norddahl.com

What It Is

Norddahl offers a Gemma 4 31B Inference API, with the endpoint https://api.norddahl.dk/v1 and the model name gemma-4-31b. It follows the OpenAI /v1/chat/completions protocol, so projects that already use the OpenAI SDK can usually integrate it by simply changing the base_url and model.

Core Capabilities

The API is based on an NVIDIA NVFP4-quantized Gemma 4 31B model. It supports text chat, image input, video input, and SSE streaming output. The context window is 128K tokens, and can be extended via the protocol to 262144 tokens. Supported image formats include JPG, PNG, WebP, and the first frame of GIFs; supported video formats include MP4 and WebM, up to 60 seconds, sampled at 1fps. For generation controls, it covers common parameters such as temperature, top_p, stop, penalties, and max_tokens.

API, Pricing, and Limits

The documentation provides cURL, Python, and JavaScript examples, as well as a /v1/models endpoint. Authentication uses a Bearer API Key. Keys are issued per customer, and each key has its own rpm, tpm, and monthly USD budget limit via max_budget. However, the main documentation does not disclose public token pricing, plans, payment methods, or any free quota, so commercial predictability is only moderate. When rate limits are triggered, the API returns 429 with a Retry-After header; repeated 401 responses may also trigger a temporary fail2ban block.

Pros and Cons

The main advantages are strong OpenAI compatibility, low integration cost, support for multimodal input and streaming responses, and relatively clear guidance on error handling and rate limits. It is suitable for quickly replacing or supplementing an existing OpenAI call chain. The drawbacks are its single-model offering and lack of transparency around pricing and trial policies. Video support also has duration and sampling limits. On privacy, the documentation only provides API Key security recommendations, without explaining data retention, training usage, or compliance certifications.

Who It’s For and Access from China

It is best suited for developers, internal enterprise tooling teams, and multimodal prototyping teams—especially scenarios where users want to call Gemma 4 31B through the OpenAI SDK. For Chinese-language support, the documentation only states that the model is multilingual and has been tested in Danish, English, German, French, and Spanish; Chinese is not mentioned. There is also no clear information on access from mainland China or payment availability, so it is advisable to run a network connectivity test first. Alternatives to compare include OpenAI, Gemini, Claude, and China-based options such as Alibaba Cloud Bailian, Zhipu AI, and SiliconFlow.