Which Model Architectures Are Fully Verified-Native in MTPLX?

Only the Qwen 3 Next MTP architecture (qwen3-next-mtp) holds the fully verified-native support level in MTPLX, making it the sole model family with complete verification and native multi-token prediction capabilities out-of-the-box.

The MTPLX repository (youssofal/MTPLX) maintains a strict architecture catalog that categorizes model families by their implementation maturity and verification status. Understanding which model architectures are fully verified-native is critical for production deployments requiring guaranteed multi-token prediction (MTP) stability and native backend optimization.

Understanding Verified-Native Support Levels

MTPLX classifies model architectures using a tiered support system defined in the ARCHITECTURE_CATALOG. The verified-native designation represents the highest tier, indicating the architecture has passed full verification suites and implements native MTP without compatibility layers.

According to the source code in mtplx/backends/registry.py, this catalog assigns each architecture a support_level string that determines how the backend initializes the model and which optimizations are enabled.

The Only Fully Verified-Native Architecture: Qwen 3 Next MTP

As of the current MTPLX codebase, Qwen 3 Next MTP (qwen3-next-mtp) stands alone as the only architecture with verified-native status.

In mtplx/backends/registry.py at line 211, the architecture entry explicitly sets:

support_level="verified-native"

This designation means the model can run with complete verification and native multi-token prediction support immediately upon initialization. No additional configuration flags or compatibility modes are required for full functionality.

Alternative Support Levels for Other Architectures

All other architectures in the MTPLX catalog use alternative support designations that exclude them from fully verified-native status:

• verified-native-ar-only — Provides verified native support exclusively for autoregressive (AR) generation, lacking full MTP verification
• experimental-native-contract-gated — Requires explicit contract acceptance and carries experimental stability guarantees
• recognized-backend-pending — Acknowledged by the backend but pending full verification integration

For example, the laguna-s-2.1-ar architecture carries the verified-native-ar-only label, indicating it does not support the complete verified-native MTP pipeline available to Qwen 3 Next MTP models.

How to Verify Architecture Support Levels Programmatically

You can query the support level for any registered architecture using the MTPLX CLI inspector. The tool reads directly from the ARCHITECTURE_CATALOG defined in the registry.

To check if a specific model is fully verified-native:

# Query the support level for Qwen 3 Next MTP

mtplx inspect --model qwen3-next-mtp --json | jq '.support_level'

The command returns:

"verified-native"

Contrast this with a non-fully-verified architecture:

# Check a model with AR-only support

mtplx inspect --model laguna-s-2.1-ar --json | jq '.support_level'

Which returns:

"verified-native-ar-only"

Implementation Details in Source Files

Two critical files manage the verified-native classification system:

• mtplx/backends/registry.py — Contains the ARCHITECTURE_CATALOG dictionary where each model family’s support_level is defined. This is where the verified-native string is assigned to qwen3-next-mtp at line 211.

• mtplx/inspect.py — Implements the CLI mtplx inspect command that queries the catalog and returns the support level metadata for validation and debugging purposes.

Summary

• Only Qwen 3 Next MTP (qwen3-next-mtp) holds the fully verified-native support level in the MTPLX architecture catalog.
• The designation is defined in mtplx/backends/registry.py with support_level="verified-native".
• Other architectures use lower tiers like verified-native-ar-only or experimental-native-contract-gated.
• Use mtplx inspect --model <name> --json to programmatically verify any architecture's support status.

Frequently Asked Questions

What does verified-native mean in MTPLX?

Verified-native indicates an architecture has passed comprehensive verification suites and implements native multi-token prediction without compatibility layers or experimental flags. According to the mtplx/backends/registry.py source, this is the highest support tier in the ARCHITECTURE_CATALOG.

Is Qwen 3 Next MTP the only fully verified-native model?

Yes. As implemented in youssofal/MTPLX, the qwen3-next-mtp architecture is the sole entry with support_level="verified-native" at line 211 of mtplx/backends/registry.py. No other model families currently qualify for this top-tier classification.

What is the difference between verified-native and verified-native-ar-only?

Verified-native provides full multi-token prediction (MTP) support with complete verification, while verified-native-ar-only restricts verified native execution to autoregressive generation only. The latter lacks full MTP verification, as seen with architectures like laguna-s-2.1-ar in the registry.

How do I check if my model is verified-native?

Run the MTPLX CLI inspector command: mtplx inspect --model <architecture-name> --json | jq '.support_level'. If the output returns "verified-native", the model has full verified-native status. Any other value indicates limited or experimental support according to the catalog definitions in mtplx/backends/registry.py.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}