MTPLX Experimental Backends: Architecture Support and Enablement Guide

MTPLX supports eight experimental backends—including DeepSeek V3-MTP, DeepSeek V4 (AR), and GLM-MTP—that are gated behind the --experimental-mtp-cohorts flag and defined in mtplx/backends/registry.py.

The MTPLX repository extends high-performance inference to specialized transformer architectures through its modular backend system. These experimental backends provide native support for multi-token prediction (MTP) and auto-regressive (AR) model families that require explicit opt-in for testing and development.

Overview of MTPLX Experimental Backends

MTPLX categorizes experimental backends by their support_level metadata in the architecture-compatibility registry. According to the source code in mtplx/backends/registry.py, backends with support_level strings beginning with experimental- represent cutting-edge implementations that are not enabled by default.

The framework identifies three distinct experimental support tiers:

• experimental-native-contract-gated: Backends that implement contract-gated multi-token prediction
• experimental-native-ar-only: Backends restricted to auto-regressive generation without MTP heads
• experimental-mlx-lm-ar-only: Backends for MLX-LM model families lacking native MTP support

Complete List of MTPLX Experimental Backends

The following architectures are available when experimental mode is activated:

Architecture	Backend Module	Support Level	Description
DeepSeek V3-MTP	deepseek_v3_mtp	experimental-native-contract-gated	Contract-gated MTP for DeepSeek-V3 families
DeepSeek V4 (AR)	deepseek_v4	experimental-native-ar-only	AR-only native backend for DeepSeek-V4-Flash models
GLM-MTP	glm_mtp	experimental-native-contract-gated	Contract-gated MTP for GLM-MoE families
MiMo-MTP	mimo_mtp	experimental-native-contract-gated	Contract-gated MTP for MiMo family
Nemotron-H-MTP	nemotron_h_mtp	experimental-native-contract-gated	Contract-gated MTP for Nemotron-H models
Step3.5-MTP	step3p5_mtp	experimental-native-contract-gated	Contract-gated MTP for Step-3.5 models
Hy-V3-MTP	hy_v3_mtp	experimental-native-contract-gated	Contract-gated MTP for Hy-V3 models
MLX-LM-AR	mlx_lm_ar	experimental-mlx-lm-ar-only	AR-only backend for Llama-AR, iQuest-AR, LFM2-MoE-AR, and other MLX-LM families

These entries map architecture IDs to their corresponding backend modules in mtplx/backends/registry.py at lines 146, 229, and 300.

How to Enable MTPLX Experimental Backends

Experimental backends are opt-in only and disabled by default. You must activate them via command-line flag or configuration file before the server will expose these architectures.

Command-Line Activation

Pass the --experimental-mtp-cohorts flag when starting the MTPLX server:

mtplx start --experimental-mtp-cohorts

Configuration File Method

Set the boolean flag permanently in your MTPLX configuration:

mtplx config set experimental_mtp_cohorts true

This updates the experimental_mtp_cohorts value in mtplx/config.py and persists across restarts.

Python Client Verification

Once enabled, the server exposes experimental backends through the standard API:

import requests
import json

resp = requests.post(
    "http://127.0.0.1:8000/v1/chat/completions",
    json={
        "model": "mtplx", 
        "messages": [{"role": "user", "content": "Hello"}]
    },
    headers={"Content-Type": "application/json"}
)
print(json.dumps(resp.json(), indent=2))

Technical Implementation Details

According to the MTPLX source code, experimental backend identification relies on string parsing in the compatibility registry.

In mtplx/backends/registry.py, the system checks the support_level field for the prefix experimental-. For example, line 300 defines the DeepSeek V4 backend with support_level="experimental-native-ar-only", while lines 229 and 146 handle the contract-gated and MLX-LM variants respectively.

The mtplx/cli.py file defines the --experimental-mtp-cohorts argument, which propagates to mtplx/config.py where the experimental_mtp_cohorts boolean controls scheduler behavior. When enabled, the scheduler enters mtp_cohort_experimental mode as documented in docs/concurrency.md, allowing the dispatcher to route requests to these specialized backends.

Summary

• Eight experimental backends are defined in mtplx/backends/registry.py, including DeepSeek V3-MTP, DeepSeek V4 (AR), GLM-MTP, and MLX-LM-AR variants.
• Three support tiers classify experimental capabilities: experimental-native-contract-gated, experimental-native-ar-only, and experimental-mlx-lm-ar-only.
• Opt-in activation requires setting --experimental-mtp-cohorts via CLI or experimental_mtp_cohorts in configuration.
• Developer preview status means these backends are intended for testing and early adoption, not production workloads.

Frequently Asked Questions

How do I enable experimental backends in MTPLX?

You must start the server with the --experimental-mtp-cohorts command-line flag or set experimental_mtp_cohorts to true in your MTPLX configuration file. This flag is defined in mtplx/cli.py and stored in mtplx/config.py.

What is the difference between contract-gated and AR-only experimental backends?

Contract-gated backends (experimental-native-contract-gated) implement multi-token prediction with contract verification for models like DeepSeek V3 and GLM-MTP. AR-only backends (experimental-native-ar-only or experimental-mlx-lm-ar-only) support auto-regressive generation only, lacking MTP heads, as seen in DeepSeek V4-Flash and MLX-LM model families.

Which model families are supported by the MLX-LM experimental backend?

The experimental-mlx-lm-ar-only backend in mlx_lm_ar supports Llama-AR, iQuest-AR, LFM2-MoE-AR, and other MLX-LM families that do not have native MTP capabilities, according to the registry at line 146 of mtplx/backends/registry.py.

Are MTPLX experimental backends production-ready?

No. As indicated by their experimental- prefix in the support level registry, these backends are not enabled by default and their performance is not guaranteed. They are designed for developers and early adopters testing cutting-edge model parallelism features.

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