Back in late September, IQuest put the weights for IQuest-Q1 on Hugging Face: 320 billion total parameters, 15 billion activated per token, 524,288 tokens of context. At the time it read like a community changelog entry — open weights, day-0 support in vLLM and SGLang, and then everyone left to reverse-engineer the model card themselves.

What pushed the story to the front was the full technical report, published on October 22. For the first time it laid out results across 8 agentic and reasoning benchmarks, and it earned a slot on the Hugging Face Daily Papers page. More importantly, it explained how the model was actually built.

The most striking numbers are in the long-horizon tasks. IQuest-Q1 scores 83.2 on Terminal-Bench 2.1, with the same field of models spread across 66.0 to 89.1 — normal-looking. Then on version 3.0 it drops to 17.9, and on version 4.0 to 22.7, while Claude Opus 5 posts 42.7 and 53.9 on those same two versions. That does not look like a benchmark being retired. It looks like a model that was only ever tuned against one generation of harness.

The Spec: 320B on the Shelf, 15B in the Loop

According to the technical report from IQuest Research, the model was developed by the IQuest team and is positioned for coding and general-purpose agents in the command line. It carries roughly 320 billion total parameters with about 15 billion activated per token. The usual sparse MoE arrangement is 256 experts with 8 live at a time — roughly 3 percent active. The model card carries the full specification.

The backbone is 88 Transformer layers with a hidden dimension of 3072, 48 query heads, 8 key-value heads, and a head dimension of 128. The attention stack interleaves 3 sliding-window attention layers with 1 full attention layer, the window capped at 4096, and rotary position embeddings are applied to only 32 dimensions per head. The very first Transformer layer drops sparse routing entirely in favor of a dense MLP, which the report gives as a way to mitigate expert load imbalance in the earliest layer of the model.

Training: Stability as a First-Class Concern

Pre-training combined Stable SplitNorm, the Muon optimizer, and muP parameterization. The idea behind Stable SplitNorm is that the residual stream entering the feed-forward network stays unnormalized, buying back the depth utilization associated with Post-Norm; muP handles learning-rate transfer across scale. The paper reports hyper-parameter transfer validated across a 4x range in width and a 4x range in depth, with the search performed on a small proxy model of hidden width 896 and 18 layers and then transferred zero-shot to the target scale.

One controlled comparison is worth remembering: against a Qwen3-30B-A3B-style MoE design using identical training data, the paper reports 1.93x higher scaling efficiency, reaching the same training loss on roughly 52 percent of the training compute. The paper also extrapolates a "critical model size" — about 572 billion by activated parameters and about 12 trillion by total parameters — meaning the recipe retains stability headroom well beyond the scales studied.

Post-Training: Tasks, Harnesses, Four Experts

Post-training is where the real work sits. Tasks are synthesized together with their environments: general-agent tasks run against real APIs, MCP servers, and workspace files, while coding tasks run in executable environments built from repositories. RL uses a multi-harness, multi-version strategy, training one shared policy across Claude Code, Codex, OpenCode and their different versions, with every episode specifying task domain, harness and version, environment, and evaluator. Reward only recognizes verified terminal correctness, and failures are attributed before optimization: errors caused by the policy supply learning signals, while environment or harness faults are masked and evaluator faults block the sample entirely.

RL ultimately yields four experts — agentic user experience, multi-harness work, long-horizon tasks, and general agentic work. Multi-teacher on-policy distillation (MOPD) consolidates those four frozen experts into a single student initialized from the SFT Stage 2 checkpoint, and residual model merging then folds checkpoints across stages and expert branches into the final weights. Reward hacking is defended on three layers: environment preparation, execution-time checks, and trajectory auditing.

Benchmarks: Strong on Coding, Weak on Long Horizons

In the official table: DeepSWE v1.1 at 64.6, SWE-bench Multilingual at 80.3, NL2Repo-Bench at 63.0, CyberGym at 84.5, Terminal-Bench 2.1 at 83.2, GPQA Diamond at 93.3, IMOAnswerBench-Verified at 93.0, and BrowseComp at 85.5. In that same official table, CyberGym comes in above the 83.6 posted by GPT-5.6 Sol, AutomationBench at 52.7 is above the 50.3 from Claude Opus 5, and tau-cubed-Banking at 51.8 is also above the 44.7 from Claude Opus 5.

But set 17.9 on Terminal-Bench 3.0 and 22.7 on version 4.0 against 42.7 and 53.9 for Claude Opus 5, and the gap is hard to explain as a harness generational shift. These are also vendor-reported numbers: the harness and runtime limits differ across models (six hours for CyberGym, eight for Terminal-Bench 2.1, ten for versions 3.0 and 4.0), and one entry, IQuest-CLIBench, is an in-house benchmark.

It Helped Build Its Own Successor

The most telling passage sits in Section 6 of the paper: an agent built on the model iterating on itself under human supervision. It produced 35 tool-call errors in one long-horizon task from mismatched tool names and parameters, then traced failures across 6,054 tool calls, found the training data heavily biased toward a single tool-call schema with some records containing mismatches between tool definitions and call parameters, and proposed cleaning those records while augmenting with diverse schemas. Human experts reviewed the diagnosis before approving the training run. That run used 2.35 billion tokens and 938.7 GPU hours.

There is a practical deployment too: the paper reports that during a three-day deployment, human security experts assisted by the model produced 30 vulnerability reports, all independently validated by fellow security experts, with findings affecting open-source projects responsibly disclosed to their maintainers.

So What

The limitations are stated just as plainly in the model card: text-only input, with no image, audio, or video; tool calls use an IQuest-specific format in the chat template and require compatible serving parsers; and real-world CLI tasks involving iterative debugging often trip the model up, so it may overlook constraints, repeat failed approaches, or leave issues unresolved, necessitating human review.

So the real story is not a leaderboard position but whether the loop in which a model contributes to its own successor actually closes. If a 512K context and 88 layers of hybrid attention mean CLI agents are turning from an application wrapper into a training target in their own right, then the next question is not who scores higher but who first gets that pipeline right. Benchmarks will get new versions and harnesses will get retired, but that coupling between training and evaluation is what this round of engineering is really betting on.