# Muse Spark 1.2, Qwen3.8-Max, DeepSeek-V4-Flash: naming has stopped helping

- Published: 2026-07-23
- Authors: CORTEXA
- Category: Opinion
- HTML: https://researchhub-vert.vercel.app/blog/model-naming-versioning-mess

Muse Spark 1.2, Qwen3.8-Max, DeepSeek-V4-Flash, GPT-5.6 Sol. Six releases in a week, and not one name tells you what was actually run.

One week of releases: Muse Spark 1.2, Qwen3.8-Max, DeepSeek-V4-Flash, GPT-5.6 Sol, Claude Opus 5, Kimi K3.

## The problem is not aesthetics

A name should let a reader recover what was used. These do not. Is Max larger than Flash? Is 3.8 newer than V4? Does Sol denote a capability tier or a release train? You cannot tell without a table, and the table is not in the paper.

## Why it damages research

Papers cite model names. A reader in two years, holding "Qwen3.8-Max", cannot determine which weights ran — the endpoint has moved on and the name was never a version.

This is a reproducibility problem wearing a marketing costume.

## What would fix it

A date and a hash, alongside the marketing name. `qwen3.8-max @ 2026-07-30 / sha…` costs nothing and makes a citation recoverable.

Until then the burden falls on authors: record the date you ran it and, where you can, pin open weights. Nobody else is going to do it for you.

## What a version identifier is for

A version exists so that a reader can recover what ran. That requires it to be immutable, ordered, and unambiguous.

Current model names are none of these. They are reassigned as the underlying weights change, they carry no reliable ordering, and the same name serves different capability tiers across providers.

## The specific ambiguities

Is Max larger than Flash, or just more expensive? Is 3.8 newer than V4, when they come from different labs with unrelated numbering? Does a suffix denote a capability tier, a release train, or a marketing distinction?

You cannot answer any of these without a lookup table, and the table is not in anyone's paper.

## Why this is a research problem, not an aesthetic one

Papers cite model names. Two years later a reader holding "Qwen3.8-Max" cannot determine which weights produced the result — the endpoint has moved on, and the name was never a version in the first place.

This makes a large body of published results permanently unverifiable, and the failure is invisible at the time of writing.

## The minimal fix

A date and a content hash alongside the marketing name. `model-name @ 2026-07-30 / sha256:abc…`. It costs nothing to publish and makes a citation recoverable.

Providers will not do this voluntarily, because ambiguity is commercially convenient — it lets a name accrue reputation while the thing behind it improves.

## What authors can do unilaterally

Record the date you ran it. Record any version string the API returns. Where the result is load-bearing, use open weights at a pinned revision, and say so.

None of that requires cooperation from a vendor, and it is the difference between a citation and a gesture.

## The counter-argument

One could argue that for most applications the exact weights do not matter — you want the current best model, whatever it is called. That is fair for products. It is not fair for a paper claiming a measurement.