Julia Routing Strix Halo

Fine-tuned Julia-1 checkpoint for fast bounded-decision routing on AMD Strix Halo-class local machines.

Runtime status:

• CPU: tested
• GPU: expected to work through the Julia/PyTorch runtime, not separately benchmarked for this release
• NPU: not validated yet; AMD/XDNA NPU packaging and execution are future work

This model is intended for advisory routing and gating metadata, not free-form generation and not safety authority.

Intended use

Good candidates:

• task routing over fixed options
• eval-case routing: category and deterministic grader type
• documentation-update triage
• context-management metadata
• non-authoritative tool-policy hints

Do not use this model as the final authority for:

• destructive command approval
• credential/secret handling
• production deploy approval
• security replay safety
• durable memory writes
• conversation summarization or final prose

Deterministic policy and the main agent must remain authoritative for those cases.

Training summary

• Base model: SupersonicLabs/Julia-1
• Fine-tuning method: native partial fine-tune, no LoRA
• Trainable parameters: 7,634,563
• Unfrozen modules: decision head plus last 2 encoder layers
• Learning rate: 1e-5
• Epochs: 1
• Training material: synthetic/adversarial bounded-decision examples for local agent routing and gating

Local eval summary

Qualification suite:

Separate holdout suite:

The holdout result is the better estimate of generalization. The qualification score is useful but should be treated as a tuned benchmark.

All latency numbers above were measured on CPU. NPU acceleration has not been validated for this checkpoint.

Usage

Install Julia runtime from the base model package, then load this checkpoint. CPU is the validated path for this release:

from julia import load_model

engine = load_model(
    "./julia-routing-strix-halo",
    device="cpu",
    strict_encoding=True,
    max_length=1024,
    head_length=512,
)

result = engine.predict(
    state={"request": "Fix the failing TypeScript test and run the suite."},
    questions={
        "route": {
            "type": "choice",
            "instructions": "Which worker route should handle this request?",
            "criteria": {
                "code_edit": "modify repository files",
                "browser_qa": "interact with a web UI",
                "research": "gather information only",
                "chat": "answer conversationally",
            },
        }
    },
)
print(result)

Files

• model.safetensors — fine-tuned Julia decision weights
• julia_config.json — Julia runtime configuration
• encoder/config.json — encoder configuration
• tokenizer/ — tokenizer files

Limitations

• The model is specialized for bounded decisions and fixed option sets.
• It is not a general-purpose embedding model.
• It is not a generative model.
• It can be confidently wrong outside the routing distributions represented in training.
• Use confidence thresholds, deterministic guards, and fallback to a main model for ambiguous cases.
• NPU execution has not been validated yet. Do not advertise this as an NPU-ready checkpoint until an XDNA/NPU runtime path is tested.