Prima Era API
GLM-5.3 and Kimi-K3 on NVIDIA B300 GPUs, through OpenAI- and Anthropic-compatible endpoints.
Endpoint
Base URL https://api.primaera.xyz/v1
/v1/chat/completions: OpenAI format, with streaming, tool calls, logprobs and reasoning returned asreasoning_content./v1/messages: Anthropic format, so Claude Code-style clients work unchanged./v1/models: the model list, public.
Usage in every response reports prompt, cached and completion tokens.
Models
glm-5.3: zai-org/GLM-5.3 with FP8 weights and FP8 KV cache.
kimi-k3: moonshotai/Kimi-K3 with its official FP4 weights.
Both are served by sglang on NVIDIA B300 GPUs in the United States that we operate. Context per request,
prompt plus output, is 1,048,576 tokens for kimi-k3 and 131,072 for glm-5.3, as
/v1/models reports. We already provide inference on the
Liquid Inference exchange.
Prices
| GLM-5.3, per million tokens | Input | Cached input | Output |
|---|---|---|---|
| List | $0.55 | $0.18 | $1.05 |
| Wholesale, best effort (500+ requests an hour) | $0.35 | $0.07 | $0.90 |
| Wholesale, always on | $0.35 | $0.08 | $1.50 |
| Kimi-K3, per million tokens | Input | Cached input | Output |
|---|---|---|---|
| List | $0.30 | $0.10 | $5.50 |
| Wholesale (500+ requests an hour) | $0.25 | $0.08 | $4.50 |
Availability
Our GPUs serve whenever they have free capacity. For GLM-5.3, list and always-on keys then continue
automatically on a third-party GLM-5.3 provider at the same price, and best-effort wholesale keys instead receive
429 or 503 with Retry-After, for routers that fail over on their own. Kimi-K3 is served only on our
GPUs, with the same 429 or 503 when they are full. There is no uptime commitment unless agreed in writing.
Keys, trials and billing
Ask for a key or a trial key at accounts@primaera.xyz. We invoice with EU VAT in EUR or USD, prepaid or monthly, by bank transfer.
We do not log prompts or completions; see the privacy notice and the terms of service.