Community

Since our October 20 update, giz.ai has a new default model that runs on our own GPU, turns pictures into 3D objects, makes longer videos, and adds voice typing.

Qwen3.8 27B Fast is the new default model

Agents and Auto now start on Qwen3.8 27B Fast, served by giz.ai on our own GPU. It is Swift 1.5, a Qwen3.8 27B tuned to reach its answer without overthinking, with a 262K context, image input, and tool use for coding and agent work.

• Strong on terminal tasks. On Terminal-Bench 2.1, the agentic terminal-coding benchmark, Swift 1.5 scores 72.1% (published by its developer, five runs per task). GPT-6 Luna, our previous default, scored 57% to 60% in our own Terminal-Bench 2.1 run with the Codex agent.
• Much faster. In our tests on the same ta

Hey quandale Dingle over here

Hi Grandma

You like that new song i have out called imlosin it?

what have u seen

Hi

I am willing to fight you and will win

hey

hey

How to make A Tesla Roadster

Jump of a cliff

Hello.

On November 8, the hourly limit for DeepSeek V4.1 Flash 1017 (120 requests) and the 20 enabled Qwen3.7 Plus calls were exhausted very quickly, after which the requests went to credits. Please help me figure out if this is normal for my use case.

What I checked on my end, based on the client logs:

Prompt caching is not working. The cache_read_tokens field is zero in every single call, including repeated steps within the same session. Each step resends the entire context.
Usage increases with dialog length. Across 18 sessions on DeepSeek, the median input is 149,000 tokens per step, with up to 254,000 in long sessions. The input-to-output ratio is about 400 to 1.
The main volume is not the dialog text, but the resending of tool results and tool schemas. The schemas of three connected