Which Experimental Native Backends Are Contract-Gated in MTPLX?

MTPLX gates six experimental native backends—including Step‑3.5 MTP, Nemotron‑H, and DeepSeek—behind the experimental-mtp-cohorts flag, requiring explicit opt-in before the backend registry exposes them to the inference engine.

According to the youssofal/MTPLX repository, the inference platform uses a strict contract‑gating mechanism to isolate unstable features from production workloads. Understanding which experimental native backends are contract‑gated in MTPLX helps developers safely evaluate cutting‑edge model architectures that are not yet generally available. These backends are identified by the support_level value "experimental-native-contract-gated" in the central registry and remain inaccessible until explicitly enabled via command‑line flag or Python API.

What Are Contract‑Gated Experimental Native Backends?

Contract‑gated backends are native implementations that require users to accept experimental terms before activation. In mtplx/backends/registry.py, the registry assigns these backends a support_level of "experimental-native-contract-gated", which prevents the loader from instantiating them unless the experimental-mtp-cohorts flag is active. This design ensures that users explicitly acknowledge the experimental nature of these features, preventing accidental deployment of unfinished backends in production environments.

Complete List of Contract‑Gated Backends

The following six backends are currently marked as contract‑gated experimental native implementations in the MTPLX source tree:

• Step‑3.5 MTP – Implementation in mtplx/backends/step3p5_mtp.py; registry entry defined around lines 229‑236 of registry.py.
• Nemotron‑H MTP – Implementation in mtplx/backends/nemotron_h_mtp.py; registry entry defined around lines 39‑46.
• MIMO MTP – Implementation in mtplx/backends/mimo_mtp.py; registry entry defined around lines 39‑46.
• HY‑V3 MTP – Implementation in mtplx/backends/hy_v3_mtp.py; registry entry defined around lines 52‑59.
• GLM MTP – Implementation in mtplx/backends/glm_mtp.py; registry entry defined around lines 44‑51.
• DeepSeek MTP – Implementation in mtplx/backends/deepseek_mtp.py; registry entry defined around lines 45‑52.

All entries share the identical support_level metadata, causing the registry to reject instantiation attempts unless the experimental cohorts contract is accepted.

How to Enable Contract‑Gated Backends

You can activate the experimental backends by passing the --experimental-mtp-cohorts flag through the CLI or setting the equivalent parameter in Python.

Command‑Line Interface

Enable all contract‑gated experimental native backends when starting the server:

mtplx serve --experimental-mtp-cohorts

To use a specific gated backend, specify it with the model flag:

mtplx serve \
  --model step3p5_mtp \
  --experimental-mtp-cohorts

Python API

When invoking the server programmatically through mtplx.cli, pass the flag as an argument:

from mtplx.cli import main

# Enable experimental‑mtp‑cohorts for a programmatic run

main(["serve", "--experimental-mtp-cohorts"])

For OpenAI‑compatible handler initialization, set the boolean parameter:

from mtplx.server.openai import OpenAIHandler

handler = OpenAIHandler(
    model="nemotron_h_mtp",
    experimental_mtp_cohorts=True   # Activates the contract‑gated experimental path

)

Checking Backend Availability Programmatically

Before attempting to load a model, verify whether the contract‑gated backend is accessible using the BackendRegistry class:

from mtplx.backends.registry import BackendRegistry

registry = BackendRegistry()

# Default check returns False for experimental contract‑gated backends

print(registry.is_enabled("deepseek_mtp"))   # → False

# Check with experimental flag returns True when cohorts are enabled

print(registry.is_enabled("deepseek_mtp", experimental=True))  # → True

This pattern allows conditional logic in deployment scripts that must handle both stable and experimental backend configurations.

Summary

• Six experimental native backends are contract‑gated in MTPLX: Step‑3.5 MTP, Nemotron‑H MTP, MIMO MTP, HY‑V3 MTP, GLM MTP, and DeepSeek MTP.
• Registry marker: Each backend is tagged with support_level="experimental-native-contract-gated" in mtplx/backends/registry.py.
• Activation method: Pass --experimental-mtp-cohorts via CLI or set experimental_mtp_cohorts=True in the Python API.
• Default behavior: Without the flag, registry queries return False for these backends, and instantiation attempts fall back to stable alternatives or raise an availability error.

Frequently Asked Questions

What happens if I try to use a contract‑gated backend without enabling the flag?

The backend registry will block instantiation and return False for availability checks. Depending on your configuration, the system either falls back to a stable non‑experimental backend or raises an explicit error indicating that the contract‑gated feature is unavailable until you enable experimental-mtp-cohorts.

Where is the contract‑gating logic defined in the source code?

The gating logic resides in mtplx/backends/registry.py, which evaluates the support_level metadata field and checks for the presence of the experimental cohorts flag. The CLI parser in mtplx/cli.py and the OpenAI handler in mtplx/server/openai.py propagate this flag to the registry during initialization.

Are contract‑gated backends considered production‑ready?

No. The experimental-native-contract-gated designation explicitly marks these backends as unstable and subject to breaking changes. They are intended for testing and evaluation only, and the maintainers require explicit opt‑in via the cohorts flag to acknowledge the experimental status.

How do I check if a specific backend is enabled in my current session?

Query the BackendRegistry.is_enabled() method with the backend name and the experimental=True parameter. This returns a boolean indicating whether the contract‑gated backend is currently active based on your session’s flag configuration.

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