MangoBoost has launched Mango Inference, a pay-per-token service offering developers several open-weight AI models through one API endpoint, on AMD Instinct GPUs. The pitch is practical: switch models with a configuration change, rather than rebuilding the application around each provider.
AMD Watch analysis
What happened
AMD announced the launch on 9 October. Mango Inference currently offers GLM-5.2, DeepSeek-V4 and MiniMax-M3 through APIs compatible with OpenAI and Anthropic clients. It supports streaming, tool calling, structured outputs and reasoning, with a console for choosing models, checking context lengths and managing usage.
Customers fund prepaid accounts and are charged by token use. Responses report token consumption, and cached input is priced separately at a lower rate, which could help with workloads that repeatedly send long prompts. MangoBoost says its infrastructure spans Korea and the United States, including a Seoul data centre it operates, and that customers can use Korean-located infrastructure for data residency needs.
Why it matters
A shared endpoint and familiar APIs could make it easier for teams to compare open models or move between them without rewriting integrations. The service also gives developers another route to AMD GPU capacity, rather than making model access synonymous with one chip or cloud provider.
The useful test will be the bill and the workload, not the launch-page promise. Per-token pricing and lower-cost cached input give buyers concrete terms to examine, while local infrastructure may matter to organisations with regional data requirements.
Our read
This is a credible new option for teams curious about open models, especially those who want to try several without first assembling their own serving stack. Start with a small workload, compare the actual cost and performance against alternatives, and check the service’s data-location terms before sending anything sensitive. Convenient APIs are lovely; they are not a substitute for procurement homework.
What to watch
- Which models MangoBoost adds to the catalogue, and how quickly.
- Public pricing details and how costs compare across models and cached input.
- Whether the service publishes workload-specific performance results and methodology.
- What data-residency and availability options it offers in each region.
Discussion spark: Would you choose a shared API that makes switching open models easy, or would you rather manage each model and provider directly for greater control?
Sources and evidence
- MangoBoost Delivers Open AI Models Powered by AMD – AMD (9 October 2026, 17:33 UTC)
Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.