Parallax Changelog
Issue 15 · live Subscribe
Menu
← Changelog № 15 · September 21, 2026

OPEN WEIGHTS

Open models are four months behind, not years

Kimi K3 took 84 days to reach the frontier OpenAI had in April. You can download it free. Almost nobody has.

The primer

Epoch AI puts open models four months behind the closed frontier. The gap is small, but it has stopped shrinking. We show what those four months hide.

PublishedSeptember 21, 2026
Reading time5 min
Sources13 cited
DeskChangelog
The dekAnd the four months is the cheap part.
— 01
WHAT YOU ASSUME

Two measurements, one answer

Four months behind, and the gap has stopped shrinking.

What most people think

Open models are hobby projects. The paid ones are a generation ahead, and closing that gap will take years.

What the data shows

4months

Epoch AI, a research group that scores models on hard tests, measures the best open model four months behind the best closed one.

A second measure asks which model people prefer. The closed lead there grew from 0.5% to 3.3%.

Four months behind, and the gap has stopped shrinking.

In plain terms Two panels: on the left what most people assume, on the right what the numbers show, with the one figure that settles it.Source · Epoch AI, open-closed capability gap · May 2026

— 02
THREE WORDS FIRST

What the labels actually mean

Four months, then. Three words decide what those four months are worth, and one of them is doing work you did not agree to.

Open weights
You get the finished model file to download and run. That is the cake, not the recipe.
Open source
Weights plus the code and the data that made them. Almost no large model ships all three.
ECI
Epoch AI's capability index. One score from many hard tests, so two models rank on the same scale.
— 03
FIVE DATES

Three months behind, then four

Two readings of the same gap, seven months apart. In between, the day an open model caught April's best closed one.

2025-10-30
Epoch AI reads the gap
Three months behind, averaged across nearly three years.
2026-04-23
GPT-5.5 · ECI 158.22
The closed frontier sets a mark.
2026-05-29
Epoch AI reads it again
Four months now. The gap widened, it did not close.
2026-07-16
Kimi K3 · ECI 158
Announced 16 July. Weights released 27 July.
April's frontier, reached 84 days later
2026-09-03
GPT-6 Astra · ECI 166
Eight points ahead again.

Three months behind in October 2025. Four months behind in May 2026.

In plain terms The key events in order down a rail, with coloured dots flagging the pivotal moments.Source · Epoch AI data insights and model pages · September 2026

— 04
INSIDE THE AVERAGE

Eight points, spread very unevenly

Eight points ahead on the index, four months ahead in time. Those are two readings of the same gap: the points say how far behind, the months say how long the catch-up took. Only these four tests score both models on the same terms.

Near parity on graduate science. On a puzzle game the open model scores 26% against 84%. Epoch AI warns this reading may understate the gap, because open models are tuned hard on public tests.points
Mystery Game Puzzles Open 26% · closed 84% 58 Fifty-eight apart on a puzzle game
Chess Puzzles Open 39% · closed 72% 33
SimpleQA Verified Open 51% · closed 76% 25
GPQA Diamond Open 93% · closed 96% 3 Three points apart on graduate science

Near parity on graduate science. On a puzzle game the open model scores 26% against 84%. Epoch AI warns this reading may understate the gap, because open models are tuned hard on public tests.

In plain terms Each bar is the gap between two scores on one test, the best open model against the best closed one. Longer bar, bigger gap.Source · Epoch AI model pages · September 2026

— 05
WHO IS ACTUALLY THERE

Not the labs you have heard of

Those gaps run from three points to fifty-eight. So who releases these open models? Three Chinese labs: DeepSeek, Moonshot, which makes Kimi, and Zhipu, which makes GLM. Each has one column below, and a line joins every model to the one built from it.

Eight of these ten releases are open weights, and the newest landed in July.history
  • deepseek-v3-2 DeepSeek · V3.2 open
  • deepseek-v4-pro DeepSeek · V4-Pro, MIT licence open
  • deepseek-v4-flash DeepSeek · V4-Flash, MIT licence open
  • kimi-k2-thinking Moonshot · Kimi K2 Thinking open
  • kimi-k2-6 Moonshot · Kimi K2.6 open
  • kimi-k3 Moonshot · Kimi K3, 2.8T open
  • glm-5 Zhipu · GLM-5 open
  • glm-5-1 Zhipu · GLM-5.1 open
  • gpt-5-5 OpenAI · GPT-5.5 closed
  • gpt-6-astra OpenAI · GPT-6 Astra closed

Eight of these ten releases are open weights, and the newest landed in July.

In plain terms Each row is one release. A line joins a model to the one built from it, and the colour marks open weights against closed.Source · Epoch AI Benchmarking Hub model pages · September 2026

— 06
THE WHOLE STACK

Free is only the top layer

Eight of those ten releases are free to download. Nothing under them is, so the rest of the stack stays somebody else's bill, and that kharcha does not come with the file.

Training data is almost never released, and the compute to serve a model is not downloadable at any licence.

In plain terms Five slabs, stacked. The top one is what you can download today. Each slab below it is another thing the word open could mean.Source · Stanford HAI, Hugging Face open-model census, AI Now Institute · September 2026

Training data is almost never released, and the compute to serve a model is not downloadable at any licence.architecture
Weights Downloadable today. Kimi K3, DeepSeek V4, Qwen. The layer everyone means.
Licence Permissive for 81% of Chinese releases above 20B. Only 29% of American ones.
Code Inference code usually. Training code rarely.
Training data Almost never released. You cannot see what the thing was trained on.
Compute and serving Not downloadable at any licence. The cluster is the real bill.
— 07
WHAT IT COSTS

Three prices for the same million tokens

The one layer you can rent has a price. Here it is, per million tokens, which means the small chunks of text a model reads and writes.

A million output tokens costs $50 from the closed frontier and $15 from the best open model.

GPT-6 Astra, the closed frontier

$50per million output tokens

Ask the same question in an Indian language and it costs about five times the English price.

  • The same million from the best open model, $15 Less than a third of the frontier price

  • From DeepSeek V4-Flash, $0.28 The cheapest open model on the list

A million output tokens costs $50 from the closed frontier and $15 from the best open model.

In plain terms One number, large, and beside it the everyday things it equals, so the size can be felt rather than read.Source · Epoch AI model pages · September 2026

— 08
WHO TOOK IT

The offer nobody took

Renting the open model costs less than a third of renting the closed one, and downloading it costs nothing at all. One count of state-backed AI projects found 139, across 56 countries. About 40% picked Llama, Meta's open-weight model. Not one picked a model from the labs actually at the open frontier. Think of a free textbook nobody opens because the shelf sits in another building.

That means the four months is not what keeps people away. The bottom slab of that stack does, the compute to run the model, and no licence makes that part free. Of all downloads, 83% go to models under a billion parameters, the smallest on offer. India's pull is different. Sarvam AI adapts open-weight models to Indian languages, work you can only do when you hold the weights instead of renting them by the token.

Source · Rest of World, Hugging Face open-model census · September 2026

The four months is free. Almost nobody takes it.

Reader signal

How did this land?

A private signal to the editor. Pick as many as apply — nobody else sees it, and it shapes what runs next.

Letters

Sixty days of argument.

Longer responses from readers. The editor reads each one before it appears here.

No letters yet Newest first

Be the first to write one. Letters open for sixty days after an issue is published.

Write a letter

0 / 2000 · reviewed before it appears

Sources

13 cited.

Every number above traces to one of these. Primary sources first.

  1. 01 The gap between open and closed models epoch.ai Epoch AI · primary
  2. 02 Epoch Capabilities Index (ECI) epoch.ai Epoch AI · primary
  3. 03 Kimi K3 — model page epoch.ai Epoch AI · primary
  4. 04 GPT-6 Astra — model page epoch.ai Epoch AI · primary
  5. 05 AI Index Report 2026, Chapter 2: Technical Performance hai.stanford.edu Stanford HAI · primary
  6. 06 Open-Weight Models Aren't Enough for Open Science hai.stanford.edu Stanford HAI · analysis
  7. 07 The State of Open Models, Summer 2026 huggingface.co Hugging Face · primary
  8. 08 China's Moonshot gave away Kimi K3. Sovereign AI programmes are not taking it. restofworld.org Rest of World · secondary
  9. 09 India's frugal AI bet: Sarvam, Krutrim and the sovereignty question restofworld.org Rest of World · secondary
  10. 10 The Openness Imperative ainowinstitute.org AI Now Institute · analysis
  11. 11 DeepSeek V4 simonwillison.net Simon Willison · secondary
  12. 12 Kimi K3: the open-weights escalation interconnects.ai Interconnects · analysis
  13. 13 H.10 Foreign Exchange Rates federalreserve.gov Federal Reserve · primary
0% · 5 min left
Press S

One email. Two things.

The dispatch when each issue ships — and a free shelf that saves your place. One click in your inbox does both.

Enable JavaScript to subscribe.