Skip to content
Research · October 2026

The thinking, on your Mac.

Sting is our own engine for running AI on an ordinary 16 GB MacBook. For the two jobs Smart Jelly does most, finding what you owe and noticing when it’s done, it comes close to the cloud, in under a second.

See the results Coming to the app
81%Finding what you owe: matches the cloud’s answer.the cloud matches itself 93% · the same model untrained 66% · 200 test conversations
92%Noticing when it’s done: matches the cloud’s answer.the cloud matches itself 99% · untrained 65% · 100 test conversations
0.6 sTo read a conversation for tasks.the cloud call takes 2.1 s
16 GBAn ordinary MacBook with an M4 chip.no server, no data centre
Results

Nearly as good as the cloud, on the jobs that matter.

Each job is asked exactly the way Smart Jelly asks it today. The score is how often the answer matches the cloud model the app uses. The white line is the cloud matched against itself: the most anyone can score.

StingThe same model, run the usual wayThe cloud, matched against itself
Finding what you oweThe tasks in a conversation that are yours to do.200 test conversations · 4 languages
The cloud matches itself 92.6% of the time
Noticing when it’s doneWhich open tasks the new messages finish.100 test conversations · 4 languages
The cloud matches itself 98.9% of the time
Does this need a reply?A yes or no for each new message.14 real messages · a small first check
The cloud matches itself 85.7% of the time

Scored on test conversations no model learned from. Ranges and how we measured are in the research notes below.

Speed

Faster than asking the cloud.

No round trip to a server. Median times on a 16 GB MacBook with an M4 chip, against the app’s cloud call for the same job.

Read a conversation for tasks
Sting0.6 sCloud2.1 s
Does this need a reply?
Sting0.1 sCloud4.1 s
Is this a group message, not for you?
Sting0.1 sCloud1.2 s
Memory

It builds your memory too.

The same small model, taught each job, now does the work that builds your memory: who and what your conversations are about, and how your day splits up. It is up against two models nearly seven times its size, on the same laptop.

People and projectsWho and what a day of conversations is about: the people, projects and organisations.86 requests from real use
The cloud matches itself 90.4% of the time · times are medians on a 16 GB M4
How your day splits upWhere one piece of work ends and the next begins.75 requests from real use
The cloud matches itself 86.6% of the time · times are medians on a 16 GB M4

95% ranges, by request: people and projects, Sting 74% to 84%, Saluki 69% to 78%; how the day splits, Sting 44% to 81%, Bonsai 52% to 79%. The day-splitting ranges are wide, so those three are close to even.

Compared

A small specialist against a big generalist.

On 7 October, Underdog released Saluki: a 27-billion-parameter model squeezed to fit a 16 GB laptop. We ran it on the same MacBook, with the same requests and the same scoring. Sting runs a model about a seventh the size, taught each job, and answers several times faster.

MeasureSaluki · 27B · same 16 GB M4Sting · 4B · same 16 GB M4What it means
Noticing when it’s doneSaluki · 27B · same 16 GB M488.9% (81–96)Sting · 4B · same 16 GB M491.8% (84–98)What it meansSting scores higher; the ranges overlap.
Finding what you oweSaluki · 27B · same 16 GB M486.1% (80–91)Sting · 4B · same 16 GB M481.1% (75–87)What it meansA statistical tie: the ranges overlap.
People and projectsSaluki · 27B · same 16 GB M473.3% (69–78)Sting · 4B · same 16 GB M479.6% (74–84)What it meansThe small model, taught the job, is ahead.
Reading a conversation for tasks, medianSaluki · 27B · same 16 GB M45.1 sSting · 4B · same 16 GB M40.68 sWhat it meansAbout 7.5 times faster.
Noticing when it’s done, medianSaluki · 27B · same 16 GB M45.9 sSting · 4B · same 16 GB M40.9 sWhat it meansAbout 6.5 times faster.
People and projects, medianSaluki · 27B · same 16 GB M428 sSting · 4B · same 16 GB M45.4 sWhat it meansAbout 5 times faster.
The modelSaluki · 27B · same 16 GB M427 billion parameters, about 7.9 GBSting · 4B · same 16 GB M44 billion, at 4 bits, plus a small add-on per jobWhat it meansA model about a seventh the size.

Bold marks the side ahead on that row. Scores are agreement with the cloud model the app uses, with 95% ranges. Both measured by us, on the same machine.

How it works

Most local AI runs any prompt. Sting runs one job.

It is written for one model, one kind of Mac and one task at a time. Knowing the job ahead of time is what makes a small model good enough.

Built for one modelHand-written code for the graphics chip, shaped to one 4-billion-parameter model (Qwen3.5-4B, at 4 bits). No Python and no server: a small library made to ship inside the app.
One pack per jobEach job comes as a pack: the model, a small add-on trained for that one job, the job’s instructions prepared once and kept in memory, and the app’s own safety checks.
Yes or no in one stepFor a yes-or-no question, Sting reads the answer straight from the model in a single pass instead of writing it out: about 0.1 seconds.
Stops earlyMost conversations have nothing to report. Sting can tell after one pass, and stops there.
Same answer, every timeEvery change to the engine is checked word by word against a reference. A change that alters an answer is a bug, never a speed-up.
Careful when unsureThe checks drop a task nobody asked of you, and won’t close a task without proof. When in doubt, a task stays open.
Compared

Sting and Husky.

Husky, from Underdog, is the fastest engine for one model on Apple silicon that we know of. Husky is built for speed on any prompt. Sting is built for the right answer to one job, on the laptop people already own.

MeasureHusky · M5 Max, 128 GB · their figuresSting · M4, 16 GB · measuredWhat it means
Checking 8 drafted words, cost vs 1Husky · M5 Max, 128 GB · their figures1.5×Sting · M4, 16 GB · measured1.03–1.1×What it meansSting checks a draft almost for free, so every accepted word is nearly pure gain.
A yes or no for a jobHusky · M5 Max, 128 GB · their figuresnot offeredSting · M4, 16 GB · measured~90 msWhat it meansSting reads the answer in one pass; there is no text to write.
“Nothing to report”Husky · M5 Max, 128 GB · their figureswrites the answerSting · M4, 16 GB · measuredone pass, then stopsWhat it meansMost chats hold no task; Sting stops as soon as it knows.
Scored on the app’s real jobsHusky · M5 Max, 128 GB · their figuresnot publishedSting · M4, 16 GB · measured81% / 92% vs cloud 93% / 99%What it meansEvery Sting number here is scored against the cloud model the app uses.
Words kept per read, copy-heavy jobHusky · M5 Max, 128 GB · their figures5.4Sting · M4, 16 GB · measured2.3–3.3What it meansHusky trains its drafter; Sting drafts from the conversation itself.
Plain writing speed vs MLXHusky · M5 Max, 128 GB · their figuresup to 4.5×Sting · M4, 16 GB · measured~1× (31 vs 33 words/s)What it meansBoth are limited by memory speed; Husky’s code gets closer to the limit.
Reading a new promptHusky · M5 Max, 128 GB · their figures5,150 tokens/sSting · M4, 16 GB · measured275–310 tokens/sWhat it meansThe M5 has matrix hardware the M4 lacks; Sting keeps job instructions in memory instead.

Bold marks the side ahead on that row. Husky’s numbers are Underdog’s published figures on different hardware; Sting’s are measured on a 16 GB M4.

Research notes

How we measured it.

The jobsEach request is built by Smart Jelly’s own code, in the app’s exact layout, and scored against the cloud model the app uses for that job.
The test conversations300 conversations written for this test, in English, Hindi, Hinglish and Marathi-English: work, family, small business and college chats, with real tasks, traps and noise. No real person’s messages. No model learned from them.
The scoreF1 over the tasks found or closed, after the app’s own safety checks. Most conversations have nothing to report, so answering “nothing” every time scores 0.
RangesFinding tasks: 81% (75% to 87%). Noticing when done: 92% (85% to 97%). 95% ranges, by conversation.
The machineA 16 GB MacBook with an M4 chip. Greedy decoding, thinking off. Times are medians.
Same answersSting matched the reference implementation, running the same model and add-on, on 116 of 119 test requests; the other three differ where two next words score almost the same.
The memory jobsRequests from the owner’s own use, with banking and payment pages removed. Each was answered by a pack that never trained on it. Scored as F1 over the people, projects and organisations found, and over where the day splits.
SalukiUnderdog’s Saluki ran in llama.cpp, its intended runtime, on the same 16 GB MacBook, with the same requests and scoring, thinking off as Underdog recommends for fast answers. Saluki is by Underdog, Apache 2.0, built on Qwen3.8-27B and ISTA-DASLab’s 2-bit release.
Next

What we’re working on.

How you workTeaching Sting procedural memory: your routines, how you write to each person, and what you want Jelly to do differently.
A trained drafterDrafting from the app’s own answers, to keep 5 or more words per read.
Faster codeFrom about 70% to about 90% of the chip’s memory speed.
Newer chipsPrompt code for the M5’s matrix hardware.
Coming soon

On your Mac, in Smart Jelly.

Sting is coming to Smart Jelly as a third way to run it. For your to-dos, your conversations will be read on your Mac and nothing will be sent to a server. Everything else keeps working the way you choose today.

Download Smart Jelly today macOS 14+ · Apple silicon · On-device mode will need 16 GB of memory
Our relayToday · all featuresOnly the few lines a question needs leave your Mac.
Your own keysToday · all featuresThe same lines, sent straight to your own AI account.
On your Mac, with StingComing soon · your to-dosFinding and closing what you owe, with nothing sent anywhere.