Fixstars Inference Cloud | Fast, Japan-Hosted Inference for Open-Weight Models

Overview

An OpenAI-compatible API for the tools you already use.

Fixstars runs and optimizes open-weight LLMs on servers it manages in Japan and offers them as an API. Use your API key to access the models from the tools you already use.

Model: DeepSeek-V4.1-Flash-1017 (as of November 2026)
API: OpenAI-compatible (Chat Completions format), API key authentication (HTTPS)
Infrastructure: GPU server performance optimized by Fixstars
Operated by: Fixstars Corporation (TSE Prime: 3687)

Any IDE, CLI, or agent that supports OpenAI-compatible APIs works. Connect it by setting the endpoint URL, API key, and model ID.

example.py Python / OpenAI SDK

from openai import OpenAI

client = OpenAI(
    base_url="https://api.fastmetal.ai/v1",
    api_key="<YOUR_API_KEY>",
)
res = client.chat.completions.create(
    model="japan-deepseek-v4.1-flash-1017",
    messages=[{"role": "user", "content": "..."}],
)

from openai import OpenAI

client = OpenAI(
    base_url="<ENDPOINT_URL>",
    api_key="<YOUR_API_KEY>",
)
res = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V4.1-Flash-1017",
    messages=[{"role": "user", "content": "..."}],
)

For direct contracts, we provide the endpoint URL when we issue your API key.

Performance

Median time to first token: 0.41s.
Optimized in-house.

Fixstars, whose core business is performance engineering, optimizes its inference infrastructure in-house.

TTFT of Fixstars Inference Cloud

0.41s

Half of all requests start producing output within this time.

Model: DeepSeek-V4.1-Flash-1017
Workload: Agentic (avg. ~100K input tokens, ~96% cache hit rate)
Concurrency: 240 requests
Measured from: Tokyo
Measured in: October 2026

Results may differ when using FastMetal.

Inference infrastructure optimized for speed

Precomputed KV cache shared across all servers
Prefill and decode separated (PD disaggregation)
Scheduling based on input length and load

Our white paper explains how we reduced P90 time to first token for GLM-5.1 to roughly one-fifteenth of its original value.

White paper: Bringing the Large-Scale MoE Model GLM-5.1 to Practical Speeds on a Single NVIDIA H200 Node

Data Residency

Servers and storage, all in Japan.

Source code and specifications you send are processed only on servers in Japan.

Easy to document for audits and compliance

You can state in audits and internal reviews where your data is processed: on Fixstars-managed servers in Japan.

Less exposure to laws and policy changes outside Japan

Less affected by data disclosure requests from outside Japan or by service policy changes at overseas providers.

Operated by a TSE Prime-listed company

Fixstars Corporation (TSE Prime: 3687) operates the service on GPU servers it manages in Japan.

No inputs or outputs retained (Zero Data Retention)

On Fixstars servers, prompts and outputs are not saved to persistent storage or recorded in persistent logs, and are never used to train AI models.

When using FastMetal

Prompts and inference stay in Japan via FastMetal as well. On FastMetal, specify the model ID.

Pricing

Priced in line with major US providers.
No premium for hosting in Japan.

We charge the standard token price for DeepSeek-V4.1-Flash-1017, with no markup for running in Japan. You pay only for what you use — no fixed fees and no minimum charge.

Price per 1M tokens (excl. tax)
Model	INPUT	OUTPUT	CACHE READ
DeepSeek-V4.1-Flash-1017	$0.30	$1.20	$0.006
Models and prices are as of November 2026.

* The same prices apply via FastMetal and under a direct contract.
* Based on DeepSeek's official peak-hour pricing as of October 2026. Prices may be revised due to exchange rates and market conditions.

About 8–17x lower monthly cost than Claude.

For the same usage, the difference in unit price becomes the difference on your monthly bill. Here is a comparison with Anthropic's standard API pricing.

Price per 1M tokens (excl. tax) and estimated monthly cost
Item	DeepSeek-V4.1-Flash-1017Fixstars Inference Cloud	Claude Opus 5.5Anthropic	Claude Sonnet 5.5Anthropic
INPUT	$0.30	$4.00	$2.00
OUTPUT	$1.20	$20.00	$10.00
CACHE READ	$0.006	$0.20	$0.10
Estimated monthly cost for 100M tokens	~$5	~$83	~$41

* Claude Opus 5.5 and Claude Sonnet 5.5 use Anthropic's official standard API pricing (as of October 29, 2026). Cache write charges and batch discounts are not included. Estimates assume a mix of 90.6% cached input, 7.7% new input, and 1.7% output. Each model differs in capabilities and strengths.

How to use

Try it today on FastMetal,
or contract directly for invoicing and engineering support.

Item	Via FastMetal	Direct contract
Best for	Trying it out or getting started right away	Teams that need invoicing or Fixstars technical support, or want to connect directly to Fixstars servers
Where inference runs	Japan (Fixstars servers)	Japan (Fixstars servers)
Network route	Through the FastMetal gateway (in Japan)	Direct connection to Fixstars servers
Contract with	FastMetal	Fixstars Corporation
Payment	Prepaid (card / bank transfer)	Billed monthly in arrears (invoice)
Getting started	Same day as sign-up	API key issued after application and review
Technical support	Contact FastMetal (Japanese-language support available)	Provided by Fixstars engineers
	Get started on FastMetal	Apply for a direct contract

* Direct contracts are for businesses only. We do not accept applications from individuals.

FAQ

Frequently asked questions

DeepSeek is a Chinese model. Are requests sent to DeepSeek or to servers in China?: Fixstars runs the publicly released model weights on GPU servers in Japan. No requests are sent to DeepSeek or to servers in China.
Is my data stored? Is it used for AI training?: On Fixstars servers, prompts and outputs are not saved to persistent storage or recorded in persistent logs, and are never used to train AI models.
Which tools can I use it with?: Any IDE, CLI, agent, or internal tool that supports OpenAI-compatible APIs. Just set the endpoint URL and API key.
Should I choose FastMetal or a direct contract?: To try it out, we recommend FastMetal, which you can use the same day you sign up. If you need invoicing (billed monthly in arrears), technical support from Fixstars engineers, or a direct connection to Fixstars servers, apply for a direct contract.
Is there anything to watch out for when using FastMetal?: Specify japan-deepseek-v4.1-flash-1017 as the model ID. If you specify deepseek-v4.1-flash-1017 without japan-, or auto, your requests may be routed to models hosted outside Japan.
Can individuals use the service?: Direct contracts are for businesses only. For eligibility via FastMetal, please check FastMetal's terms.

Get your API key and start today.

With FastMetal, you can start the same day you sign up. For businesses that need invoicing or Fixstars technical support, apply for a direct contract. Tell us your expected number of users and token volume, and our engineers can help you plan your deployment.

Get started on FastMetal Apply for a direct contract

To discuss your number of users or expected token volume, use our contact form.

Fast, Japan-hosted inference for open-weight models.