Model news · 2026-10-24
Kimi K3.1 Pro: open weights by November 4
Moonshot AI introduced Kimi K3.1 Pro, a post-trained Kimi K3 with a 1M-token context. Today it is API-only. Weights land by November 4, and that day we try to run it on 8 B300 GPUs.
First release · macOS Apple Silicon
The best AI models, under your control.
SophyAI picks, starts and publishes open-weight models on your Mac or on cloud GPUs rented by the hour.
The agents and editors you already use talk to an OpenAI endpoint that exists only on 127.0.0.1.
How it works
1 · Catalog
Measured recipes
Each entry is model + inference engine + hardware, with measured hourly cost and speed. Choose by budget, VRAM or quality.
2 · Run
Local or cloud
On your Mac with Ollama, ds4 Metal or Colibrì. When you need a large model, a GPU on Vast.ai starts, works and shuts down.
3 · Publish
Active models catalog
One OpenAI endpoint on 127.0.0.1:8800/v1: several sessions and local models together, selected by ID (sophy/… or ollama/…). Cursor and OpenCode never see the provider.
4 · Agents
MCP with revocable keys
AI agents read the catalog, sessions and costs with read or manage scoped keys that expire.
From whitelist to app in three steps
1
Ask for access
Submit a whitelist request. We review each one by hand and reply by email.
2
Activate your account
If accepted, choose your password from the email link and download the app. The Pro trial starts.
3
Choose and pay
Max, Pro or Team, monthly or yearly. Payment goes through Stripe; the license reaches the app online.
Architecture
Desktop app
Native app (Tauri) for Mac. It holds the offline catalog, chat, the local gateway and model management. Windows and Linux come later.
Control plane
A Docker service that keeps accounts, licenses, wallet and cloud sessions. The app connects with a personal key; the monthly license is signed and can be verified offline.
Model NAS
A private store of model weights, checked with SHA-256 against the originals. A wrong checksum and the model won't start.
Two separate bills, no surprises
App license
Free, Max, Pro or Team: it sets how many hours a day you use the app. You pay us, monthly or yearly.
Cloud GPU
Pay as you go: a registered provider key (Vast.ai) and credit in your wallet. Every session shows its cost before it starts and stops when you stop it.
From the blog
Model news · 2026-10-21
DeepInfer, the engine DeepSeek uses to serve its own models
DeepSeek publishes DeepInfer, the inference framework behind its own production. This release covers DeepSeek-V4.1-Flash offline, on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.
Model news · 2026-10-21
LFM2.5-24B-A2B, a 24B agent model with 2.45B active
Liquid AI publishes LFM2.5-24B-A2B on Hugging Face. It is a reasoning model with 24 billion parameters, 2.45 billion active, a 131,072-token context, and the LFM Open License v1.0.