Open models,
the day
they ship.
- 01Latest open frontier modelsKimi, GLM, Qwen, DeepSeek, MiMo. Live the day weights drop.
- 02Free to startNo card. No forms. No sales call.
- 03Run entirely by your agentSetup, keys, models, budgets. There is no dashboard.
01 / Day zero
Weights drop. You're live before the thread trends.
| DeepSeek-V4.1-Flash | DeepSeek | Sep 10 | 39 | LIVE |
| K2 Horizon 375B | IFM | Sep 03 | 31 | LIVE |
| GLM-5.3 | Z.ai | Aug 29 | 45 | LIVE |
| Hy4 preview | Tencent | Aug 28 | -- | LIVE |
| GLM-5.3-Flash | Z.ai | Aug 26 | 42 | LIVE |
02 / Benchmarks
Know which model wins before you call it.
Every model we serve comes with independent benchmark scores, speed and price, on day one. Your agent reads the same numbers through the API and picks for you.
Compare all models→Intelligence Index: Artificial Analysis Intelligence Index v4.3.2, as of Sep 23, 2026. GPQA Diamond: lab model cards. Source ↗
- 0146MiMo-V2.6-ProXiaomi
- 0245GLM-5.3Z.ai
- 0344Kimi K3Moonshot AI
- 0442GLM-5.3-FlashZ.ai
- 0540Qwen3.8-2.4TAlibaba
- 0639DeepSeek-V4.1-FlashDeepSeek
- 0736DeepSeek V4 ProDeepSeek
03 / Router
model: "auto"
Set a quality bar once. For every request, the router picks the cheapest open model that clears it, and tells you which one it chose and why.
- Label 40,000 support tickets→GLM-5.3-FlashShort input, one-word output. The fast model clears the bar.
- Quick answer inside a chat app→Qwen3.8-27BSmall dense model, lowest latency.
- Summarize a 300-page contract→DeepSeek-V4.1-FlashLong context with a small memory footprint.
- Refactor a 2,000-line module→Kimi K3Hard, long-horizon coding. Worth the bigger model.
One setting
quality_floor: 0.85. Raise it for hard work, lower it for bulk jobs.
Pin any time
Name a model instead of auto and it never changes under you.
Every answer explains itself
x-st-routed-to and x-st-reason on every response.
04 / Headless
Works with your agent.
Claude Code / Codex / Cursor / any MCP client / the OpenAI SDK / plain curl
MCP
Docs→Your agent gets tools for accounts, keys, models and usage.
claude mcp add --transport http scaledthought \ https://api.us.scaledthought.com/mcp \ --header "Authorization: Bearer $ST_API_KEY"
CLI
Docs→One command: free account, key in .env, newest models listed.
npx scaledthought@latest init --region us
API
Docs→OpenAI-compatible. Change the base URL and keep your code.
from openai import OpenAI client = OpenAI( base_url="https://api.us.scaledthought.com/v1", api_key=os.environ["ST_API_KEY"], )
05 / Data residency
Your prompts never leave the country you pick.
US
region: "us"us-east · us-central
CA
region: "ca"ca-central · ca-west
✓Zero data retention by default
✓Never used for training
✓Region pinned per API key
06 / Built for the switch
Moving to open models should be boring.
01
Shadow mode
Mirror 1% of your closed-model traffic. Get a per-task report on quality and cost before you switch.
02
Your evals, not ours
Upload 50 real examples. Every model gets scored on your work, and the router tunes to your pass rate.
03
Hard budgets for agents
Per-key spend caps and daily limits. An agent can hold a key without holding your card.
04
Models never vanish
Pinned versions stay up 12 months after a successor ships.
05
Tool calls that just work
Every model's tool-call dialect normalized to one format, with strict JSON output.
06
Your fine-tunes stay yours
Bring LoRAs, download weights any time. Open should stay open.