1. Get an approved evaluation build
Ask your organization for Razor Manager and Razor Terminal.
Razor SDK 0.4.0 · Local inference
Razor is a C++ inference SDK for Apple Silicon and Android. Run models on device, with CPU inference on both platforms and GPU inference on Apple Silicon.
Razor for Mac
The Mac apps put model setup and chat in two familiar windows. Razor Manager shows which model files are ready. Razor Terminal opens the selected model for chat and computer actions.
Ask your organization for Razor Manager and Razor Terminal.
Open the downloaded folder and drag each app into Applications.
Open Razor Manager, choose a chat profile and check its file status.
Select Open in Razor Terminal. Model weights are downloaded separately from their listed sources.
Current SDK terms allow evaluation within the licensee's organization. They do not allow public distribution or sale of a product containing Razor. A one-time Mac app purchase is planned after commercial distribution rights are available.
Supported model families
Published results cover five configurations across three model families. Other profiles are available in the CLI and may use external runtimes or remain experimental.
SmolLM2
Published fp32 and int8 results. Quality and speed vary with model size and weight format.
Qwen3
Instruction chat support with tool calling through the configured chat format.
Gemma 3
Published results include fp32, int8 and int4 formats. Gemma model terms apply.
Measured performance
SmolLM2 135M decode results below use greedy generation. Weight formats and hardware differ by row. Treat the values as workload measurements.
| Device and runtime | Model and format | Decode |
|---|---|---|
| Apple M4 Max · Razor CPU | SmolLM2 135M · int8 | 155 tok/s |
| Apple M4 Max · Razor MLX | SmolLM2 135M · MLX | 563 tok/s |
| Snapdragon 8 Gen 3 · Razor CPU | SmolLM2 135M · int8, cold | 151 tok/s |
| Snapdragon 8 Gen 3 · LiteRT | SmolLM2 135M · q8 | 64 tok/s |
Apple GPU comparison: 64-token prompt, 128 generated tokens. Android result: cold device. Results change with prompt length, model format and device temperature.
Evaluation
On a 68-question factual suite with identical prompts and greedy decoding, Razor scored 58 correct and MLX-LM scored 60. The reported 95% confidence intervals overlap. Model and weight format affect output quality.
Modes
One command covers chat, tools, coding, shell work, memory across sessions, diffusion models, images and voice. Some profiles use external runtimes or remain experimental, and reply quality depends on the model.
A conversation in the terminal with more than 20 slash commands. Switch models in the middle of a conversation.
Eight built-in tools, plus your own in OpenAI format. Razor shows each command and its risk and asks before it runs.
Project-aware file and command tools for coding models. Edits and commands wait for approval.
Small models turn a plain request into a shell command. Piped input joins the prompt.
Razor archives a project's chat on disk and recalls relevant earlier turns, within the model's context limit.
Masked-diffusion language models with six pass budgets, exact-length replies and infill.
Multimodal profiles take a picture, a clip or a recording. A speech profile replies with audio.
Optional fast decoding on eligible models and backends. Standard decoding stays available.
Give a workload and limits. Razor predicts which models fit on this device before you download them.
Describe a task and Razor ranks the models you have, with the command to start each one.
Using Razor
Razor provides a C++ SDK and command-line tools for running supported language models on Apple Silicon and Android arm64. Published validation covers SmolLM2 135M, 360M and 1.7B, Qwen3 0.6B, and Gemma 3 1B. Other profiles have separate availability and validation status.
Tool calls work with configured model formats. The model determines whether it can produce reliable tool requests.
Disk-backed history retrieves relevant earlier turns into the prompt. The model's configured context window still applies.
Inference runs on the device. Razor supports CPU inference on Apple Silicon and Android, with MLX GPU inference on Apple Silicon.
The 30-day evaluation includes the SDK and benchmark harness. Model licences apply to their weights.
Request an evaluation