Grok Model Parameters, Size & Identifier: Complete xAI Guide (2026)

TL;DR

Every Grok model at a glance — parameter counts, API identifier strings, and context windows:

How Many Parameters Does Grok Have?

The answer depends on which Grok model you mean. xAI has released multiple generations, each with different parameter counts:

Grok-1 (Open Source, 2023)

Grok-1 was xAI's first public model and remains the only Grok model that is fully open-source. It uses a Mixture-of-Experts (MoE) architecture with 314 billion total parameters, of which 86 billion are active per inference pass. This means only ~27% of the model's weights fire for any given token.

• Total parameters: 314 billion
• Active parameters: 86 billion
• Architecture: MoE with 8 experts, 2 active
• Context window: 8,192 tokens
• License: Apache 2.0 (open-source)
• API string: grok-1 (deprecated)

Grok-2 (2024)

Grok-2 was a significant jump in capability. xAI never disclosed the exact parameter count, but external estimates place it around 300 billion parameters in a dense or sparse architecture. It introduced image understanding and improved reasoning.

• Total parameters: ~300B (estimated, not confirmed)
• Context window: 128K tokens
• API string: grok-2 (deprecated)

Grok-3 (Late 2024 / Early 2025)

Grok-3 marked xAI's leap to trillion-scale models. Trained on the Colossus supercomputer in Memphis with 100,000+ NVIDIA H100 GPUs, Grok-3 uses a Mixture-of-Experts architecture with approximately 3 trillion total parameters.

• Total parameters: ~3 trillion (MoE)
• Active parameters: Not disclosed (estimated 300-600B active)
• Context window: 128K tokens
• API string: grok-3

Grok-4.1 (Current Flagship, 2025-2026)

Grok-4.1 is the current production model available through the xAI API and the Grok chatbot on X. It builds on Grok-3's architecture with additional training and refinements. The parameter count remains in the ~3 trillion range with improved training data and techniques.

• Total parameters: ~3 trillion (MoE)
• Context window: 256K tokens
• API string: grok-4.1
• Mini variant: grok-4.1-mini (~400B MoE, 128K context)

Grok-5 (Upcoming)

Grok-5 is expected to double the parameter count to 6 trillion parameters. It will feature native multimodal capabilities (text, image, video, audio) and real-time data access from Tesla's fleet and X.

• Total parameters: 6 trillion (expected)
• Context window: 512K+ (expected)
• API string: grok-5 (expected)
• Release: Q1-Q2 2026

How to Find the Grok Model Identifier String

If you need the exact model identifier string for API calls, here is what to use:

xAI API (Direct)

The xAI API follows OpenAI-compatible formatting. Your API call looks like this:

curl https://api.x.ai/v1/chat/completions \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.1",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Available model strings as of March 2026:

Listing Available Models

You can query the API to get all currently available model identifiers:

curl https://api.x.ai/v1/models \
  -H "Authorization: Bearer $XAI_API_KEY"

This returns a JSON list of all model objects, each with an id field containing the model identifier string.

Using Grok in Third-Party Tools

Many tools and frameworks support xAI's Grok models. The model identifier string is the same — you just need the xAI API base URL:

• Base URL: https://api.x.ai/v1
• Model ID: grok-4.1 (or whichever variant you need)
• Authentication: Bearer token with your xAI API key

Grok vs Other Frontier Models (March 2026)

How does Grok stack up against the competition?

Key Takeaways

• Grok leads on real-time data — access to X and Tesla fleet data is a genuine differentiator no other lab can match.
• Grok trails on reasoning — the ARC-AGI-2 gap (55% vs 77.1% for Gemini) is significant. Grok-5 needs to close this.
• Grok is competitively priced — cheaper than GPT-5.2 and Claude Opus, more expensive than Gemini 3.1 Pro.
• Parameter count doesn't determine quality — Grok-4.1 has the most disclosed parameters but doesn't lead benchmarks. Architecture, training data, and RLHF matter more.

Understanding Mixture-of-Experts (MoE) Parameters

When xAI says Grok has "3 trillion parameters," that number deserves context. In a Mixture-of-Experts architecture, only a fraction of parameters are active during each forward pass.

1. The model has many "expert" sub-networks
2. A router network decides which experts to activate for each token
3. Typically 2-4 experts out of 8-16+ are active per token
4. Total parameters are large, but compute cost is closer to a smaller dense model

Grok-1 is the only model where xAI confirmed the exact numbers: 314B total, 86B active (8 experts, 2 active). The later models have not had their expert configurations publicly confirmed.

Frequently Asked Questions

How many parameters does the current Grok model have?

Grok-4.1, the current flagship model, has approximately 3 trillion parameters in a Mixture-of-Experts architecture. The active parameter count per inference is lower, likely 300-600 billion.

What is the model identifier string for Grok?

The current model identifier for the xAI API is grok-4.1 for the flagship model and grok-4.1-mini for the fast variant. Use these strings in your API calls.

Is Grok open source?

Only Grok-1 (314B parameters) is open-source under an Apache 2.0 license. All subsequent models (Grok-2, 3, 4.1, and the upcoming 5) are proprietary.

How does Grok's parameter count compare to GPT-5?

Grok-4.1 at ~3 trillion parameters is the largest disclosed parameter count among frontier models. GPT-5.2's parameter count has not been confirmed but is estimated around 2 trillion. However, parameter count alone does not determine model quality.

When will Grok-5 be released?

xAI has indicated a Q1 2026 release for Grok-5 with 6 trillion parameters. As of March 2026, no exact date has been announced. It could arrive any day or slip to Q2.

What context window does Grok support?

Grok-4.1 supports 256K tokens. Grok-4.1-mini supports 128K tokens. Grok-5 is expected to expand this to 512K or more.

Build With Any AI Model

The AI model landscape in 2026 is fragmented — Grok, GPT, Claude, Gemini all have different strengths. The smart move is building products that can swap between models as the landscape shifts.

• xAI API Documentation
• xAI Grok-1 Open Source Release
• KuCoin: Grok 5 with 6 trillion parameters
• SentiSight: What to expect from Grok in 2026