Nace AI Open-Sources Drex Max v1: A 27B Multimodal Decision Model That Runs Locally

SAN FRANCISCO — Providing autonomous agent engineers with an ultra-efficient alternative to massive cloud generative models, AI research lab Nace AI has officially open-sourced Drex Max v1, a dense 27-billion-parameter multimodal decision model explicitly built for agent choice-scoring and environment state evaluation. Designed to run locally on consumer-accessible hardware—including Apple Silicon Macs with 24GB of unified memory and desktop NVIDIA RTX GPUs—Drex Max v1 discards sluggish token-by-token text generation in favor of a single forward-pass probability scoring architecture, handling massive context windows up to 262,144 tokens while slashing agent decision latency by more than 80%.

Why Autonomous Agents Need Decision Models Instead of Chatbots

Modern agent frameworks like AutoGPT, Cursor, and LangChain spend substantial compute budgets generating conversational prose that no human ever reads. For instance, when an agent must choose whether to execute git diff, npm test, or inspect an AST node, an LLM typically outputs 50 to 100 tokens of conversational reasoning before finally outputting the tool call JSON.

Drex Max v1 eliminates this conversational middleman. By formulating the agent's problem as a multi-choice preference evaluation across candidate states, Drex Max accepts the current environment context along with an array of proposed next actions. In a single forward pass, it calculates the softmax probability distribution across the choices. If an action score exceeds a pre-set confidence threshold (e.g. 0.85), the harness triggers execution instantaneously.

from drex import DrexModel

# Initialize locally on Apple Silicon or RTX 4090
model = DrexModel.from_pretrained("nace-ai/drex-max-v1", device="mps", quant="int4")

# Evaluate multi-candidate agent action in a single forward pass
scores = model.score_actions(
    context="Compiler failed: TS2345 Argument of type 'string' is not assignable to 'number'",
    candidates=[
        "run: npm run test",
        "patch: cast value to Number(val) in src/calc.ts",
        "search: grep -rn 'calc.ts' ."
    ]
)

print(scores)
# Output: {"patch: cast value...": 0.942, "search: grep...": 0.048, "run: npm...": 0.010}
# Selected action executed in 38 milliseconds.

Benchmark Comparison: Decision Model vs. Generative LLMs

In benchmark trials measuring autonomous agent decision speed and memory overhead across 500 repository debugging challenges, Drex Max v1 delivered transformative efficiency gains compared to standard commercial frontier models.