English documentation · runtime 7.24.4 · SDK 2.6.7. Content is maintained with runtime development; see each guide's scope and review date.

MTPLX and oMLX local runtimes

Status: Active

Scope: source-checkout integration

Last reviewed: 2026-09-19

Owner: ax-code runtime

The source checkout includes separate MTPLX (mtplx) and oMLX (omlx) presets under /connect → Local LLM runtime. These presets are not included in the v7.19.3 packaged binaries. They connect to an already-running server; AX Code does not install, launch, or tune these runtimes.

Connect

1. Start your runtime and load a model that supports coding tools. Follow the MTPLX instructions or oMLX instructions.
2. Select MTPLX or oMLX in AX Code’s local runtime menu and confirm the server endpoint.
3. Select a coding-capable model. Model IDs come from the server’s GET /v1/models response.

Both upstream servers default to port 8000. Run them one at a time, or assign different ports and save those endpoints separately. AX Code appends /v1 once if omitted. A saved endpoint takes precedence over the host override. Listing a preset does not probe or activate it. Existing provider enable/disable lists remain authoritative; if you use an allowlist, add the new ID without removing your other providers.

Examples with explicit ports:

mtplx serve --model /path/to/your/mtplx-model --host 127.0.0.1 --port 8000
omlx serve --model-dir /path/to/your/models --host 127.0.0.1 --port 8001

Use the endpoint corresponding to the server you started. Inspect its model IDs with curl http://localhost:8000/v1/models (change the port for your configuration).

Coding tools and authentication

MTPLX’s native chat listing identifies owned_by: "mtplx" and capability: "chat" separately from retrieval models. AX Code admits that chat transport for tool use unless it explicitly disables tools. A generic or unrecognized listing retains conservative defaults.

oMLX’s model listing does not declare per-model tool support and can include non-chat models. Enable tools only for a model whose chat template supports them. Add this to your AX Code configuration, replacing the exact model ID returned by your server:

{
  "provider": {
    "omlx": {
      "options": { "baseURL": "http://localhost:8001/v1" },
      "models": {
        "your-tool-capable-model-id": { "tool_call": true }
      }
    }
  }
}

The explicit tool choice survives discovery. Other discovered models are not promoted to coding models. oMLX’s native max_model_len is used for the context limit when present. Runtime sampler settings, MTP enablement, and server-side output limits remain owned by the runtime.

For a server requiring authentication, set provider.mtplx.options.apiKey or provider.omlx.options.apiKey in your configuration, for example "{env:OMLX_API_KEY}". AX Code sends that configured bearer credential for discovery as well as inference. The local endpoint dialog does not prompt for a token. Do not put credentials in the endpoint URL.

AXQuant Qwen MTP in oMLX

AXQuant’s Qwen MTP packs include a separate mtp.safetensors head and a mtplx_runtime.json execution contract. Recent metadata also includes axquant_omlx_compat.json, which lists the 15 MTP tensors for this 27B pack. These files describe compatibility; turning on Lightning MTP alone does not import the head.

For oMLX 0.6.4, use a complete, writable local model copy. In Model Settings, select Import MTP side-car, then enable Lightning MTP. The upstream importer prepares an MTP shard and updates that local copy’s checkpoint index. Keep the original downloaded pack intact for other runtimes; do not edit a shared Hugging Face cache snapshot in place. See the AXQuant model instructions. AX Code connects to the prepared server and does not modify model weights or perform this import.

Performance and scope

Loopback MTPLX and oMLX connections receive the local prefix-stability behavior: deterministic tool ordering and transient context after conversation history. Remote endpoints and AX Trust-managed connections retain their existing behavior; cloud and AX Trust settings are not changed by these presets.

See the AX Code/OpenCode client retest for the complete MTP-enabled matrix after the local prefix fix, including hardware, versions, model identity, cache state, timing definitions, and code checks. The earlier runtime measurements retain native decode replays and historical client results. A working OpenAI-compatible connection does not establish a model’s coding quality or guarantee a decode rate.