[KongchangAI]
Productstreethub

Streeta

An open-source project that offloads model weights and expert components to memory and SSD in order to run large models on consumer-grade hardware, lowering video memory capacity requirements through tiered storage.

Timeline (last 90 days)

Nov 1

Streeta is an open-source project (the name is also spelled inconsistently as Streeta/Streethub) that offloads the parts of a model that do not fit in VRAM and RAM to an SSD, so that larger models can run on consumer-grade hardware.

Unverified50%
Nov 1

In the Streeta demo, the model's VRAM usage at runtime is about 21G, CPU memory usage is about 60G, and the background shows about 67% of memory in use.

Expired50%
Nov 1

Streeta adopts an offloading strategy: layers or expert weights that are not active in the current computation are temporarily stored in memory and on the SSD, and loaded back into VRAM when needed.

Unverified50%
Nov 1

setup.sh is the environment initialization script of the Streeta project; it is responsible for pulling model weights, installing Python dependencies, configuring the inference backend and starting the service, and the first run requires downloading tens of GB of model files.

Unverified50%
Nov 1

The set-up service exposes a standard Claude Code API and an OpenAI interface, which can be configured for use in Codex or Claude Code, with the model name selected as deepseek v4.1 flash.

Unverified50%
Nov 1

The combination of Streeta plus Cloud Studio is suitable for local setup experience and offline experiments, but not for scenarios that pursue rapid development speed.

Unverified50%
Nov 1

The Streeta solution requires a machine equipped with an SSD and large-capacity memory, and its speed cannot compare with pure GPU inference, but the barrier to entry is low.

Unverified50%

All Facts (6)

Source Articles