Model Card: IQuest-Q1

🌏 中文版

Model Details

Key Capabilities

• 320B total parameters with only 15B active (8 of 256 experts routed per token), scoring 84.5% on CyberGym (real-world CVE remediation) — second only to DeepSeek-V4.1-Flash's 88.1%, and ahead of GLM-5.3, DeepSeek-V4-Pro, and Hy4-preview
• A 524,288-token context window with native support for Claude Code and Codex CLI — swapping models only needs an environment variable change (e.g. ANTHROPIC_MODEL="IQuest-Q1[1m]"); the [1m] is just a client-side label, the actual context ceiling stays at 512K. It's positioned as a self-hosted alternative to the two mainstream coding-agent tools
• At inference, a single recursive MTP (multi-token prediction) layer runs 8 times paired with EAGLE speculative decoding, keeping latency down despite the 15B active-parameter budget
• The training and R&D pipeline lets the model itself take part in capability diagnosis, training-plan design, and parts of experiment execution, with humans reviewing only at key checkpoints — research direction changes, high-cost experiments, version adoption. IQuest calls this the model "taking part in its own development"

Benchmark Results

⚠️ All figures are IQuest's own self-reported benchmarks (harness: mini-SWE-agent for DeepSWE v1.1, Claude Code 2.1.140 / Codex 0.142 for the rest; a 6-hour cap for CyberGym and an 8-hour cap for Terminal-Bench 2.1), with no independent third-party reproduction yet.

Versus Predecessor / Competitors

IQuest-Q1 is IQuest's first public model, so there's no predecessor to compare against — it can only be placed on the existing board of open-weight agentic-coding models. At 15B active parameters, it sits on the lighter end of its generation — DeepSeek-V4.1-Flash, GLM-5.3, and Hy4-preview are all larger or similarly sized open models — yet IQuest-Q1 lands near the top of the pack on CyberGym at 84.5% (just behind DeepSeek-V4.1-Flash's 88.1%), and its 83.2% on Terminal-Bench 2.1 trails Hy4-preview's 85.4% closely. A smaller active-parameter budget doesn't appear to cost it much task execution capability.

Against the strongest closed model on the board, Claude Opus 5, the gap is still clear: 12.3 points behind on NL2Repo, 10 points behind on JobBench, and 5.9 points behind on Terminal-Bench 2.1. In other words, IQuest-Q1's position is "solid upper-mid-tier among open-weight models of its size," not a challenge to the closed-model ceiling.

What stands out is the team itself: IQuest had no prior public model track record, yet its first release shipped weights, inference code, and training methodology (MOPD, Multi-Teacher On-Policy Distillation) together. That kind of complete, "all at once" release is unusual for a lab this new.

What It Means for Agent Development

This model positions itself as a self-hosted, open-weight alternative to Claude Code or Codex CLI, not another general-purpose chat model.

• If you're building a coding agent that needs data to stay on-premises (finance, government, regulated industries): IQuest-Q1 is natively compatible with Claude Code's Anthropic Messages interface and Codex's OpenAI Responses interface — swapping models only requires changing environment variables and the gateway, not rewriting agent logic
• If you're running real-world CVE remediation or security-audit style agentic tasks: 84.5% on CyberGym is an uncommonly high score among open-weight models, worth evaluating as a replacement or supplement to your existing toolchain in a self-hosted setup
• Not a fit: workloads needing multimodal input (this checkpoint is text-only — the team is explicit that it has no image, audio, or video input capability), or teams without something like 8x H200-class GPUs for tensor parallelism — the BF16 weights run about 640GB, so the self-hosting bar is not low
• If you're still deciding whether to trust a model from a brand-new team: IQuest published its full training pipeline and a description of the "model takes part in its own development" R&D process, which is a point in favor of trust — but the lack of independent reproduction is still a risk, so validate against your own task set before shipping to production

Today's Takeaway

A lab with no public model history three days ago shipped weights, inference code, training methodology, and a description of a process where the model participates in diagnosing its own training — all at once, on its first release. That's a level of completeness that doesn't lag behind teams with years of release history. It's a reminder that judging whether a new model is trustworthy isn't just about "how high did it score this time" — how much reproduction detail the team is willing to publish may be the more stable signal.

References

• IQuest-Q1 Technical Blog (with benchmark data)
• HuggingFace: IQuestLab/IQuest-Q1
• GitHub: IQuestLab/IQuest-Q1 (inference code and deployment guide)
• HuggingFace: IQuest-Q1 LICENSE (Modified MIT License)
• Pandaily: IQuest Research Open-Sources IQuest-Q1
• NetEase (in Mandarin): a dark horse open-source model makes a striking debut