AI Models
Open-source deep-research agents are advancing rapidly as reinforcement learning teaches models to search, navigate, use tools, evaluate evidence, and complete long-horizon research tasks. This guide compares the strongest available models and the infrastructure needed to train and deploy them reliably at scale.
Open-source deep-research agents are advancing rapidly as reinforcement learning teaches models to search, navigate, use tools, evaluate evidence, and complete long-horizon research tasks. This guide compares the strongest available models and the infrastructure needed to train and deploy them reliably at scale.
Reinforcement learning is turning deep-research agents from brittle scripted systems into adaptable models that can plan, recover from errors, evaluate evidence, and use live web tools. The strongest open models span compact 3B and 4B systems through sparse 30B and dense 32B models, giving teams meaningful choices across quality, latency, and training cost.
Modal is a serverless cloud platform for compute-intensive AI workloads rather than a foundation model. For browser-agent teams, it supplies the infrastructure layer needed to generate trajectories, execute isolated browser sessions, train policies on GPUs, evaluate checkpoints, and deploy inference endpoints without operating a fixed cluster.
Agentic RL alternates between model inference, environment interaction, reward computation, and parameter updates. Modal lets each stage run with the hardware and scaling behavior it needs while sharing code, datasets, volumes, and secrets across the pipeline.
Browser-agent RL is unusually bursty: rollout generation needs high concurrency, policy updates need tightly coupled GPUs, and evaluation needs reproducible environments. Modal separates these resource profiles so teams can scale each phase independently instead of provisioning for peak demand across the entire loop.
Modal's platform is designed for production AI systems that need rapid container startup, high concurrency, observability, and secure execution. Browser-agent developers can use the same primitives for experimental rollouts and customer-facing workloads, reducing the gap between research infrastructure and deployment infrastructure.
Tongyi-DeepResearch-30B-A3B is an open deep-research agent from Alibaba built for long-horizon information seeking. Its mixture-of-experts design activates roughly 3B parameters from a 30B-parameter model, combining broad capacity with more manageable inference cost than a similarly sized dense model.
The model is designed to plan searches, inspect evidence, revise its strategy, and synthesize answers across many interactions. That makes it a strong base for RL environments where rewards depend on both final-answer correctness and efficient tool use.
Tongyi-DeepResearch reports a broad set of competitive web-agent results. Scores should be compared with attention to each benchmark's environment, tool configuration, and evaluation protocol.
LiteResearcher-4B targets deep research at a substantially smaller parameter count. It is particularly relevant for teams that need to generate many trajectories, iterate quickly, or deploy multiple concurrent agents under tighter GPU budgets.
LiteResearcher's training emphasizes large-scale tool interaction and retrieval grounded in a local corpus. A local collection of 32 million documents supports reproducible search, while 73.2 million tool calls provide substantial behavioral data for learning when and how to retrieve information.
Despite its compact size, LiteResearcher reports results that compete with much larger research agents.
WebSailor-V2-30B-A3B is built for complex web navigation and information seeking. Like Tongyi-DeepResearch, it uses a sparse 30B-A3B configuration, making it well suited to difficult tasks that benefit from substantial model capacity without requiring all parameters to be active for every token.
WebSailor-V2 focuses on maintaining useful reasoning across long, partially observable trajectories. Its architecture and post-training are aimed at agents that must decide which pages to visit, extract relevant evidence, and continue searching when an early path fails.
WebSailor-V2 reports 35.3 on BrowseComp EN and 44.1 on BrowseComp ZH. The stronger Chinese result makes it especially relevant to multilingual web-agent programs, while its sparse execution profile supports larger rollout batches than a dense 30B model under comparable infrastructure constraints.
BrowserAgent-RFT is an adjacent approach that improves browser behavior through rejection fine-tuning rather than online reinforcement learning. It generates candidate trajectories, evaluates them, retains successful examples, and fine-tunes the model on the filtered data.
Rejection fine-tuning can be simpler and more stable than full RL because it does not require an online policy optimizer. It is useful when teams can define reliable success checks but do not yet have the infrastructure or reward shaping needed for continuous RL.
The approach can improve action selection, recovery behavior, and adherence to browser workflows without introducing policy-gradient instability. Its limitation is that it learns from accepted samples rather than directly optimizing a reward through exploration, so gains depend heavily on the quality and diversity of generated trajectories.
WebDancer-32B is a dense web-agent model designed for autonomous information seeking. Its work is notable not only for the resulting checkpoint but also for a training recipe that turns large-scale web interaction into progressive agent improvement.
WebDancer uses a four-step pipeline that moves from task construction to supervised initialization and reinforcement learning. Representing each stage explicitly helps teams reproduce the workflow with their own environments and reward functions.
WebDancer reports 51.5 on GAIA and 47.9 on WebWalker. These results show useful performance on both general-assistant research tasks and structured website traversal, though its dense 32B architecture requires more memory and compute per rollout than sparse or compact alternatives.
ToolMind-Web-3B is a compact tool-using web agent that prioritizes efficient deployment. At 3B parameters, it can support high rollout concurrency and lower-latency serving while retaining strong benchmark performance.
The model is suited to search, retrieval, and structured tool-use workflows where teams want to train specialized policies without the cost of a large foundation model.
The model card states that ToolMind-Web-3B can execute long sequences of tool calls, making it suitable for extended web research workflows that require repeated navigation, retrieval, and synthesis.
ToolMind-Web-3B reports 75.1 on Xbench and 67 on GAIA. Its quality-to-size ratio makes it a strong candidate when infrastructure cost and agent concurrency matter as much as peak benchmark performance.
OpenWebRL-4B focuses on reinforcement learning for interactive online web tasks. Its compact size and available training artifacts make it useful for researchers who want to study policy optimization rather than only consume a finished browser-agent checkpoint.
OpenWebRL's value lies in connecting model weights with an RL-oriented workflow for web environments. Accessible artifacts can help teams inspect data, reproduce experiments, and adapt rewards or environments for proprietary tasks.
OpenWebRL-4B reports 67 on Online Mind2Web and 64 on DeepShop. These evaluations emphasize actual interaction rather than text-only research, making the model relevant to transactional and workflow-oriented browser agents.
Choosing a model is only one part of building a reliable deep-research agent. RL training requires thousands of environment instances, scalable model serving, durable data storage, GPU-intensive updates, and secure execution. Modal brings those pieces together in a code-first platform.
Modal lets teams define compute, dependencies, secrets, storage, and scaling behavior in Python. Browser environments can run separately from policy servers and trainers, so each component remains reproducible and independently scalable.
Rollout workloads expand and contract rapidly. Modal can scale from zero to large fleets of containers, allowing teams to collect trajectories in parallel without paying for an always-on browser cluster. GPU workers can scale independently for batched inference or training updates.
A single Modal application can coordinate data generation, online environments, rewards, training, evaluation, and deployment. This reduces handoffs between separate systems and makes experiments easier to reproduce.
Web agents interact with external content and may execute unpredictable actions, so isolation and secret management are essential. Modal provides security controls, encrypted secrets, network configuration, and gVisor-based sandbox isolation for workloads that require stronger boundaries.
The same platform primitives used for training can serve production agents with autoscaling, observability, and fast container startup. Teams can move from small experiments to high-concurrency rollout generation and customer-facing inference without rebuilding the entire infrastructure stack.
Explore the Modal documentation to start building scalable training and deployment pipelines for deep-research agents.
Read the Modal DocumentationOpen-source models give teams control over weights, training data, reward design, deployment, and privacy. They also make it possible to inspect behavior, reproduce experiments, fine-tune for proprietary workflows, and choose an architecture that fits a specific cost and latency target rather than depending on a closed API.
Reinforcement learning optimizes behavior against outcomes produced through interaction. For deep-research agents, that means a model can learn to choose better searches, navigate multi-step workflows, recover from failed actions, evaluate evidence, use tools efficiently, and trade off exploration against task completion. Outcome rewards also capture behavior that is difficult to express through next-token supervision alone.
The answer depends on the task. Tongyi-DeepResearch-30B-A3B is a strong general choice for deep research and long-horizon information seeking, while WebSailor-V2-30B-A3B targets difficult web navigation. For interactive automation under tighter budgets, OpenWebRL-4B and ToolMind-Web-3B offer attractive compact alternatives. Teams should evaluate candidates in their own browser environment because benchmark tools and protocols differ.
Key requirements include high-concurrency browser environments, secure isolation, scalable GPU inference, durable trajectory and checkpoint storage, reliable reward computation, reproducible container images, and observability across long-running tasks. A platform such as Modal can run these stages with independent scaling while keeping them in one code-defined workflow.
Yes. Open weights can be adapted with supervised fine-tuning, rejection fine-tuning, or reinforcement learning using trajectories collected from a proprietary environment. Teams should define safe action spaces, deterministic success checks where possible, representative tasks, and held-out evaluations that measure both completion quality and policy safety.
Larger dense models generally require more GPU memory, produce lower rollout throughput, and cost more to train and serve, but they may offer stronger reasoning capacity. Compact 3B and 4B models support more concurrent agents and faster iteration. Sparse MoE models such as 30B-A3B checkpoints hold greater total capacity while activating only a subset of parameters per token, although they still require memory for the full model and specialized serving support.