Apodex: The AI That Fact-Checks Itself Before It Answers You

Links & Resources

• Website: apodex.ai
• Blog Post: apodex.com/blog/apodex-1.0
• API Platform: platform.apodex.ai
• Hugging Face: huggingface.co/apodex
• GitHub: github.com/ApodexAI
• AgentHarness: github.com/ApodexAI/AgentHarness
• Technical Report: apodex.com/pdf/20260608

A new AI just scored 90.3 on BrowseComp and 94.4 on DeepSearchQA - beating GPT-5.5 and Claude Opus 4.8 on the hardest deep research benchmarks that exist. It's called Apodex, it dropped June 8, 2026, and the whole thing is open-source.

Why Apodex Is Different

Every other AI research tool works the same way: one model, one chain of thought, one output. If it hallucinates halfway through, that hallucination carries forward into everything that follows.

Apodex doesn't work like that. It runs up to 150 sub-agents in parallel, coordinated by an orchestrator, and then passes the entire output through a dedicated verification team before anything reaches you.

The result isn't "AI-generated research." It's research that's been internally peer-reviewed before you ever see it.

The Architecture - How 150 Agents Work Together

Apodex runs a system called AgentOS - a task-agnostic runtime that coordinates the entire agent team:

The Pipeline

1. Orchestrator receives your query, decomposes it into sub-tasks, and asynchronously dispatches specialized sub-agents. Each sub-agent gets its own context, prompt, and tool set.

2. Sub-agents (up to 150) explore in parallel - running retrieval, analysis, and verification tasks simultaneously. Their reports flow into a shared pool. The orchestrator reads asynchronously, never blocking on the slowest agent.

3. Verification Team kicks in when reports disagree, claims need grounding, or a draft is ready for final review. This team consists of three dedicated agents:

• Conflict Reviewer - resolves disagreements between sub-agents
• Fact Checker - verifies every claim against source evidence
• Draft-Report Reviewer - audits the final assembled report

4. Global Verifier reasons over all assembled evidence to produce the final answer. Only verified conclusions make it through.

The system coordinates up to 15,000 steps within a single task. And here's the counterintuitive part: heavy-duty mode often spends fewer total steps than a single-agent approach, because the verifier filters out steps that produce no information gain.

The Benchmarks

Deep Research (Where Apodex Dominates)

The heavy-duty mode adds +14.8 points over base on BrowseComp (75.5 → 90.3). That's not an incremental improvement - that's a different class of system.

Science Benchmarks (Where It Gets Ridiculous)

These aren't close contests. Apodex is beating the next-best model by 8 to 12 points on every science benchmark.

Coding (Solid, Not the Focus)

Apodex isn't built to be a coding model. It holds its own, but deep research and scientific reasoning is where it's built to dominate.

The Open-Source Models

Apodex open-sourced a full family of models, including one small enough to run on your own machine:

The 4B model deserves special attention: it outperforms every open-source model up to 30B parameters on both BrowseComp and BrowseComp-ZH. A 4-billion parameter model competing with models 7x its size.

All models preserve general capabilities - they track their matched-size Qwen3.5 bases within roughly a point on MMLU-Pro, AIME 2026, IFEval, and LongBench v2.

How to Use Apodex

Option 1: Use It Online (Easiest)

Go to apodex.ai → sign in → ask your question. The system returns step-level reasoning traces, citations in every report, and full-text search across threads.

Option 2: API Access

Use the Apodex Platform API to integrate Apodex into your own tools and workflows.

Option 3: Self-Host with SGLang

pip install sglang

python3 -m sglang.launch_server \
  --model-path apodex/Apodex-1.0-mini \
  --tp 8 \
  --host 0.0.0.0 \
  --port 1234 \
  --context-length 262144

Option 4: Self-Host with vLLM

vllm serve apodex/Apodex-1.0-mini

Query via OpenAI-Compatible API

curl -X POST "http://localhost:1234/v1/chat/completions" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "apodex/Apodex-1.0-mini",
    "messages": [{"role": "user", "content": "Your research question here"}]
  }'

Recommended Settings

temperature: 1.0
top_p: 0.95
repetition_penalty: 1.05
max_context_length: 262,144
max_tokens: 32,768

The model supports native function calling - pass tool schemas via the tools parameter of the chat-completions API.

AgentOS - The Runtime Behind It

Apodex runs on AgentOS, a task-agnostic runtime that provides:

• Scheduling - coordinates agent dispatch and parallelism
• Model and tool routing - routes sub-agents to the right models and tools
• Event streaming - real-time progress tracking
• Checkpoints and traces - full auditability of every reasoning step
• Cost accounting - tracks token spend across the agent team
• Permission enforcement - controls what each agent can access

Adding a new application to AgentOS is a folder of plugin code, not a patch to the kernel. Strict separation between workflow policy and runtime execution.

Where to Get It

Follow AI Adventure YT for more AI breakdowns and guides.