IQuest-Q1 320B coding model is a card with architecture and no ship date

The sourced description of IQuest-Q1 320B coding model is a Hugging Face model card, and that card does not carry a release date. This page will not assert one. What the card does carry is a shape. About 320B total parameters. About 15B active. 88 layers. Hidden size 3,072. 48/8 attention heads. Hybrid attention of 3 sliding-window layers plus 1 full-attention layer. Sliding window 4,096. 256 experts with 8 active. Context 524,288. Text only. The card says outputs can be wrong and real CLI tasks may need human oversight.

TL;DR

What actually happened

Two pages, two jobs. The card is the spec sheet. MindStudio is a later writeup that repeats part of the outline and carries an editor and a date. The date on MindStudio is 30 Sep 2026, and the byline is Luis Chavez-Mattos, as editor. That is an edit date on a writeup. It is not printed here as the model’s release date, because the card had none to copy, and this assignment refuses a guessed one.

The size claim is “about,” twice. About 320B total parameters. About 15B active. About means the card is not handing you an integer parameter census in these facts. This page does not replace “about” with a sharper count.

The stack under that size is specific. 88 layers. Hidden size 3,072. Attention heads written as 48/8. Hybrid attention described as 3 sliding-window layers plus 1 full-attention layer. Sliding window of 4,096. Experts: 256, with 8 active. Those are card lines. MindStudio, in the facts used here, repeats the 320B / 15B pair and the 256/8 expert pair. It is not credited with repeating every layer and head figure. If you need 88 layers or hidden size 3,072, cite the card.

Text only is on both. The card says text only. MindStudio repeats the text-only limit. No image claim, no audio claim, and no extra modality is added.

The caution is the card’s, and it fits a coding model more tightly than a parameter count does. Outputs can be wrong. Real CLI tasks may need human oversight. That is not a benchmark score. It is a limitation statement. A writeup that drops it and keeps only 320B has kept the boast and lost the instruction.

Read the card at Hugging Face and the writeup at MindStudio.

What the model card lists

Walk the card in the order a reader can check.

Scale: about 320B total, about 15B active. The gap between those two “abouts” is the card’s way of saying most parameters are not active on a given step. This page does not draw the routing picture beyond the expert counts the card already gives: 256 experts, 8 active.

Depth and width: 88 layers, hidden size 3,072. Heads: 48/8. The slash is the card’s notation. Expanding it into a theory of grouped heads would add a mechanism the facts do not spell out, so the notation stays 48/8.

Attention pattern: hybrid, stated as 3 sliding-window layers plus 1 full-attention layer. The sliding window length is 4,096. The facts give the pattern. They do not give a multiplied tally of how many layers in the stack of 88 are sliding versus full, and this page will not multiply the pattern into a new pair of totals. The pattern is the claim.

Context on the card: 524,288. Text only, repeated because people skip modality. Outputs can be wrong. Real CLI tasks may need human oversight. If the use you care about is a command line, the second sentence is the one the card wants in the room. Human oversight is the card’s phrase, not a policy this article invented.

IQuest-Q1 320B coding model, as a phrase, can tempt a reader to hear “best coder.” The card does not say that. It says a shape, a text-only limit, and a warning. Coding shows up in the caution about real CLI tasks, not in a score.

What a later writeup repeated

MindStudio, edited by Luis Chavez-Mattos, 30 Sep 2026, repeats four things you can lean on as shared: 320B / 15B, 256 experts with 8 active, a context written as 512K, and the text-only limit. Shared does not mean the writeup re-measured the weights. It means the outline was printed again.

The context strings differ on the page, and they should be shown differently. The card says 524,288. MindStudio says 512K. This note does not collapse those writings into one numeral. A reader who sees both can decide whether the shorthand and the integer refer to the same length. The article’s job is to keep each source’s spelling.

What MindStudio is not asked to carry, in these facts, is the full layer list. Do not cite the writeup for 88 layers, hidden size 3,072, 48/8 heads, the 3-plus-1 hybrid pattern, or the 4,096 sliding window unless you have gone back to the card. The repeat list is the short one.

The edit date stays an edit date. 30 Sep 2026 tells you when that page was edited by Luis Chavez-Mattos. It does not tell you the day the weights appeared, because the card had no release date and this piece will not supply one. A URL that says “agentic coding” is a title on that writeup’s path. It is not a score, and it is not a release stamp.

There is no market on this card in the facts used here. No Yes price on a ship date, which is convenient, because there is no ship date to bet. Do not add one.

What to do as a reader (not a trade)

Open the card first. Copy the about-320B and about-15B lines with the word about still attached. Copy 524,288 from the card and 512K from MindStudio without forcing them to look identical. Copy text only. Copy the warning that outputs can be wrong and that real CLI tasks may need human oversight.

Then decide what the writeup is allowed to confirm. It can confirm that a 30 Sep 2026 edit by Luis Chavez-Mattos repeated 320B / 15B, 256/8, 512K, and text only. It cannot confirm a release date the card never stated. If a thread says the model “launched on” a day you cannot find on the card, the thread is ahead of the sources.

This is not a trade. There is no token call, no valuation, and no “use this instead of your current tool” order. A large total parameter count is not a profit story and not a quality score. The card’s own quality comment is the caution, not a leaderboard.

If you are actually going to point the model at a command line, the sourced advice is the card’s: outputs can be wrong, and real CLI tasks may need human oversight. That is a reading instruction. It is not a benchmark, and it is not a promise that oversight makes every command safe. It is the limitation the card bothered to print.

Other notes on this site are different models and different posts. The DogeOS public testnet is not this card. The Gemini 4 Argon launch is not this card. Do not borrow their dates to fill the missing ship date.

Say IQuest-Q1 320B coding model only with the card’s hedges in reach. About 320B, not an audited integer. About 15B active. Text only. Context written two ways, 524,288 and 512K. Experts 256 with 8 active. And no release date on the card. A spec you cannot date is still a spec. Dating it from memory is how a card becomes a rumor.

Use IQuest-Q1 320B coding model as a label for the card, not for a launch day. The architecture block is worth repeating once in compact form so a reader can check the card quickly. 88 layers. Hidden size 3,072. 48/8 attention heads. Hybrid attention: 3 sliding-window layers plus 1 full-attention layer. Sliding window 4,096. If any of those fail to match the card when you look, trust the card and treat this paragraph as stale. The writeup’s job was the short repeat. The card’s job is the long one. Neither job was to name a launch day.