IQuest-Q1

Introduction

IQuest-Q1 is a Mixture-of-Experts (MoE) model developed by IQuest for agentic coding, reasoning, and multi-step tool use. It comprises approximately 320B total parameters, with an estimated 15B parameters activated per token.

Model Specifications

Performance

Benchmark Notes

For reproducibility, we recommend using a temperature of 1.0, top-p of 0.95, and top-k of 20, with Claude Code 2.1.140 or Codex 0.142 as the respective harness.

For each model, we report the publicly reported score; otherwise, we evaluate the model using the corresponding benchmark setup: (1) Harness: For agentic coding tasks, we use mini-SWE-agent for DeepSWE v1.1, and Claude Code for other tasks. For Agents' Last Exam, we evaluate our model using Claude Code 2.1.258. Because our models currently do not have multimodal capability, we replace multimodal content inputs in the agent conversation with placeholders during tokenization. (2) Runtime: We set six-hour limits for CyberGym, eight hours for Terminal-Bench 2.1.

Quick Start

Start a server as described in Deployment, install the client with pip install openai, and then call the OpenAI-compatible API:

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="sk-iquest")
response = client.chat.completions.create(
    model="IQuest-Q1",
    messages=[
        {"role": "user", "content": "Hello! Can you briefly introduce yourself?"},
    ],
    temperature=1.0,
    top_p=0.95,
)
print(response.choices[0].message.content)

Deployment

For production deployment, we recommend using SGLang or vLLM.

SGLang

Use our prebuilt image iquestlabworkspace/sglang-iquest-q1:cu130:

docker pull iquestlabworkspace/sglang-iquest-q1:cu130

# without MTP
MODEL_ROOT="$(hf download IQuestLab/IQuest-Q1 --quiet)" && \
python -u -m sglang.launch_server \
  --model-path "$MODEL_ROOT" \
  --served-model-name IQuest-Q1 \
  --tp-size 8 \
  --dtype bfloat16 \
  --attention-backend fa3 \
  --mem-fraction-static 0.85 \
  --disable-prefill-cuda-graph \
  --enable-metrics \
  --tool-call-parser iquest_q1 \
  --reasoning-parser iquest_q1 \
  --enable-torch-compile \
  --load-format fastsafetensors \
  --speculative-use-rejection-sampling

# with recursive MTP
MODEL_ROOT="$(hf download IQuestLab/IQuest-Q1 --quiet)" && \
python -u -m sglang.launch_server \
    --model-path "$MODEL_ROOT" \
    --served-model-name IQuest-Q1 \
    --tp-size 8 \
    --dtype bfloat16 \
    --attention-backend fa3 \
    --mem-fraction-static 0.85 \
    --disable-prefill-cuda-graph \
    --enable-metrics \
    --tool-call-parser iquest_q1 \
    --reasoning-parser iquest_q1 \
    --speculative-algorithm EAGLE \
    --speculative-num-steps 5 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 6 \
    --speculative-draft-model-path "$MODEL_ROOT/mtp" \
    --enable-torch-compile \
    --load-format fastsafetensors \
    --speculative-use-rejection-sampling \
    --speculative-draft-attention-backend fa3

vLLM

Use our image iquestlabworkspace/vllm-iquest-q1:cu130:

docker pull iquestlabworkspace/vllm-iquest-q1:cu130

# without MTP
MODEL_ROOT="$(hf download IQuestLab/IQuest-Q1 --quiet)" && \
vllm serve "$MODEL_ROOT" \
  --served-model-name IQuest-Q1 \
  --tensor-parallel-size 8 \
  --reasoning-parser iquest_q1 \
  --enable-auto-tool-choice \
  --tool-call-parser iquest_q1

# with recursive MTP
MODEL_ROOT="$(hf download IQuestLab/IQuest-Q1 --quiet)" && \
vllm serve "$MODEL_ROOT" \
  --served-model-name IQuest-Q1 \
  --tensor-parallel-size 8 \
  --reasoning-parser iquest_q1 \
  --enable-auto-tool-choice \
  --tool-call-parser iquest_q1 \
  --enable-prefix-caching \
  --speculative-config '{
    "method": "eagle",
    "model": "'"$MODEL_ROOT"'/mtp",
    "num_speculative_tokens": 5,
    "draft_sample_method": "probabilistic",
    "rejection_sample_method": "standard",
    "enforce_eager": false
  }'

Tool Use and Agent Integration

Use a gateway that supports tool calling via Anthropic Messages for Claude Code and OpenAI Responses for Codex.

Claude Code

We recommend version 2.1.140 and the IQuest-Q1[1m] client-side setting does not change 512K context limit.

export ANTHROPIC_MODEL="IQuest-Q1[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="IQuest-Q1[1m]"
export ANTHROPIC_DEFAULT_OPUS_MODEL="IQuest-Q1[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="IQuest-Q1[1m]"
export CLAUDE_CODE_SUBAGENT_MODEL="IQuest-Q1[1m]"
export CLAUDE_CODE_MAX_OUTPUT_TOKENS="131072"
export CLAUDE_AUTOCOMPACT_PCT_OVERRIDE="80"
export CLAUDE_CODE_AUTO_COMPACT_WINDOW="524288"
export API_TIMEOUT_MS="3000000"
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC="1"
export CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS="1"
export ANTHROPIC_BASE_URL="http://example-iquest-q1-link"
export ANTHROPIC_AUTH_TOKEN="sk-iquest-q1"
claude --model IQuest-Q1

Codex CLI

We recommend version 0.142.0.

export OPENAI_API_KEY="sk-iquest-q1"
export MODEL_ID="IQuest-Q1"
# Add /v1 if required by your gateway's Responses endpoint.
export BASE_URL="http://example-iquest-q1-link"

(
  set -eu
  CONFIG_DIR="${HOM