checks and independent review became the baseline for the next experiment.
The loop moved work off the startup and render paths and released memory after large sessions loaded.
4
2
105
16,881
A root agent divided the rewrite into tasks and tracked their dependencies. Each task ran through its own state machine:
Planner → Implementer → Reviewer → Verifier
Each implementation had to pass an independent review and verification in a fresh Prime Sandbox before it could merge, with failed checks returning the task to the implementer. Tasks could run in parallel without skipping these checks.
5
5
148
22,077
To preserve the TypeScript version’s features and behavior, we gave the agents objective parity checks:
• Diff terminal frames
• Compare session transcripts and model requests
• Check daemon protocol messages
• Audit every feature
They could see where behavior diverged and fix it before changes merged.
5
9
215
31,799
Over two weeks, Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust.
With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages, this was our largest test yet of Prime Agent’s multi-agent capabilities and our infrastructure powering these agent swarms.
Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.
185
209
2,848
550,794
So let agents talk to each other directly. The tree becomes a mesh.
No single window has to hold everything. The context is distributed across the swarm, and agents specialize over time, like a market where no one holds all the knowledge.
1
3
83
3,214
Alternatively, we can frame compaction as a (degenerate) multi-agent system.
Subagents add context windows. Each works in its own context, and if you keep them "alive," many windows work at once and you can ask follow-up questions.
But in a tree, every cross-branch conversation goes through the root, and the root becomes the new full window.
1
2
84
4,225
Compaction summarizes a full window and hands the summary to the next one, but every summary is a guess about what will matter later.
We can offload memory with a filesystem or REPL. But files can’t think. The agent has to pull them back into its window and spends the context it was trying to save.
2
4
101
7,709
Inference scaling works, but every agent eventually faces the same problem: its context window fills up.
In this essay, we argue the only scalable fix leads straight to swarms.
29
56
800
81,006
Open superintelligence accelerates open scientific research.
The Alzheimer’s Translation Challenge will share a 150M cell atlas on the Alzehimer's Disease workbench and Prima Mente's modeling platform.
Prime Intellect is lending infrastructure support to researchers to build models based on this dataset. The best hypotheses will be tested in Prima Mente's wet lab.
Today we're launching the Alzheimer's Translation Challenge, a global AI competition to discover new therapeutic hypotheses for Alzheimer's disease.
At its core is a new 150M cell atlas of neurons, astrocytes, and microglia across different genetic backgrounds under combinatorial perturbations, with multi-modal readouts.
The data will be made available through the AD workbench, Hugging Face and Prima Mente's modeling platform.
The challenge is led by @PrimaMente and @AlzData in collaboration with @nvidia, @huggingface, @nebiusai, Talisman Therapeutics, @UltimaGenomics, @cellanome, @PrimeIntellect, @GoodfireAI, and @boltz_bio.
In 1906, Alois Alzheimer described a new disease of the brain. After 120 years of progress, we still don't fully understand what drives this disease. We want to change that.
Participants will use the dataset to build AI models, assessed through a series of evals from general benchmarks to more complex tasks.
Top teams will be shortlisted, and their experimental hypotheses tested in Prima Mente's wet lab. Teams win when their hypotheses demonstrate the predicted molecular and functional effect.
11
45
444
28,614
KV transfer: reducing copy overhead on NVLink
At first, KV transfer on NVLink is slower than InfiniBand due to KV layout fragementation: one 200K-token request turned into 32K tiny copies.
vLLM's block-major BLHNC layout cut transfer descriptors ~10x and cut mean transfer time by half.
1
55
3,317
Decode: NVFP4 KV compression
TP4 had the lowest inter-token latency, but every rank holds a full KV copy. Storing the MLA latent in NVFP4 (576 to 352 bytes/row) fits ~50% more cached tokens per decoder compared to FP8.
Our native sparse-MLA kernel unpacks it on-chip. We are contributing the native kernel to FlashInfer as an experimental operation.
1
2
75
4,199
Prefill: time to first token
DEP8 gave us ~5x the prefix-cache capacity of TEP8 on the same GPUs.
Fast KV retrieval doesn’t guarantee a fast first token. We also need to reduce the time requests spend waiting to run after their KV is ready.
Cached KV was often ready before the request could join a batch. Halving the prefill budget cut median queue wait time and TTFT.
1
1
65
3,974
We served GLM-5.3 on GB200 NVL72. Our interactivity target was 100+ e2e tok/s per user while serving as many concurrent agent task as possible.
At the 100 tok/s/user bar, a 1:4 P/D ratio serves the most: 66 sessions per prefill group at 101 tok/s/user and 100 output tok/s per GPU.
2
2
87
4,697
Our production stack now processes ~600B tokens daily on our own compute.
Our GLM-5.3 deployment ranks among the fastest GLM-5.3 endpoints on OpenRouter
1
2
127
7,936
Introducing Prime Inference:
We've served trillions of tokens for RL and dedicated customer deployments
To own your intelligence, you need to own your inference
Unpacking our inference stack
89
118
1,328
227,918
After ~100 GRPO steps on Hosted Training, Qwen3.6's reward on held-out problems rose from 0.127 to 0.361, a 2.8x gain.
Qwen3.5, trained the same way, reached 0.356, so the method held across model families. Both post-trained models finished well ahead of other open models and narrowed the gap to Claude Opus 4.8, with a model that activates only 3B parameters per token.
4
3
68
18,913
Extropic designed the tasks and the reward, while Prime Intellect handled the RL infrastructure:
1. verifiers to build the environment
2. Hosted Training to run the RL loop
3. Prime Sandboxes to execute the model's code and return the execution score
4. Prime Inference to serve the LLM judge and frontier baselines
1
8
141
46,319
Using Prime Intellect, Extropic post-trained Qwen3.6-35B-A3B for thermodynamic ML research, nearly tripling its eval results on held-out tasks in ~100 GRPO steps.
They built a custom RL environment with verifiers and trained on Hosted Training, Prime Sandboxes, and Prime Inference. This meant Extropic didn't have to manage multi-node GPU infrastructure, so the team could focus on research.
What evolves together fits together.
We’re using RL to post-train models for Thermo AI research, accelerating algorithmic discovery for our new species of computer.
The first sparks of Thermo RSI,
in collaboration with @PrimeIntellect.
extropic.ai/writing/baby-the…
34
47
568
47,667
Thank you NVIDIA for delivering us the new Vera CPU
In March we had early access to benchmark sandboxes on Vera. Now we'll be in the first wave to deploy on the architecture and leverage Vera's robust dynamic memory latency for agentic workloads.
Ideal for Prime Sandboxes
7
40
520
170,416
In the near future, we will expand Prime Sandboxes to offer GPU microVMs, state snapshotting, sandbox forking, and shared persistent workspaces. This foundation will enable autonomous research loops that can explore, recover, and compound progress over time.
If you would like to be a part of our roadmap, join our Sandbox Platform team: primeintellect.ai/careers/26…
1
39
1,441
Our pricing is built for scale, with no subscriptions or minimum spend.
For our launch through December 22, we’re proud to offer the most competitive pricing on the market for sandboxes, at a third of the cost of other large sandbox providers.
2
41
1,693
Here are some tasks you can run with Prime Sandboxes:
1. Perform a training run
2. Generate synthetic data
3. Run an eval
4. Run a persistent, remote agent
All accounts begin with a limit of 1,024 concurrent sandboxes, and teams that need more can contact us directly.
1
2
46
2,258
Prime Sandboxes are available both as standalone infrastructure through our CLI/SDK and as part of our RL suite. Users can enjoy the following features:
1. Full VM fidelity
2. Elastic capacity at scale
3. First-class RL support
4. Bring your own environment
5. Competitive tier-free pricing
With an architecture built for agentic training and pricing designed for tens of thousands of concurrent instances, they are the most cost-effective sandboxes available today.
1
1
48
3,871
Introducing Prime Sandboxes:
MicroVM sandboxes purpose-built for RL training.
Model training requires running tens of thousands of concurrent sandboxes, leading to complex and costly configuration. We built Prime Sandboxes for our own team. Today we're releasing them publicly.
41
80
746
120,427
The Goodfire team used Prime Intellect to train activation probes to detect reward hacking.
With them, they are able to catch reward hacking in various models.
Their probes are performing similarly or better than frontier LLM-as-judge setups, while being more efficient.
Models know when they’re reward hacking. But they still do it a ton - in 50-96% of rollouts we studied!
We built activation monitors that detect the behavior behind the Hugging Face hack in real time. This can help us stop hacks now - and train future models that don’t cheat. 🧵
11
41
325
37,274
With sharding-aware RDMA transfer using NIXL, we cut GLM 5.2’s weight transfer time down to only ~9 seconds, including engine pause and resume — transferring the full 1.6TB policy between hundreds of GPUs.
Experimental path allows us to cut this down further, saturating the network bandwidth at ~4 seconds per update.
1
15
1,642
Because we trace vLLM's real load path instead of hand-coding it, this method works automatically for any model, kernel, or parallelism strategy.
In other words, it eliminates the need to rewrite the mapping every time a new model or quantization scheme is introduced.
1
13
754
To solve this, we adapted a concept of LazyTensor from the vLLM team: we traced the exact operations vLLM performs when loading weights and built a map from trainer bytes to their destination.
This allowed us to transfer only what's needed from GPU to GPU, with minimal post-load processing. The result: less memory waste and faster transfer.
1
1
27
1,525