Apodex
Beats Claude-Opus-4.7 on deep research
Last verified:
What is Apodex?
Apodex is a multi-agent research system and self-evolving heavy-duty solver that performs deep research by verifying every claim before committing to an answer. Built on a fork of the slime reinforcement learning framework, it generates research reports with auditable evidence chains, achieving state-of-the-art results on deep research benchmarks. The system reasons through questions step by step, verifying every conclusion before moving to the next, producing verified briefs rather than simple chat replies.
Key features include a verification-centric agent team with up to 150 sub-agents coordinating over 15,000 steps per task, a global verifier that audits assembled evidence before any answer is committed, and auditable-by-construction reports where every claim traces back to an evidence graph node. The system includes AgentOS (a runtime for building and evaluating agent workflows), multiple model variants (Apodex-1.0-H flagship, Apodex-1.0, Apodex-mini), and open-weight Smol models at 0.8B, 2B, and 4B for independent verification. It has native function calling, 262K token context window, and achieves 90.3 on BrowseComp, 94.4 on DeepSearchQA, and 60.8 on HLE-Text.
Apodex is designed for researchers, scientists, developers, and builders who need long-horizon research capabilities including planning investigations, gathering evidence, using tools, testing hypotheses, and synthesizing grounded conclusions. It excels at macroeconomics questions, medical evidence research, ML research comparisons, clinical evidence analysis, and regulatory shifts tracking. The system is built for mission-critical tasks where knowledge must be shown, supported, and understood—not merely asserted.
Apodex pricing
Pricing model: Freemium
The website does not explicitly display pricing details. The Terms of Service mention service credits can be purchased for some services, and billing information is required for paid accounts. The API Platform is available at platform.apodex.ai. Open-weight models (0.8B, 2B, 4B) are released under Apache 2.0 license for free use. Specific free tier limits, paid plan costs, and credit pricing are not publicly stated on the main website.
Apodex pros
- State-of-the-art performance on deep research benchmarks
- Verifies every claim before committing to an answer
- Generates research reports with auditable evidence chains
- Coordinates up to 150 sub-agents in a single task
- 262K token native context window for long-horizon research
- Native function calling with tool invocation capabilities
- Global verifier audits assembled evidence before delivery
- Every claim traces back to evidence graph node
- Reports are auditable, retractable, and forkable by construction
- Preserves general capabilities while focusing on deep research
- Open-weight Smol models available for independent verification
- AgentOS runtime for building and evaluating agent workflows
- Multiple model variants including flagship H and mini versions
- Apache 2.0 license allowing commercial use
- Performs parallel exploration with specialized sub-agents
Apodex cons
- New release with limited established user base
- Requires technical setup for self-deployment (SGLang/vLLM)
- Heavy-duty mode needs asynchronous agent team infrastructure
- Open-weight Smol models have lower benchmark scores
- 0.8B model only achieves 13.9 on BrowseComp
- May coordinate over 15,000 steps per task (resource intensive)
- Limited documentation publicly available
- No clear free tier pricing information on website
Frequently asked questions about Apodex
What is Apodex and what does it do?
Apodex is a multi-agent research system and self-evolving heavy-duty solver built on a fork of the slime reinforcement learning framework. It performs deep research by planning investigations, gathering evidence, using tools, testing hypotheses, and synthesizing grounded conclusions. The system verifies every claim before committing to an answer and generates research reports with auditable evidence chains.
What benchmarks does Apodex achieve on?
Apodex-1.0-H achieves state-of-the-art results across multiple research benchmarks: 90.3 on BrowseComp, 60.8 on HLE-Text, 94.4 on DeepSearchQA, 46.7 on FrontierScience Research, 87.4 on FrontierScience-Olympiad, and 74.2 on SuperChem. The appliance also achieves 84.1 on BrowseComp-ZH and 82.2 on DeepSearchQA for Apodex-1.0-mini.
How does the verification system work?
Apodex uses a verification-centric agent team where an orchestrator dispatches specialized sub-agents that explore in parallel. A global verifier audits the assembled evidence before any answer is committed. Every claim in the final answer traces back to a node in the evidence graph and is independently checked before delivery. The report pool records every finding, verdict, and intervention.
What model variants are available?
The release includes Apodex-1.0-H (flagship deep research model), AgentOS (runtime for building and evaluating agent workflows), Apodex-1.0, Apodex-mini, and open-weight Smol models at 0.8B, 2B, and 4B. The Smol models are optimized for independent verification of other agents' outputs rather than task execution itself.
What is the context window size?
Apodex has a native 262K token context window (262,144 tokens), enabling it to handle long-horizon research tasks that require processing extensive documentation and evidence.
Can I deploy Apodex locally?
Yes, Apodex follows the Qwen3.5 chat template and can be deployed with SGLang or vLLM for an OpenAI-compatible endpoint. The recommended deployment includes tensor parallel size 8, context length 262,144, tool-call-parser qwen3_coder, and reasoning-parser qwen3. The models are released under Apache 2.0 license.
What use cases is Apodex best for?
Apodex is built for long-horizon research including macroeconomics questions (like Fed rate cuts), medical evidence research (GLP-1 cardiovascular effects, SGLT2 inhibitor combinations), ML research comparisons (Transformer vs. SSM), clinical evidence analysis, and AI policy/regulatory shifts tracking. It excels at mission-critical tasks requiring verified, grounded conclusions.
How many sub-agents can Apodex coordinate?
In deployment, Apodex coordinates up to 150 sub-agents over 15,000 steps in a single task. The sub-agents specialize in retrieval and verification, route their reports through a shared evidence pool, and feed the global verifier.
What is the license for Apodex models?
Apodex-1.0 is released under Apache 2.0 license, which allows commercial use. The open-weight Smol models (0.8B, 2B, 4B) are also under Apache 2.0/MIT licenses.
What are the recommended settings for agentic tasks?
For optimal performance in agentic tasks, the recommended settings are: temperature 1.0, top_p 0.95, repetition_penalty 1.05, max_context_length 262,144, and max_tokens 32,768. The system is trained for native function calling with tool schemas passed via the tools parameter of the chat-completions API.