No one runs open models more efficiently than Dyson — so no one runs them cheaper. Find a lower published rate for the same model anywhere, and we'll match it, then hand you a $100 credit for the tip.
# one CLI. both orbits. $ pip install dyson # call a model by the token… $ dyson run deepseek-v4-flash "summarize" ✓ 312 tok · $0.0006 · 340ms # …or ship your own container to a GPU $ dyson deploy ./app.py --gpu h100 ✓ live in 142ms · scales to zero when idle
The inference market has bifurcated — token APIs racing cost-per-million to the floor on one side, serverless GPU clouds fighting over cold-start milliseconds on the other. Pick the API and you can't ship a custom container. Pick the cloud and you're wiring up model serving yourself.
Send a prompt, get an answer — but you're locked to their menu of models.
Bring your own code and weights — but you own the serving stack and the scaling.
Serverless inference for DeepSeek, Llama, Qwen and GLM — OpenAI-compatible, per-million pricing, zero contracts.
$0.08 / M IN · $0.16 / M OUTShip a Dockerfile or define the environment in Python. Custom fine-tunes and agentic workflows the endpoint can't hold.
GPU PODS FROM $0.59 / HRMemory-snapshot restore boots a warm GPU container in under 150ms. Your endpoint costs nothing between requests.
<150MS · SCALE-TO-ZERORun hundreds of Low-Rank Adaptations concurrently on one base model — custom behavior without reloading weights.
FP4 / FP8 · H100 · B200Burst into your own AWS, GCP or bare metal for data sovereignty. SOC 2, ISO 27001, and a zero-retention I/O policy.
BYOC · ZERO-RETENTIONRequests route to the nearest of 31 regions in real time. Per-second billing, integrated logs, versioned rollbacks.
31 REGIONS · SUB-SECOND AUTOSCALEThe disruptors are excellent — at exactly one half of the job. Coverage synthesized from public platform documentation.
| Capability | Dyson | DeepInfra | Fireworks | Together | Runpod | Modal | Beam |
|---|---|---|---|---|---|---|---|
| Per-token model API | ✓ | ✓ | ✓ | ✓ | – | – | – |
| Bring-your-own container | ✓ | – | – | – | ✓ | ✓ | ✓ |
| Scale-to-zero billing | ✓ | api | api | api | ✓ | ✓ | ✓ |
| Sub-200ms cold start | <150ms | – | – | – | ✓ | ✓ | ✓ |
| Multi-LoRA on one base | ✓ | ltd | ✓ | ltd | – | ✓ | – |
| Bring-your-own cloud (BYOC) | ✓ | – | – | – | – | – | ✓ |
| Zero-retention · SOC 2 / ISO | ✓ | ✓ | part | part | part | ✓ | part |
A bursty workload keeps a GPU idle 70–90% of the time. On an always-on plan you fund every one of those idle seconds.The efficiency play
Dyson restores from a memory snapshot in under 150ms, so scaling from zero is invisible to your users — and the meter only turns while a token is actually being generated.
One API key covers token endpoints and your own GPU containers. No contracts, no idle billing, no infra team.