> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Chinese fund Ubiquant's AI lab releases IQuest-Q1, an open model for coding agents
- URL: https://globalfeed.ai/en/chinese-fund-ubiquants-ai-lab-releases-iquest-q1-an-open-model-for-coding-agents/
- Published: 2026-09-30T03:45:08.000Z
- Updated: 2026-09-30T03:45:08.000Z
- Description: IQuest, the AI unit of Chinese quantitative fund Ubiquant, released IQuest-Q1 for coding agents and multi-step tool use. The model works with Claude Code and Codex CLI and runs on your own servers, but it needs eight high-memory data center GPUs.
- Author: GlobalFeed Editor
- Tags: IQuestLab, IQuest-Q1, open models, coding agents, x-iquestlab, x-huggingface, dil-en, elle

**IQuest** released [IQuest-Q1](https://huggingface.co/IQuestLab/IQuest-Q1?ref=globalfeed.ai) on Hugging Face and GitHub on September 28, 2026\. The team, listed as IQuestLab on Hugging Face, is the AI research unit of the Chinese quantitative investment fund Ubiquant and previously released the IQuest-Coder family of coding models. According to the model card, IQuest-Q1 is built for agentic coding, reasoning and multi-step tool use, and it is offered as a model you can run on your own servers.

## A model for coding agents

IQuest-Q1 is a Mixture-of-Experts (MoE) language model. It has 320 billion parameters in total; about 15 billion of them are active for each token, with 8 of 256 experts selected. The context window is 512K tokens (524,288), which suits long code repositories and long agent sessions. The architecture mixes sliding-window and full attention layers, with three of every four layers using a sliding window. A multi-token prediction (MTP) layer drafts tokens ahead to speed up generation. The model accepts text only; there is no image, audio or video input.

The card walks through using the model inside **Claude Code** and **Codex CLI**. The server exposes an OpenAI-compatible API, and a gateway that supports Anthropic Messages for Claude Code and OpenAI Responses for Codex is needed. This lets developers point the command-line agent they already use at a model running on their own servers.

## Hardware and setup

At full precision (BF16) the model takes about 650 GB. The setup commands on the card assume a server that splits the model across 8 GPUs; since the load per card exceeds 80 GB, that means high-memory data center cards of the H200 or B200 class. The model does not fit on a single GPU, and IQuest has not released a quantized version. The recommended serving tools are SGLang and vLLM: [SGLang support](https://github.com/sgl-project/sglang/pull/41590?ref=globalfeed.ai) was merged into the main project on September 29, and for vLLM the company provides its own fork and ready-made Docker images. The card has no support information for llama.cpp or Ollama.

The license is an MIT variant that IQuest calls "Modified MIT". Commercial use is allowed; according to the [license text](https://huggingface.co/IQuestLab/IQuest-Q1/blob/main/LICENSE?ref=globalfeed.ai), the only added condition is that if the model or a derivative is used in a commercial product or service, "IQuest-Q1" must be displayed prominently in its user interface.

In the company's charts, IQuest-Q1 mostly trails Claude Opus 5 and DeepSeek-V4.1-Flash on coding and agent tests such as DeepSWE, NL2Repo, CyberGym and Terminal-Bench, and comes close to them on CyberGym and Terminal-Bench. The same charts place it ahead of models such as DeepSeek-V4-Flash and GLM-5.3-Flash on some tests. For companies and research teams that want to run coding agents in their own data center without sending code outside, and that have hardware at this scale, it is a new option.

IQuest writes on the model card that IQuest-Q1 is still at an early stage: it may overlook constraints in real command-line tasks, repeat failed attempts and need human oversight, and the card recommends verifying generated code with tests. The Codex example configuration on the card turns off approval prompts and the sandbox, so it is worth tightening that setting for your own environment. All benchmarks come from the company, and one of them (IQuest-CLIBench) is an in-house test. To see how the model does on real work, it is worth following evaluations from independent researchers.

Sources: [Hugging Face: IQuest-Q1 model card](https://huggingface.co/IQuestLab/IQuest-Q1?ref=globalfeed.ai), [GitHub: IQuestLab/IQuest-Q1](https://github.com/IQuestLab/IQuest-Q1?ref=globalfeed.ai), [GitHub: SGLang support for IQuest-Q1](https://github.com/sgl-project/sglang/pull/41590?ref=globalfeed.ai).