// in-browser models

AI models that run in your browser. Vetted, hosted once, cached for every app.

Every model here is an approved parameter file — ONNX, TFLite, a WASM kernel — served from models.skillsafe.ai at an immutable URL. Download it in one SkillSafe app and the next app already has it. Inference happens on your device, nothing you enter is uploaded to run it, and it never costs a credit. Builders declare a model on a release and get the download disclosure, the licence notice and a verified loader for free.

7models
6tasks
4apps running them
58.6 MBof weights, hosted once
Checking your browser… your device
Request a model Apps that run in-browser AI Build with the registry

Runs in your browser

MB · downloaded once, then cached
Task
Runs on

Every file is a vetted parameter file served from models.skillsafe.ai at an immutable URL, so a visitor who loaded a model in one app already has it for the next — and "Preload" above warms that same shared cache. There is no charge: the bytes are on the platform's bill and the compute is the visitor's device. Apps declare what they load with models on a release, load it with ss.models.load() (verified against the catalogue hash) and ship their own tokenizer and config files — see the deploy guide.

Why a shared registry

01

Vetted bytes, not a URL

Every file is approved by exact SHA-256 after a structural audit of the graph or module — the same allowlist that gates what an app may ship. A withdrawn hash answers 410 everywhere at once.

02

Hosted once, cached for every app

One immutable URL per file, credential-free and byte-identical for every app, so the browser keeps a single copy across *.skillsafe.ai. Model bytes never count toward a builder's bundle cap or storage quota.

03

Loaded efficiently

ss.models.load() streams with progress, verifies the hash, keeps a durable copy in the app's Cache API and tells you whether it was a download or a cache hit. Edge caches hold files up to 500 MB per region.

04

Disclosed and attributed

Declaring a model on a release puts "downloads N MB · runs on your device" on the app page and generates /models.txt from the licence's attribution — Llama and Gemma notices included, without hand-writing them.

Requested models

0 open · vote for what gets vetted next

No open requests yet — propose the model your app needs.

Propose a model

Sign in to request a model for the registry and vote on others' requests. Duplicate asks for the same repository merge into one, so demand adds up.

Have the weights already? Publishers can file the exact bytes for vetting with POST /v1/vetting-requests — see the deploy guide.

Runs on SkillSafe servers

Provider list price · USD per 1M tokens

GPT-5.6 Sol

gpt-sol gpt-5.6-sol
OpenAI
$4.00 in / 1M tokens · list
$20.00 out / 1M tokens · list

GPT-5.6 Terra

gpt-terra gpt-5.6-terra
OpenAI
$2.00 in / 1M tokens · list
$12.00 out / 1M tokens · list

GPT-5.6 Luna

gpt-luna gpt-5.6-luna
OpenAI
$0.20 in / 1M tokens · list
$1.20 out / 1M tokens · list

Gemma 4 26B (Workers AI)

gemma-fast @cf/google/gemma-4-26b-a4b-it
Workers AI
$0.10 in / 1M tokens · list
$0.30 out / 1M tokens · list

GPT Image 2.5 Flare

gpt-image gpt-image-2.5-flare
OpenAI
~$0.060 list, per 1024×1024 image

Model cost is pass-through: these are the providers' published list prices, unchanged — see OpenAI pricing and Workers AI pricing. What SkillSafe charges for is the computation and API handling on its own servers: a 10% platform fee on each run's base — the model's list cost, with no per-job overhead. A creator's markup is a separate cut of the same base — they keep 100% of it. Only the latest model of each tier is shown; configure the alias and your app follows upgrades. Tiers that cannot run on this deployment are not shown.

How server pricing works

  • Pay for what a run actually uses. Each run is charged on the tokens the model actually consumed, rounded up to whole credits (minimum 1 credit ≈ $0.0001). There is no fixed per-job overhead, so a short turn on a cheap model bills a handful of credits.
  • The model is pass-through; the fee pays for the servers. A run's base is the provider's list cost — charged exactly as the provider publishes it, no markup and no overhead. SkillSafe adds a 10% platform fee on that base for the computation and API handling it does on its own servers; the creator takes their markup — up to 100% — of the very same base, shown on each app's detail page. At the default 10% the two are equal, and nothing is deducted from the creator's side: they keep 100% of their markup.
  • BYOK apps bill the 1-credit minimum ($0.0001) per run. When a publisher brings their own provider key, their key pays the provider directly and the platform charges only the minimum billable unit — no 10% platform fee and no creator markup.
  • Spend is capped per job. Every model carries hard caps on input tokens, output tokens, and wall-clock time, so a single run can never overrun its hold.
  • Image models bill per image. An image-generation run produces one 1024×1024 image. The hold reserves the worst-case cost of that image and the run settles down to the provider's actual (usage-reported where available) cost — same base, same two cuts as text runs.
  • No model configured? Apps without a model run on the Workers AI default — cheap, keyless, and never silently billed at premium rates.
Turn your skill into an app Browse apps See plans