Information Technology · Tuesday, 18 August 2026

01 · Briefing · what happened

Alibaba's new laptop-sized model claims to match one ten times bigger

Information Technology 5 min 22 sources

The industry's main cost lever right now is shrinking AI models onto ordinary hardware. This week also showed where the shrinking stops working.

10x

size gap Alibaba claims to close

a 27-billion-parameter model, vendor claim

53.8%

hard agent tasks finished

by a model topping the leaderboards

151,448

Qwen-based versions downloaded

2.6 times Meta's total, per Hugging Face

69%

of firms use half their chips or less

and under half track the cost at all

At a glance

Alibaba released Qwen3.8-27B, built to run on a laptop, and says it matches a model ten times its size.
It also opened the weights of its most powerful model, days after Meta open-sourced its own and shipped laptop models.
Alibaba is defending a lead: Hugging Face counted 151,448 Qwen-based versions, 2.6 times Meta's total.
The week also showed the floor. DeepSeek's leaderboard-topping cheap model finished 53.8% of hard agent tasks, and its prices went up.
Gartner expects agent workflow costs to more than quintuple by 2028 even as model prices fall.
The squeezing has moved to the code around the model: Writer's researchers cut costs 40% by changing the harness, not the model.
Two-thirds of surveyed firms run AI in production; fewer than half can measure what it costs.

Forces in play

Shrink-it pressure High↑

Alibaba, Meta and Nvidia all shipped laptop-sized models inside eight days

The quality floor Building↑

the cheap leaderboard leader finished only 53.8% of hard real-world jobs

Price war Steady→

Google halved Gemini Flash for now, DeepSeek put its prices up

Cost blindness High→

fewer than half of firms track what their AI compute costs or returns

China's open-weight lead High→

Qwen versions outnumber Meta's 2.6 to one on the main download site

In play Alibaba — shipped a laptop-ready model and opened its top model's weights Meta and Nvidia — answered a week earlier with their own open and laptop-sized models DeepSeek — topped leaderboards, then raised prices and stumbled on real tasks Gartner — forecasts agent costs rising fivefold despite cheaper models Writer and Kog — squeezing cost out of the plumbing rather than the model

How it unfolded

Last week Meta open-sources its top model and ships laptop models; Nvidia does the same
Thu DeepSeek launches V4 Pro and raises prices on both its models
Sun Independent testing puts the cheap leaderboard leader at 53.8% on hard tasks
Mon Alibaba answers with Qwen3.8-27B and opens Qwen3.8 Max's weights

Where this points

Watch whether independent testing backs the ten-times claim on real multi-step work. If it does not, the race moves from model size to the code around the model.

Full briefing

Alibaba puts the frontier on a laptop

On Monday Alibaba released Qwen3.8-27B, a model built to run on consumer hardware like a laptop [1]. The company says it matches the performance of a model ten times its size, and handles coding, research and long multi-step work [1]. That is a vendor claim, not an independent result.

Alibaba also opened the weights of Qwen3.8 Max, its most powerful model [1]. Weights are the numbers that decide how a model behaves. Opening them lets anyone download and run the thing, though the training data and methods stay private [1].

The timing is the story. Last week Meta open-sourced its most powerful model and launched Muse Glimmer, a family designed for laptops [2]. Nvidia released open models the same week, and is separately building a trillion-parameter model called Nemotron 4 [2][3]. More than 20 US tech companies had asked Washington not to put “premature restrictions” on open-weight models, even Chinese ones [2].

Alibaba is defending a lead, not chasing one. Hugging Face, the main place developers download open models, counted 151,448 Qwen-based derivatives last week - 2.6 times Meta’s total [1]. Two analysts told CNBC the next fight is on-device: models running on your phone or laptop instead of in a data centre [1].

Where the squeeze stops working

Shrinking is the industry’s main cost lever right now. This week it hit resistance in three places.

DeepSeek’s V4 Flash has topped leaderboards since July. Put on real work, it finished 53.8% of a batch of deliberately hard agent tasks [4]. The testing firm Composio ran it through eight different tool setups on 30 multi-step jobs across Gmail, GitHub, Slack and Sheets [4]. Of 240 runs, 129 passed, and only six of the 30 jobs were completed by every setup [4]. The same model behaved very differently depending on the plumbing around it [4]. DeepSeek then raised prices on both V4 Flash and the larger V4 Pro, and added peak and off-peak rates [5]. Flash has 284 billion parameters; Pro has 1.6 trillion [4].

Gartner expects the cost of agent workflows to more than quintuple by the end of 2028, even as models get cheaper [6]. The reasoning is that cheaper tokens tempt people to build heavier workflows, which burn far more of them. Routing one task to a reasoning model raises inference costs at least fivefold, Gartner says [6]. Inference is what it costs to run a finished model to answer something, as opposed to training it once.

So the squeezing has moved to the code around the model. Writer launched Palmyra X6, built by further training Z.ai’s open GLM-5.2, and says the package cuts customer costs by up to 50% on basic tasks [7]. Its own researchers found that changing the harness - the code that feeds and steers the model - cut costs 40% on average, more reliably than changing the model [7]. French startup Kog is chasing the same gain on GPUs firms already own, claiming 3,000 tokens a second per request on standard AMD and Nvidia data-centre chips [8].

Prices at the cheap end keep falling anyway. Google shipped Gemini 3.7 Flash three weeks after 3.6, at a temporary $0.75 per million input tokens and $3.75 per million output [9]. From 1 January 2027 both double [9]. OpenAI added an “Ultrafast” mode it says runs GPT 5.6 Sol at 14 times the speed, up to 750 output tokens a second [10].

Almost nobody can see whether any of it pays. Across 170 enterprises surveyed, two-thirds run AI in production, but fewer than half track rigorously what their compute costs and returns [11]. Of those running their own GPUs, 69% report using half their capacity or less [11].

A quieter piece this week made the same trade visible in a place everyone can see. Ultra HD Blu-ray can devote up to 108 megabits a second to video on a 66GB disc, and 128 on a 100GB one [12]. A stream throws far more of the picture away, because it has to keep playing as your connection wobbles [12]. Same pixel grid, different amount of information kept.

The chip side of the same trade

SMIC, China’s largest contract chipmaker, passed $3 billion in quarterly revenue for the first time [13]. Average selling prices rose 5.7% and wafer shipments 14% [13]. Profit more than tripled from a year earlier [14].

Investors were harder on the specialists. Cerebras fell 16% after results; its hardware revenue dropped to $54.1 million from $70.3 million a year ago [15]. Groq raised $350 million at a $3.5 billion valuation, roughly half what it was worth a year ago, before Nvidia struck a licensing deal with it [16]. And Stripe has reportedly agreed to buy OpenRouter, the gateway that routes developer traffic across more than 400 models, for over $7 billion [17]. If that closes, the switch between models is worth more than most of the models.

Regulators had a week too

A German regulator ordered Apple to stop designing its data-consent pop-ups in ways that steer people away from third-party apps [18]. US District Judge James Donato gave Google one week to strip “anticompetitive friction” out of downloading rival app stores inside Google Play [19]. The US Federal Trade Commission opened an antitrust investigation into Epic Systems, which holds medical records for hundreds of millions of Americans [20]. And a French press association asked its competition watchdog to act against Google’s AI article summaries, which it says take readers away from newspapers [21].

The archive that nearly went down with its host

Nine PBS, a public television station in St. Louis, has spent five months trying to get back 50 terabytes covering 70 years of local broadcasting [22]. Its cloud storage provider stopped answering in February, let the contract lapse on 6 March, and cut off access without the 30-day retrieval window the contract allowed [22]. The station traced the servers to an Iron Mountain data centre in Colorado and sued [22]. A judge has now ruled the station owns the data and cleared it to retrieve it [22].

02 · Lesson · why it matters

Why nothing can shrink everything

A compressor saves space by betting on what your data looks like, and every bet it wins on one file it loses on another.

How it works

A file is only long because it is unpredictable
A compressor spends few bits on what it can guess
That is a bet about what your data looks like
Win the bet on one file, lose it on another
Once the repetition is gone, nothing is left to remove
Push past that floor and real information goes

The twist

No scheme can shrink everything, because there are fewer short names than long ones - so every file a compressor makes smaller forces another one bigger.

Where you've seen this

Photos and music

a JPEG or MP3 shrinks by dropping detail your senses rarely notice, and stops working on noise

Summarising a meeting

you can cut the repetition for free; past that you start losing what someone actually said

Zipping a folder

text collapses to a fraction, already-compressed video barely moves and sometimes grows

The catch

The floor is set by the data, not the cleverness - and what counts as predictable was decided by whoever chose the training set.

Full lesson

The claim hiding inside the announcement

Alibaba’s new model is small enough to run on a laptop, and the company says it does the work of a model ten times its size. Meta said something close about its laptop family a week earlier. Nvidia is building both a trillion-parameter giant and something you could run at your desk.

Strip the branding and there is one claim underneath all of it: most of what was in the big model was not needed. That is a claim about information, and information has laws.

What making something smaller actually means

Take any file. It is long because it is unpredictable. Compression works by spending very few bits on the parts a machine can already guess, and saving the long descriptions for the parts it cannot.

English text collapses to a fraction of its size because English is repetitive. The letter after a “q” is almost always “u”, so you barely need to write it down. A photograph of a wall compresses hard because most of it is the same grey. A recording of static barely compresses at all, because nothing in it predicts anything else.

So a compressor is not a machine that removes waste. It is a bet about what your data is likely to look like.

Every winner forces a loser

Here is the part that surprises people. No compression scheme can shrink every possible input, and this is not a limit of current cleverness. It is arithmetic.

Count the short descriptions. There are far fewer of them than there are long files. If a scheme maps some inputs to shorter outputs, and every output must map back to exactly one input, then some other inputs have to get longer. There is nowhere else for them to go.

The bet is always paid for by somebody. A tool tuned for text makes video slightly bigger. Try to zip a folder of photographs and you will watch it happen: the files barely move, and a few come out larger than they went in.

The floor is set by the data, not by the effort

Once a scheme has removed the repetition, there is a hard limit under which it cannot go. That floor is a property of the data itself, not of the engineer. A file with genuine variety in it needs a certain number of bits, and no amount of work finds a shortcut that is not there.

This is why “we made it smaller” is only half a sentence. Smaller than what, and by throwing away which part? A streaming service and a Blu-ray disc show the same film at the same pixel count. One keeps several times more of the picture. The grid is identical; the information is not.

Where the shrinking starts to cost

A model is a file too. Cutting its precision, or training a small model to imitate a big one, follows the same shape. Early on the savings are close to free, because a large model carries a great deal of redundancy. Past a point the cuts stop landing on repetition and start landing on things the model actually knew.

That point is hard to see from a leaderboard, because leaderboards test the common cases, and common cases are exactly what survives compression. It is the rare, awkward, multi-step job that reveals what went missing. This week, a cheap model that had topped the rankings finished barely half of a set of deliberately difficult real tasks.

Whose data counted as typical

There is a choice buried in every compressor, and it rarely announces itself. Somebody decided what “typical” data looks like, and everything unusual pays for that decision.

A speech system trained mostly on one accent will compress that accent beautifully and mangle another. A shrunk model keeps what appeared often in the data it was measured against and quietly drops what did not. The person on the wrong side of that line does not get told the scheme was not built for them. They are told their case is unusual.

That arrangement can be perfectly reasonable and still uneven. Serving the common case is how anything gets cheap enough to be useful. Cheapness is real: it puts a capable model on an ordinary laptop instead of behind a data centre bill. Both things are true at once.

What we are all inside

The pull toward smaller reaches further than the labs. It sets which model answers your bank’s chat window, what a rural clinic can run on hardware it can afford, what your phone can do without a signal. Every one of those choices is somebody deciding which details are rare enough to lose.

None of us sees the whole ledger. The engineer sees the benchmark, not the person whose case fell off the edge. The user sees an answer, not the parts that were never in the file. The rest of us find out only when the missing piece happens to be ours.

03 · Lab · your turn

Spend the Short Codes

Rehearse the compression bet and feel why a scheme that shrinks one file must grow another.

04 · Hope · carry this

There is a floor under every file, and it holds: what is genuinely there cannot be squeezed away. Seventy years of one city's television came back from a company that no longer exists.