Whallm runs language models larger than your Mac’s memory on Apple Silicon. It keeps common weights in memory and reads experts from SSD as needed. Includes a chat interface and an OpenAI-compatible API for clients such as Codex.

• Apple Silicon Mac, macOS 15 or later.
• 64 GiB memory recommended; actual use depends on the model and settings.
• Fast SSD: internal, Thunderbolt or USB4. Model weights are downloaded separately; the app checks required free space, including partial downloads.

DeepSeek V4.1 needs about 458 GiB for experts and Engram alone, plus other weights. Qwen sizes include image weights. See model settings and input limits for details.

1. Download Whallm-macOS-arm64.zip from GitHub Releases, unzip it and open Whallm.app.
2. Open Model, pick a model and select Download Model. The app checks free space first. If a download stops, you can resume it.
3. Open Server and select Start Server.
4. Open Chat and pick your model, or connect an API client as shown below.

By default the server runs at http://127.0.0.1:11434. A model loads the first time you use it. The server keeps one model loaded and handles one request at a time. If a model you just installed is missing from the Chat list, restart the server.

A user measured Swift1.5-Qwen3.8-Flash-Next on v1.1.13 with a 16K-token input and a 512-token answer. Prefill is how fast Whallm reads your input; Decode is how fast it writes the answer. Both are in tokens per second.

Automatic mode chooses the expert cache size; the smaller fixed cache used here trades speed for lower memory use. Slots are the number of expert cache entries.

Reference only: user-provided measurements, without the prompt type, full settings, cache state or complete environment. Not a controlled comparison between versions. Peak memory is reported process memory, not total Mac memory.

Full results and test instructions. Use Throughput in the app to test your own Mac.

Start Whallm's server, then add this to ~/.codex/config.toml:

model = "deepseek-v4-flash-0731"
model_provider = "deepseek-v4-ssd"
model_reasoning_effort = "high"

[model_providers.deepseek-v4-ssd]
name = "Whallm"
base_url = "http://127.0.0.1:11434/v1"
wire_api = "responses"
requires_openai_auth = false

Set model to an API model ID, or to an Alias you set in Model → Advanced Settings. Then restart Codex.

Find API model IDs with GET /v1/models or in the model ID table. Display names are not request IDs.

This example uses the default address and no API key. If you set a key in Whallm, set the same key in your client.

See the Codex configuration reference.

• One loaded model and one request at a time. Long prompts need more cache memory.
• MiMo is a preview; video and audio output are not supported. Qwen accepts still images, not audio, video or documents.
• Tool calls are executed by your client, not Whallm. Full Pi/Codex sessions and broad model quality are not fully validated.
• Speed depends on SSD, prompt and settings; acceleration is not faster for every prompt.

Privacy: inference runs on your Mac. Downloads, updates and API connections use the network; connected clients may send data elsewhere. Debug logs omit full prompts, tool results, images and API keys, but can contain local paths and network addresses. Check logs before sharing.

• Model settings, API IDs, image limits and API reference
• Complete benchmarks and historical results
• Release notes and upgrade steps
• MiMo preview status

Whallm uses the MIT License. Model weights have their own terms. Whallm is not affiliated with DeepSeek or Qwen.