The stack behind the platform

We build on open models and standard interfaces, and spend our own engineering where it changes the economics: fast, exact inference on hardware you control.

Stack

  • Models

    Open-weight models (mixture-of-experts and dense, from compact to frontier scale) for local work; leading LLMs via API where a policy allows it.

  • Compute

    Apple silicon workstations for local inference today. Next: NVIDIA GPUs on-premise and in the cloud, starting with unified-memory systems such as NVIDIA DGX Spark.

  • Cloud

    Target providers for cloud orchestration: Microsoft Azure, AWS and Google Cloud, plus Scaleway and OVHcloud for EU-sovereign deployments.

  • Edge

    Planned: compact models on devices and on-site servers where connectivity is limited or data must stay on the premises.

  • Interfaces

    OpenAI-compatible API, so existing SDKs and agent frameworks work without changes.

PureDrive, our open-source inference engine

PureDrive is a patched MLX runtime with custom Metal kernels for exact, fast local inference on Apple silicon. On one Apple M4 Pro it decodes a 35B-parameter mixture-of-experts coding model at 110 tokens per second at 2K context and 63 at 128K, with greedy output identical to standard decoding.

Fig. 1. PureDrive decode speed on one Apple M4 Pro (35B-parameter mixture-of-experts coder, median of 3 runs, greedy). The baseline is the project’s first lossless configuration. Source: published benchmark on Hugging Face.
ContextPureDriveFirst lossless baseline
2K 110.3 85.8
4K 107.3 85.1
16K 102.6 70.8
32K 95.8 62.8
40K 90.2 60.4
64K 78.3 50.6
128K 62.8 34.6

tokens per second, by context length

Fig. 1. PureDrive decode speed on one Apple M4 Pro (35B-parameter mixture-of-experts coder, median of 3 runs, greedy). The baseline is the project’s first lossless configuration. Source: published benchmark on Hugging Face.

How the platform chooses models

No single model is best for every task. The platform is built to use the smallest model that meets the quality bar for each service, and to escalate only when the task requires it.

  1. I

    Define the bar

    Each service carries its own test cases and checks, written before any model is chosen.

  2. II

    Test small first

    Compact open models are evaluated on local hardware against those checks, for quality, speed and cost.

  3. III

    Escalate by policy

    Only the cases that need it go to larger models or the cloud, within the privacy rules of the service.

  4. IV

    Re-evaluate

    New models appear every month; the same checks decide whether a switch is worth it.

Roadmap

Where we are, and what comes next. We update this page as milestones ship.

  1. Shipped

    PureDrive open source

    Exact, fast local inference on Apple silicon, published with benchmarks and audit tools.

  2. In development

    Platform early access

    API, policy engine and cloud orchestration, for the first early-access organizations.

  3. In development

    Connectors and console

    Document, email and ticketing connectors; usage, cost and quality dashboard.

  4. Planned

    NVIDIA and unified-memory targets

    Porting the exact-speculation method to CUDA systems such as NVIDIA DGX Spark, and to cloud GPU fleets.

  5. Planned

    Edge packages

    Pre-configured deployments for on-site servers and devices in factories, clinics and offices.