Drex local serving needs a license check
November 1 update: Drex Max v1 on Apple silicon
The public Drex Max v1 weights are a separate 27B decision model: they return probabilities for supplied choices rather than generate prose. Its repository history predates this update, with October 30 additions for Q4, Kapitsa and MLX. This is a deployment update, not a November 1 launch.
| Local path | Documented requirement | Input limit |
|---|---|---|
| Affine Q4 | Apple silicon, at least 24 GB; Kapitsa 0.2.25+ or MLX | Text only |
| Q8_0 GGUF | 48 GB Mac; documentation recommends 2,048-token context on that machine | Text only |
| Python | NVIDIA GPU; about 49 GB for BF16 weights or 26 GB for the 8-bit backbone | Text and images |
The Q4 card distinguishes its 24 GB support statement from the example actually measured: a 48 GB M5 Pro, context 2,048, on October 30. Its comparison covers 21 probability fields. That is not a broad accuracy-equivalence result or a measurement on a 24 GB machine. Download size is not total runtime memory.
Use the model-specific loader and documented Nace forks for llama.cpp or Ollama. A normal causal chat run does not reproduce this decision model’s behavior. Check the chosen runtime, checkpoint and request schema together before planning a trial.
The Max v1 weight license has use restrictions. Check its current terms before deployment; the historical discussion below is not blanket permission.
Nace’s current hosted offering is labeled Max v1.1. Its API prices, latency claims and benchmark comparisons should not be treated as measurements of these local v1 weights. ByteForward has not installed or tested these local paths.
Original October 10 review: Drex v1.5
OpenRouter lists Drex v1.5 with an October 9 release date. On October 10, its displayed DeepInfra endpoint costs $0.04 per million input tokens and $0 for output. These are listing observations, not a guarantee of future pricing or service.
Nace announced Drex September 24 and updated v1.5 September 28. Repository commits span October 9 and 10.
Those dates make this a useful moment to examine deployment choices. Commit timestamps establish when repository changes were recorded. They do not establish when a model first became available. For developers, the important question is what they would need to operate, maintain and evaluate before putting a decision system into an application.
What local control entails
Drex scores supplied choices. Local runners include Python and Nace’s llama.cpp and Ollama forks. The documented backbone is distilled Qwen 3.5 9B.
A fork adds a practical maintenance question. Which version will the team pin, who will track changes, and how will it compare answers after an update? A familiar runner name alone should never substitute for checking the exact build. The first evaluation should record the runtime, model files, hardware and request format together so another developer can reproduce the comparison.
At the listed input rate, one million requests averaging 1,000 input tokens would imply $40 of input charges. This is simple arithmetic, not a complete operating budget. It excludes any other applicable charges and does not estimate local hardware, staff time or the cost of incorrect decisions.
Read permission before planning deployment
The code uses Apache 2.0. Weight terms are modified Nace Open RAIL M. They restrict models, derivatives and outputs above US$1 million of prior twelve month gross revenue or total equity or debt funding, including employers and affiliates. Those thresholds exempt personal use or research. Supplying products or services competing with Nace or its affiliates triggers a separate restriction. These restricted uses require a separate written license. Fully automated decisions with legal or similarly significant effects on individuals require meaningful human review.
Treat that permission check as an early dependency. A prototype that later becomes part of a customer product deserves a fresh review of its intended use. This is a summary of selected terms, not legal advice or a determination that any particular deployment is allowed. Read the full agreement before downloading or using the weights.
Test decisions and confidence separately
Nace reports 58.08 from its own run on Decision Index 0.3.1 across 37 public benchmarks, excluding private tests. Drex, Jev and Nimble fall within its cited 0.9 point tie band.
Calibration documentation explains a separate question. Among cases assigned a probability near 0.8, does the relevant outcome occur about 80 percent of the time? A model can return probabilities without those numbers being reliable confidence estimates. Threshold selection then decides which score should trigger an action, using the application’s objective and the costs of different errors.
Consider a hypothetical support queue with billing, technical and general categories. A wrong category might add a few minutes of work, while silently closing a ticket could lose a customer. Evaluate those actions separately. Include ambiguous requests and an escalation path, and inspect the cases the system handles confidently but incorrectly. An average score can hide the particular mistake that matters most.
For an authorized trial, keep the evaluation cases separate from the cases used to adjust thresholds. Compare both errors and the share sent for human review. Count how much review remains at an acceptable error level, then include that work in the cost comparison. That is a more useful deployment result than simply obtaining a valid response.
ByteForward reviewed the public documentation and license files. It did not install the runners, execute the model or reproduce the vendor’s benchmarks. The reported score provides context for an evaluation, with no guarantee of performance on a reader’s workload.
