# Fixstars GLM-5.3 in Japan: API Access, Token Costs and What to Verify

Fixstars now serves an NVFP4 build of GLM-5.3 in Japan. Check direct-contract access, billing and the actual data path before migrating.

By Paras Madan | 2026-11-21


Fixstars [added GLM-5.3 to its Japan-hosted inference cloud](https://news.fixstars.com/7618/) on November 20. The model itself is not new. What changed is where a team can run it through an API: Fixstars operates its own quantized build on servers in Japan.

That gives teams with Japan-only inference requirements another option. But the route is through a direct corporate contract, not an instant FastMetal model switch. Before moving a workload, check access, test whether the quantized model behaves as expected, and budget for generated tokens rather than comparing only input rates.

## Use the model ID for the direct route

The [Japanese service page](https://www.fixstars.com/ja/products-services/inference-cloud) lists `GLM-5.3-NVFP4` for direct customers. It describes an OpenAI-compatible Chat Completions API over HTTPS. The model is available after corporate application and review. Fixstars says FastMetal access to GLM-5.3 is planned, not available at launch.

The [English service page](https://www.fixstars.com/en/products-services/inference-cloud) still uses DeepSeek in its examples and says the direct endpoint URL comes with your API key. Do not copy its FastMetal URL into a GLM configuration or treat the gateway's DeepSeek example as a way to test the new model.

Set up a separate provider configuration with the issued URL and key and the exact model ID. Send a small non-sensitive prompt through Chat Completions before trying streaming, tools or structured output. Treat compatibility with any feature your application relies on as something to test, not as a consequence of accepting the API's basic request shape.

## Verify the whole route before sending private data

Fixstars says it serves the published model weights on its own Japanese infrastructure and does not forward requests to Z.ai. It also says prompts and outputs are not kept after inference or used for training.

That is a statement about Fixstars' part of the chain. A coding assistant might still send information to an external search tool, write a transcript to a telemetry service or fall back to another provider. Draw the actual path from your application to the model, and review each secondary connection separately.

In a network-restricted workflow, make the unapproved fallback test explicit: if the configured Japan route fails, the request should stop rather than get resent elsewhere. Check the relevant handling terms for your contract before using real documents.

## Model the bill from all three token counts

The launch lists rates of ¥10.49 per million uncached input tokens, ¥10.34 per million cache-read tokens and ¥210 per million output tokens, before tax. On this route, the cached-input rate is only slightly below the uncached-input rate. Do not apply a large cache discount from another provider to these figures.

For example, suppose a month uses 10 million uncached input tokens, 20 million cached input tokens and one million output tokens. The arithmetic is:

- Uncached input: 10 × ¥10.49 = ¥104.90.
- Cache reads: 20 × ¥10.34 = ¥206.80.
- Output: 1 × ¥210 = ¥210.
- Total: ¥521.70 before tax, using only the listed token rates.

An agent that writes twice as much output for the same task adds ¥210 in this example. Log usage by task, together with whether the work completed. A cheap request that causes another expensive attempt is not necessarily the cheaper workflow.

## Test warm and cold requests separately

Fixstars reports a median time to first token of about 1.13 seconds for a Tokyo test with 64 concurrent requests, average input near 139,000 tokens and roughly 96% cache hits. That describes a particular workload, not a cold-start guarantee or a measure of time to finish the answer.

Start with a cold document prompt, then test a repeat that should use a shared prefix. Run the same workload at realistic concurrency and capture first-token and full-response times. Keep an eye on tail latency so a good median does not hide the experience of users who wait longest.

If you already use GLM-5.3 elsewhere, hold the prompts and grading rules constant. The Fixstars route uses NVFP4 quantization; the name alone does not establish identical behavior. Compare schema validity, decision quality and the completion rate on the tasks that justify switching. A Japan-hosted route is useful only if it satisfies both the data boundary and the job the model needs to do.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn launches and infrastructure decisions into technical writing developers can use.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
