AMD Radeon Token Factory (daily free allowance, China developer portal)
AMD
The free shared API doesn't consume credits: the cap is a per-account, rolling-period USD spend limit (the official figure isn't published; third-party sources cite roughly $10/day, about 100 million tokens/day for DeepSeek-V4-Flash), with an effective rate of about 20 RPM and 8 concurrent requests. As of 2026-09-01 the free pool has 4 models; the flagship is the GA release of DeepSeek-V4-Flash with a native 1M context.
The official docs site (amd-aim.github.io/radeon-cloud-docs, updated 2026-08-31) discloses the full mechanism: the spend cap is a rolling period (starting from your first billed request, not a fixed midnight reset), and the remaining balance is visible in real time via the X-RateLimit-*-User-Daily-USD response headers or GET /api/profile/model-usage. Rate limits: 20 RPM per account at the gateway (60-second sliding window), plus platform-level limits of 30 RPM per key, 120 RPM per IP, and 8 concurrent requests. The API is OpenAI-compatible (/chat/completions) plus Anthropic-compatible (/v1/messages — the official docs explicitly say this is for clients such as Claude Code); Base URL https://developer.amd.com.cn/radeon/api/v1, and a single rc--prefixed key works across every shared model. No embeddings endpoint. The free pool changes over time (the docs say to check GET /v1/models rather than trusting a hardcoded list): as of 2026-09-01 it's DeepSeek-V4-Flash-0731 (FREE, 1M context, supports tools and reasoning-effort settings), MiniCPM5-1B (FREE, 128K), and GLM-5.3-Flash plus Qwen3.8-Flash-Next (LIMITED FREE, 256K each). The pool has narrowed twice since July (mid-August briefly had 13 models fully free, and DeepSeek-V4-Pro was added and later removed); every model is tagged experimental.
A different product from the site's amd-developer-cloud-credits entry (a one-time $100 Developer Cloud compute credit) — the two don't overlap and can both be claimed. Token Factory is labeled BETA and runs on AnRuiCloud's backend. Sign-in supports a phone number or email verification code, a password, or ModelScope/CSDN/GitHub third-party accounts — email alone is enough to register, and the free shared API works normally even while an account is under manual review. One blogger's hands-on test recorded a 22-second first-token latency, worth keeping in mind under heavy load.