IQuest-Q1 Runs Agentic Coding From a 320B MoE With Only 15B Active

IQuest Research published open weights for IQuest-Q1, a Mixture-of-Experts model aimed at agentic coding, reasoning, and multi-step tool use—about 320 billion total parameters with an estimated 15 billion activated per token, a 524,288-token context window, and documented serving paths into Claude Code and Codex CLI. The Hub card at IQuestLab/IQuest-Q1 and the companion repository github.com/IQuestLab/IQuest-Q1 are the grounding primaries. The LICENSE file is a modified MIT grant dated 2026: commercial products or services that use the Software (or derivatives) must prominently display “IQuest-Q1” in the user interface.

Architecturally the card is specific. The model uses 88 transformer layers, a 3,072 hidden dimension, 48 query / 8 key-value heads at head dimension 128, and a hybrid attention pattern of three sliding-window layers plus one full-attention layer, with a 4,096-token sliding window and partial RoPE on 32 dimensions. The MoE block exposes 256 experts and activates 8 per token. Multi-token prediction (MTP) is part of the ship: two independent MTP layers in training and a recursive ×8 setup at inference with a 512-token MTP sliding window, intended for speculative decoding. Modality is text-only—no native image, audio, or video path.

Serving is multi-GPU by design. IQuest documents SGLang and vLLM with prebuilt CUDA 13.0 images (iquestlabworkspace/sglang-iquest-q1:cu130 and iquestlabworkspace/vllm-iquest-q1:cu130), typically launched with eight-way tensor parallelism, bfloat16, and custom --tool-call-parser iquest_q1 / --reasoning-parser iquest_q1 flags. Optional EAGLE-style speculative decoding points a draft model at the checkpoint’s mtp subdirectory. For reproducibility the card recommends temperature 1.0, top-p 0.95, and top-k 20, with Claude Code 2.1.140 or Codex 0.142 as the respective harnesses when comparing agentic runs.

Harness integration is a first-class part of the release. Claude Code setup points Anthropic-compatible environment variables at a gateway that speaks IQuest’s tool-call format; the card notes that a client-side [1m] model suffix does not change the model’s 512K context limit. Codex CLI setup writes a provider block for the OpenAI Responses wire API with matching context and autocompact limits. The sample Codex configuration on the card sets approval_policy = "never" and sandbox_mode = "danger-full-access"—operators should treat those as examples to revise, not production defaults.

Evaluation notes on the card matter as much as architecture. Agentic coding tasks use mini-SWE-agent for DeepSWE v1.1 and Claude Code for other listed tasks, including Claude Code 2.1.258 for Agents’ Last Exam; runtime caps are six hours for CyberGym and eight hours for Terminal-Bench 2.1. Multimodal content in agent conversations is replaced with placeholders during tokenization because the checkpoint has no multimodal capability. Vendor charts compare against DeepSeek-V4-Flash/Pro releases and an in-house IQuest-CLIBench; those figures live in Hub-page images rather than machine-readable tables, so this piece does not invent numeric leaderboard claims beyond the prose.

Limitations are stated bluntly. Generated explanations and code can be wrong and need task-appropriate tests. Tool calls require IQuest-specific parsers and a compatible harness. Real-world CLI work may overlook constraints, repeat failed attempts, or leave issues unresolved. The authors call the release early-stage with “substantial limitations” and “a long way to go.”

The sparsity math is the product story. Total parameters set capacity; active parameters set per-token compute. Routing eight of 256 experts keeps decode closer to a mid-size dense model while the full expert bank still resides across a tensor-parallel group. Hybrid attention further shapes long-context cost: three of every four layers use a 4,096-token window, so only periodic full-attention layers pay full-sequence attention when transcripts stretch toward the half-million-token ceiling. MTP and EAGLE-style draft verification target interactive coding where long tool-call sequences make tokens-per-second feel like product quality.

Builders should read the license and the harness samples together. The modified MIT attribution requirement is unusual among recent open coding MoEs and matters for any UI that wraps the model. The Codex sample that disables approvals and sandboxes is a poor copy-paste default: it shows how IQuest expects the model to be driven inside existing agent CLIs, not how a secure deployment should look. The intended user is teams with multi-GPU nodes who already operate Claude Code or Codex-style gateways and want an open, long-context coding MoE behind the same UX, with humans still reviewing diffs.

For Writeble’s opensource desk, IQuest-Q1 is a large sparse coding MoE with explicit agent-harness glue and a commercial attribution clause—not a silent Hub dump.