SkyForge · open-weights model hosting
Your models. Your endpoint.
Your GPU.
Launch any open-weights model — GLM, Qwen, DeepSeek, Llama 4, Mistral — on a dedicated GPU with a private, OpenAI-compatible /v1 endpoint and one API key. Billed per second. Terminate anytime.
# launched from the console · boots in minutes
$ curl https://swift-mesa-4821.skyforgecompute.com/v1/chat/completions \
-H "Authorization: Bearer $SKYFORGE_API_KEY" \
-d '{"model":"glm-4-9b","messages":[…]}'
{"choices":[{"message":{"content":"Hello from your own GPU."}}]}01Models
The open-weights frontier, as a spec sheet.
Curated, versioned templates — from a 9B chat model on a single 24 GB GPU to 2.8T-parameter frontier MoE on an 8× 288 GB node. Each launches to its own private endpoint.
| Model | Architecture | Served | Max context | Min GPUs | |
|---|---|---|---|---|---|
GLM-4 9B tier S — single GPU | 9B dense | bf16 | 128K | 1× 24 GB | Launch → |
Qwen3.6 27B | 27B dense | bf16 | 256K | 1× 80 GB | Launch → |
Qwen3.8 27B | 27B dense | bf16 | 256K | 1× 80 GB | Launch → |
Gemma 3 27Bgated | 27B dense | bf16 | 128K | 1× 80 GB | Launch → |
DeepSeek-R1-Distill 32B | 32B dense | bf16 | 128K | 1× 80 GB | Launch → |
Qwen3.6 35B-A3B | 35B MoE · 3B act | bf16 | 256K | 1× 96 GB | Launch → |
Llama 4 Scoutgated tier M — multi-GPU | 109B MoE · 17B act | FP8 | 256K | 2× 80 GB | Launch → |
DeepSeek V4 Flash | 284B MoE · 13B act | FP8 | 256K | 4× 96 GB | Launch → |
Qwen3.8-Flash-Next | 125B MoE · 6B act | FP8 | 256K | 4× 96 GB | Launch → |
Qwen3.5 397B-A17B | 397B MoE · 17B act | FP8 | 256K | 8× 80 GB | Launch → |
MiniMax M3 | 428B MoE · 23B act | FP8 | 1M | 8× 80 GB | Launch → |
GLM 5.3 Flash | 320B MoE · 18B act | FP8 | 1M | 8× 80 GB | Launch → |
DeepSeek V4.1 Flash | 552B MoE · 8–16B act | MXFP4 | 1M | 8× 96 GB | Launch → |
MiniMax H3 | 33B dense omni · video | bf16 | — | 4× 141 GB | Launch → |
Kimi K2.7-Code tier F — frontier cluster | 1T MoE · 32B act | INT4 | 256K | 8× 80 GB | Launch → |
Mistral Large 3gatedlimited supply | 675B MoE · 41B act | FP8 | 256K | 8× 141 GB | Launch → |
GLM 5.2limited supply | 744B MoE | FP8 | 1M | 8× 141 GB | Launch → |
DeepSeek V4 Prolimited supply | 1.6T MoE · 49B act | FP8 | 1M | 8× 288 GB | Launch → |
Kimi K3limited supply | 2.8T MoE · 104B act | MXFP4 | 1M | 8× 288 GB | Launch → |
Text models served with vLLM behind a private, OpenAI-compatible /v1 endpoint; MiniMax H3 serves its video API (/v1/videos) via vLLM-Omni. Plus JupyterLab + PyTorch for your own code.
Full catalog →New — hardware pilot
Mac Studio M5 Ultra, DGX Spark, DGX Station — be first in line.
Desk-side machines with the unified memory to run frontier open-weights models locally. Allocations are first-come.
02How it works
Intent to a live endpoint in three steps.
Pick a model by intent
Chat, coding, reasoning, or your own notebook — the guided flow surfaces the open-weights model that fits, with a plain-English why.
Set a budget tier
Value, Recommended, or Performance. We map it to the right GPU shape and show the per-second rate before you commit.
Call your endpoint
A private, OpenAI-compatible /v1 with one API key across every instance. Point the SDK you already use at it — no rewrites.
base_url = https://<your-slug>.skyforgecompute.com/v1 · api_key = one SkyForge key, every instance
03Pricing
Pay by the second. Stop, and you stop paying.
No minimums, no tier gates, no reserved commitment to start. Live accrued, burn-rate, and projected spend in the console.
The catalog
every GPU type
$0.28/ GPU-hr
From-rates across 20 GPU types — workstation cards to datacenter Blackwell, platform fee included, billed per second.
RTX PRO 6000
inference workhorse
$2.85/ GPU-hr
96 GB GDDR7 per GPU, 1–8 GPUs per node. Launch any Tier-S template on a single GPU — billed per second from ready to terminate.
04Why SkyForge
A private endpoint you actually control.
One operator, not a marketplace
Your job runs on our own fleet, start to finish. It won't vanish mid-run on an unverified third-party host.
OpenAI-compatible, zero rewrites
Every endpoint speaks the OpenAI /v1 API. Switch a base URL and ship — same SDK, same tooling, your weights.
Per-second, no lock-in
Pay by the second and terminate anytime. No minimums, no reserved commitment to get started, live spend in the console.
Open weights you control
GLM, Qwen, DeepSeek, Llama 4, Mistral, Kimi — the model runs on your instance and the key is yours.
05 — Enterprise
Outgrown on-demand? We'll build with you.
Reserved & committed capacity for sustained workloads, allocated through our team with a named point of contact.
Buy-through-us hardware. Few teams hold Supermicro or vendor accounts — we can procure the servers and stand them up for you. In planning; tell us what you need.
Where this is going
From brokered compute to our own datacenter.
Today we run on brokered capacity we source and operate, so you can launch in minutes without waiting on hardware.
In parallel: real land, secured power, modular container deployments, and an open Supermicro purchase order — we build the datacenter as demand proves in. No stranded capacity, and no claims we can't stand behind.
06Early access
Get launch news and capacity updates.
Tell us what you're building and we'll follow up — plus a heads-up as reserved capacity and the hardware path come online.
Launch your first model in minutes.
Pick a model, set a budget, and get a private OpenAI-compatible endpoint with one key. Per-second billing, terminate anytime.