Source-linked AI summary
INTELLECT-3: Technical Report
Prime Intellect Team, Mika Senghaas, Fares Obeid, Sami Jaghouar, William Brown, Jack Min Ong, Daniel Auras, Matej Sirovatka, Jannik Straube, Andrew Baker, Sebastian Müller, Justus Mattern, Manveer Basra, Aiman Ismail, Dominik Scherm, Cooper Miller, Ameen Patel, Simon Kirsten, Mario Sieg, Christian Reetz, Kemal Erdem, Vincent Weisser, Johannes Hagemann
TL;DR
Open-source RL infrastructure still has shortcomings in frameworks, environments and evaluations, and stable training recipes. INTELLECT-3 pairs the GLM-4.5-Air base model with an open end-to-end training stack and prime-rl. It reports strong cross-domain performance, including surpassing frontier open models over 6× larger on reasoning and agentic benchmarks.
Problem
Open-source RL pipelines still have shortcomings in frameworks, environments and evaluations, and stable training recipes.
Source: §1 · p. 4Method
The report trains INTELLECT-3 on GLM-4.5-Air using an open end-to-end stack and prime-rl, with reusable environments and asynchronous training infrastructure.
Source: §1 · p. 4, §5 · p. 19Results
INTELLECT-3 performs strongly across math, code, science, and reasoning, surpassing frontier open models over 6× larger on reasoning and agentic benchmarks.
Source: §1 · p. 4, §5 · p. 19Takeaways & Limitations
The authors aim for the open-sourced model, environments, and complete training framework to narrow the gap between proprietary RL pipelines and what independent researchers, small labs, and companies can build.
Source: §5 · p. 19For sequence-length scaling, the 256k-token approach with Ncp=2 halved data parallelism and degraded accuracy, making it unsuitable for production training.
Source: §2.1.6 · p. 8
Abstract
from arXiv · showhide
We present INTELLECT-3, a 106B-parameter Mixture-of-Experts model (12B active) trained with large-scale reinforcement learning on our end-to-end RL infrastructure stack. INTELLECT-3 achieves state of the art performance for its size across math, code, science and reasoning benchmarks, outperforming many larger frontier models. We open-source the model together with the full infrastructure stack used to create it, including RL frameworks, complete recipe, and a wide collection of environments, built with the verifiers library, for training and evaluation from our Environments Hub community platform. Built for this effort, we introduce prime-rl, an open framework for large-scale asynchronous reinforcement learning, which scales seamlessly from a single node to thousands of GPUs, and is tailored for agentic RL with first-class support for multi-turn interactions and tool use. Using this stack, we run both SFT and RL training on top of the GLM-4.5-Air-Base model, scaling RL training up to 512 H200s with high training efficiency.
1 Introduction
INTELLECT-3 pairs a state-of-the-art model in its weight class with an open end-to-end training recipe and infrastructure stack. The report targets shortcomings in open-source RL infrastructure and reports strong reasoning and coding results.
1 Introduction
- RL with verifiable rewards has emerged as a dominant post-training paradigm, with cited models demonstrating gains for long-context reasoning and agentic tool use.The passage names OpenAI o3, Grok 4, and DeepSeek R1.Source: §1 · p. 4
- Open-source end-to-end RL infrastructure still has shortcomings in training frameworks, environments and evaluations, and stable training recipes.Existing frameworks can be complex and monolithic, making extension difficult and fragmenting ecosystem artifacts [43].Source: §1 · p. 4
- The report presents INTELLECT-3, built on GLM-4.5-Air, alongside a complete recipe spanning supervised fine-tuning through large-scale reinforcement learning.The model was trained end-to-end with prime-rl and the open-source infrastructure components.Source: §1 · p. 4
- 90.8% on AIME 2024 and 88.0% on AIME 2025; INTELLECT-3 also surpasses frontier open models over 6× larger on reasoning and agentic benchmarks.It scores 69.3% on LiveCodeBench v6, outperforming Z.ai’s GLM-4.5-Air post-train by 8%.Source: §1 · p. 4
- The open-source release includes INTELLECT-3, prime-rl, and environments used for synthetic data generation, training, and evaluation.The framework supports OpenAI-compatible asynchronous inference, multi-turn agentic RL, multi-node FSDP2 training, and vLLM inference.Source: §1 · p. 4
2 Training Infrastructure
The training infrastructure comprises an asynchronous RL framework, a shared environment interface and ecosystem, and secure code execution for agentic coding environments.
2 Training Infrastructure
- prime-rl is an asynchronous RL framework for large-scale supervised fine-tuning and reinforcement learning of Mixture-of-Experts models.The passage introduces it as one of the end-to-end training infrastructure components.Source: §2 · p. 4
- Verifiers and the Environments Hub provide a unified environment interface and ecosystem for agentic reinforcement learning.They are presented as a component of the training infrastructure.Source: §2 · pp. 4–5
- Prime Sandboxes provide high-throughput, secure code execution for agentic coding environments.They are presented alongside the framework and environment ecosystem as an infrastructure component.Source: §2 · pp. 4–5
2.1 prime-rl: A Framework for Asynchronous Reinforcement Learning at Scale
prime-rl coordinates disaggregated training and inference through asynchronous rollouts, continuous batching, and modular environments. Its infrastructure choices address throughput, sequence-length, optimizer-distribution, and MoE-kernel constraints in the reported training setting.
2.1.1 Architecture
- The orchestrator relays rollouts to the trainer and updated weights to inference, while FSDP 2 shards training state and vLLM serves rollouts.Trainer and inference run on disjoint GPUs so rollout generation and training can overlap.Source: §2.1.1 · p. 5, §2.1.2 · p. 6
2.1.2 Asynchronous Off-Policy Training
- Asynchronous one-step off-policy training lets inference generate rollouts from an older policy while the trainer computes an update, avoiding the synchronous inference stall.The inference service can continue with θ0 while the trainer produces θ1.Source: §2.1.2 · p. 6, Figure 3 · p. 6
2.1.3 Continuous Batching & In-Flight Weight Updates
- Continuous batching and in-flight policy updates support long-context RL with higher end-to-end system throughput while minimizing data off-policyness.Completed rollouts are immediately replaced to keep the request pool saturated; rollouts exceeding the configured number of policy updates are discarded to limit policy drift.Source: §2.1.3 · p. 6
2.1.4 Multi-Client Orchestrator
- A multi-client orchestrator distributes requests round-robin across independent inference nodes, and the reported inference throughput scales linearly with node count.This approach avoids inter-node synchronization, unlike the standard vLLM multi-node data-parallel strategy, whose throughput plateaued as nodes increased.Source: §2.1.4 · p. 7
2.1.5 Online Data Filtering
- Online difficulty pools and filtering use observed solve rates to maintain a changing curriculum and discard rollouts that are always solved or always failed.Problems are assigned to easy, normal, or hard pools, with per-step sampling controlling their mix.Source: §2.1.5 · p. 7
2.1.6 Scaling Sequence Length
- CPU activation offloading raised sequence length to 72k on the same hardware without MFU degradation; synchronous offloading was adopted after a memory leak in the asynchronous variant.The synchronous version reduced MFU by roughly 0.1%; context parallelism reached 256k but halved data parallelism at Ncp = 2 and degraded accuracy, making it unsuitable for production training.Source: §2.1.6 · pp. 7–8
2.1.7 Distributed Muon
- The adopted all-to-all Muon method reshuffles gradient shards, improving performance and avoiding the InfiniBand congestion caused by overlapping gathers at large multi-node scales.It is less flexible and may require padding tensors before communication; the main training runs use the all-to-all implementation.Source: §2.1.7 · p. 8
2.1.8 Efficient Mixture-of-Experts Support
- Expert parallelism was disabled because the training sequence lengths and hidden dimensions already saturate grouped matrix multiplication, leaving scatter/gather overhead without a corresponding reduction in kernel time.The report notes that lower sequence length or hidden dimension, or parallelizations that reduce work per GPU, might benefit from expert parallelism.Source: §2.1.8 · p. 8, Figure 5 · p. 9
2.2 Verifiers: Environments for LLM Reinforcement Learning
Verifiers provides modular environments that can be developed independently of the trainer and reused across training and evaluation. The Environments Hub packages these environments for sharing, while prime-rl integrates them into multi-environment training and online or offline evaluation.
2.2 Verifiers: Environments for LLM Reinforcement Learning
- Verifiers separates environment logic from training infrastructure, enabling environments to be developed, tested, and versioned independently while supporting compositionality and interoperability.The report describes these environments as playing a role analogous to datasets in SFT or pre-training.Source: §2.2 · p. 9
2.2.1 Environment Design
- Environments combine task data, rollout execution, loading, and reward scoring; their inheritance structure supports single- and multi-turn tasks, tool calls, and task-specific behavior.Rollouts can run asynchronously in parallel, while configurable rubrics combine reward functions and can support inter-group comparisons.Source: §2.2.1 · pp. 9–10
2.2.2 Integration with prime-rl
- prime-rl loads standalone verifiers modules from the Environments Hub and supports simultaneous training across environments through EnvGroup routing, without requiring multi-environment-aware orchestrator code.Environment modules return rollout states that include rewards, token IDs, logprobs, and attention masks.Source: §2.2.2 · pp. 10–11
2.2.3 Environments Hub
- The Environments Hub packages environments as standalone Python modules with pinnable dependencies and a standardized entry point, supporting independent versioning, sharing, testing, and iteration.The report presents the Hub alongside prime-rl, secure code sandboxes, and distributed compute infrastructure as part of an open RL research stack.Source: §2.2.3 · p. 11
2.2.4 Evaluations
- Verifiers supports both online and offline evaluation through the same rollout and Rubric entrypoints used for training.Online evaluation requests are interleaved asynchronously with training requests on the shared inference pool, hiding evaluation overhead while providing real-time feedback.Source: §2.2.1 · p. 10, §2.2.4 · p. 11
2.3 Prime Sandboxes: Code Execution for RL Training
Prime Sandboxes addresses the latency and scaling limits of running untrusted code for high-volume agentic RL. Its architecture bypasses the Kubernetes API in the execution path and combines fast readiness signaling, image distribution, dense scheduling, and isolation.
2.3 Prime Sandboxes: Code Execution for RL Training
- The system targets sub-second provisioning and millisecond execution because Kubernetes control-plane patterns cannot meet the throughput needs of thousands of concurrent sandboxes.Public image-registry rate limits and node bandwidth can also delay naive image pulls by minutes.Source: §2.3 · p. 11, §2.3.4 · p. 12
2.3.2 Prime Sandboxes Architecture
- A Rust gateway routes execution requests directly to sandbox pods, bypassing API-server and etcd bottlenecks that drove command latency to 2.5 seconds at thousands of sandboxes.The naive path sent requests through Kubernetes API Server and kubectl exec; each API operation was serialized through etcd write locks.Source: §2.3.1 · p. 11, §2.3.2 · p. 12
2.3.5 Security and Capabilities
- A sidecar uses nsenter to execute commands in the target namespace, combining local-process speed with container isolation.The system uses gVisor and configurable network policies to isolate the host from untrusted code; it also supports arbitrary TCP/UDP ports and GPU mounting.Source: §2.3.2 · p. 12, §2.3.5 · p. 13
2.3.3 Asynchronous Lifecycle Management
- Sidecar webhooks notify the training backend when a sandbox is ready, yielding cold starts under 10 seconds regardless of cluster load.Pre-warmed standard runtime images can be acquired effectively instantaneously.Source: §2.3.3 · p. 12
2.3.4 Image Distribution and Infrastructure Density
- Image distribution combines a private registry with lazy pulling for dynamic environments and warm pools for static images, reducing startup waits.Lazy pulling fetches entrypoint-required chunks first while streaming remaining layers in the background; warm pools avoid image-pull and initialization latency.Source: §2.3.4 · p. 12
- A custom autoscaler and bin-packing scheduler target 256 sandboxes per node, using Burstable QoS to oversubscribe CPU during idle periods.Burstable QoS lets pods request baseline CPU and burst above it when available, matching workloads with brief execution spikes and idle waits.Source: §2.3.4 · p. 13
2.4 Compute Orchestration: Frontier GPU Infrastructure
The training run used a 512-H200 cluster across 64 nodes, with infrastructure choices aimed at reliable distributed training, resource cleanup, and balanced storage and monitoring.
Provisioning and Fabric
- 512 NVIDIA H200 GPUs across 64 nodes formed the cluster, backed by a 400Gbps NDR InfiniBand fabric with throughput validated against a ≥160 GB/s target.Idempotent Ansible playbooks handle hardware discovery and firewall generation; automated binary search isolates straggler nodes when fabric performance degrades.Source: §2.4 · p. 13
Orchestration
- Slurm with Cgroup v2 reclaims resources by freezing and eliminating job cgroups, preventing zombie processes from retaining GPU memory.The authors describe this as container-like isolation without filesystem overhead.Source: §2.4 · p. 13
Storage
- Lustre serves high-throughput trajectories and multi-terabyte checkpoints, while NVMe-backed NFS handles metadata-heavy environments and keyless SSH.The two-tier storage architecture balances I/O performance across different workload needs.Source: §2.4 · p. 13
Observability
- DCGM GPU telemetry feeds Prometheus, with Xid-error and thermal-slowdown alerts supporting proactive node draining before component failures corrupt training progress.The monitoring stack is intended to detect issues early enough to protect training progress.Source: §2.4 · p. 13
3 INTELLECT-3 Training
INTELLECT-3 was trained from GLM-4.5-Air base through supervised fine-tuning and large-scale reinforcement learning, using a diverse mix of reasoning and agentic environments. The run included environment validation, two SFT phases, and RL with online evaluation and off-policy controls.
3 INTELLECT-3 Training
- Training used diverse environments intended to enhance reasoning and agentic capabilities, and both SFT and RL stages ran on a 512-H200 cluster over two months.The training and evaluation stack used open-source Environments Hub contributions.Source: §3 · p. 13, §3.1 · p. 13
3.1 Environments Mix
- The math, science, and logic environments provide 21.2K, 29.3K, and 11.6K problems respectively, with difficulty filtering based on model solve rates.Math answers use rule-based verification with an LLM judge checking rule-based negatives; science also uses math-verify and an LLM judge, while logic adapts tasks and verifiers from SynLogic.Source: §3.1.1 · p. 14, §3.1.3 · p. 14, §3.1.4 · p. 14
3.1.2 Code
- The code environment uses 8.6K single-turn Python challenges, executing up to 15 tests per problem in isolated sandboxes and masking completions on sandbox failure.Difficulty annotations use Qwen/Qwen3-4B-Instruct-2507 solve rates over eight generations per problem.Source: §3.1.2 · p. 14
3.1.5 Deep Research
- DeepDive uses multi-step questions with search, click, open, and finish tools; its implementation was validated by training a 4B Qwen model with SFT followed by RL.The validation used 884 SFT samples and 122 RL steps; the authors say the model’s successful Deep Research learning demonstrates the environment implementation’s correctness and usability.Source: §3.1.5 · pp. 14–15, Figure 7 · p. 15
3.1.6 Software Engineering
- The software-engineering environments adapt R2E-Gym [16] and mini-swe-agent-plus scaffolds, tasking agents with repository issue fixes verified by test-suite changes.Agents can take up to 200 turns and use Bash and file-editing tools; over 20,000 preinstalled-repository images support ephemeral rollouts.Source: §3.1.6 · p. 15
3.2 Supervised Fine-Tuning
- The two SFT phases paired broad chat-and-reasoning training with a smaller agentic phase targeting tool use and long-horizon control.The general phase drew on open-source reasoning and instruction datasets; the agentic phase used open-source and synthetic agentic datasets, with training scaled to 98K context.Source: §3.2 · pp. 15–16, Figure 8 · p. 16
3.3 Reinforcement Learning
- RL used 256 prompts per batch, 16 rollouts each, and a 65,536-token context, with online difficulty filtering and a maximum of eight off-policy steps.The authors used double-sided importance-ratio masking to address trainer-inference mismatch, rather than clipping, and masked rollouts when any token ratio fell below 1e-5.Source: §3.3 · p. 17
- At RL-training end, reasoning benchmark scores were still trending upward with no apparent plateau; early GSPO ablations also showed reward and other metrics collapsing.Online evaluations tracked AIME25, AIME24, LiveCodeBench, HLE, and GPQA every 15 steps; the authors report the GSPO collapse had also been reported in [19] [38].Source: §3.3 · p. 17, Figure 9 · p. 18, Figure 10 · p. 18
4 Evaluations
INTELLECT-3 outperformed GLM-4.5 Air, the best matching comparison model, across the tested benchmarks. It also surpassed the three-times-larger GLM-4.5 on several named benchmarks, while RL-training rewards were still rising without a benchmark plateau.
4 Evaluations
- INTELLECT-3 outperformed GLM-4.5 Air across every tested benchmark, with Table 2 summarizing results across the benchmark suite.The comparison model is Z.ai’s post-trained version of the GLM-4.5 Air base model; the table compares models of similar or larger size.Source: §4 · pp. 18–19, Table 2 · p. 19
- INTELLECT-3 outperformed the 3× larger GLM-4.5 on AIME 2024, AIME 2025, and LiveCodeBench v6.The evaluations used the same settings as API services to support fair comparison.Source: §4 · pp. 18–19
- At the end of RL training, rewards were still increasing and benchmark performance showed no sign of plateauing.The authors planned continued training with a focus on agentic environments.Source: §4 · p. 19
5 Conclusion & Future Work
INTELLECT-3 is a 100B+-parameter post-trained model built on GLM-4.5-Air, with strong performance across several benchmark areas and a supporting open-source RL infrastructure stack. The conclusion highlights further training, broader environment coverage, and long-horizon agent behavior as future directions.
Conclusion
- INTELLECT-3 achieves strong performance across math, code, science, and broader reasoning benchmarks, and is competitive with or ahead of significantly larger frontier models.It is a 100B+ parameters post-train on the GLM-4.5-Air base model.Source: §5 · p. 19
- The stack combines asynchronous RL training features with reusable, versioned environments and evaluation artifacts, supporting multi-week runs on a 512 H200 cluster.prime-rl includes disaggregated trainer and inference, continuous batching, in-flight weight updates, and support for Mixture-of-Experts models.Source: §5 · p. 19
- Open-sourcing the model, environments, and complete training framework aims to narrow the gap between proprietary RL pipelines and what independent researchers, small labs, and companies can build.The same code is available for single-node experiments, mid-scale research runs, and production-scale training.Source: §5 · p. 19
Scaling Agentic RL
- Reward and evaluation curves had not flattened by the end of the RL run, and training remained stable, leading the authors to expect further gains from continued training.They particularly expect gains for complex agentic use cases when more agentic environments are included.Source: §5 · p. 19
Richer RL environments
- INTELLECT-3 used only a small slice of the more than 500 environments released on the Environments Hub, which span areas including autonomous research, computer use, theorem proving, and browser automation.The report identifies scaling RL across more high-quality community-contributed environments, tools, modalities, and real-world workloads as a next step.Source: §5 · pp. 19–20
Long Horizon Agents
- A roadmap direction is training agents to manage their own context over long horizons, using context cutting, isolated sub-branches, and external memory.The report frames the context window as a scarce resource to manage, citing evidence of context rot in long-context models.Source: §5 · p. 20
A.1 Evaluation Environments
The evaluation environments use benchmark-specific datasets, answer verification, and sampling procedures. They cover competition math, STEM multiple-choice questions, coding contests, and broad subject areas.
MATH-500
- MATH-500 uses 500 high-school competition math problems, verifies extracted final answers with math-verify, and runs two generations per problem.The problems are distilled from the original MATH dataset, for 1,000 generations total.Source: §A.1 · p. 27
AIME
- AIME 2024 and 2025 each contain 30 competition math problems, and results use Avg@32, defined as Pass@1 over 32 generations per question.Responses are verified with math-verify without an LLM judge, making the reported numbers more conservative than evaluations using LLM judges.Source: §A.1 · p. 27
GPQA
- GPQA evaluation uses the Diamond subset's 198 hardest Ph.D.-level STEM multiple-choice questions and reports Avg@4, or Pass@1 over four generations per question.Responses are judged by exact match against the ground-truth answer.Source: §A.1 · p. 27
LiveCodeBench
- LiveCodeBench v6 evaluation uses 454 recent contest problems and reports Avg@2, or Pass@1 with two rollouts per problem.The authors integrate the official verification logic with their sandboxes for secure and scalable test verification.Source: §A.1 · p. 27
MMLU-Pro
- MMLU-Pro evaluation covers 12K general STEM multiple-choice questions from MMLU, grades by exact match, and reports Avg@1 over the samples.The questions are a challenging subset of MMLU.Source: §A.1 · p. 27
HLE
- Humanity’s Last Exam evaluation uses its text-only subset of 2,158 questions without giving models additional tools, reporting average solve rate across all samples.The benchmark spans dozens of subjects, including mathematics, humanities, and natural sciences.Source: §A.1 · p. 27
A.2 API models
The API model evaluations route requests to official providers where specified and use recommended sampling settings. The report also describes an adjustment to GPT-OSS 120B evaluation after identifying a reasoning-effort mismatch.
API models
- GLM-4.5-Air, GLM-4.5, and GLM-4.6 evaluations use OpenRouter routes enforced to the official Z.ai API and Z.ai-recommended sampling settings, including temperature 0.6.The models are identified in the passage as GLM-4.5-Air, GLM-4.5, and GLM-4.6.Source: §A.2 · p. 27
- DeepSeek R1 0528 is evaluated through OpenRouter with requests routed to the official DeepSeek API, while DeepSeek v3.2 (Thinking) is evaluated through the official API because suitable OpenRouter hosting was unavailable.Most providers, including the official DeepSeek provider on OpenRouter, hosted the chat version rather than the desired model.Source: §A.2 · p. 27
- GPT-OSS 120B (High) is evaluated through TogetherAI after OpenRouter responses suggested lower-than-advertised reasoning effort; observed scores differ slightly from the OpenAI model card.The report describes its results as an apples-to-apples comparison within the authors' evaluation harnesses, whose formatting and scoring may differ slightly from other implementations.Source: §A.2 · p. 27