Ultrafast AI optimized
for quality and speed

Tuna Ultrafast is a premium product hosted on expensive hardware, intended for use cases where paying the premium for speed is worth it. While we are researching how to serve a slower Fast tier, our current focus is on Ultrafast.

Most inference providers try to balance all 3: quality, speeds, and prices. Tuna Ultrafast trades off pricing in exchange for doing 2 things very well: preserving model quality and increasing per-user speed. For example, serving more requests together can lower the cost per token, but it can also slow down individual responses. We allocate more compute to each request.

Since we prioritize quality, we are unable to use some cost-cutting techniques that make the model weaker. For example, a popular technique is to quantize the model to a half the size (which is faster, cheaper, and dumber), but we are not willing to sacrifice any quality. Some other providers serve a watered down model without clearly disclosing the loss of quality.

We spend a lot of effort on engineering optimizations throughout the entire stack. We also use Bluefin, our AI optimization system, to focus optimization efforts on increasing per-user speed. Then, we scale up our deployment so that each request gets a bigger slice of compute than most inference providers would allocate. We prioritize speed of individual responses, rather than just trying to cheaply serve the masses with slow speeds.

New users can get a $10 promotional credit after entering valid billing details and valid payment method. No purchase is required. You will not be charged unless you choose to purchase credits or enable automatic top-ups yourself. We have systems to limit one signup bonus per user. The promotional credit amount is subject to change any time.

We support the industry-standard OpenAI Chat Completions API. For compatible tools, you can add Tuna as a provider by using the base URL https://api-us.tuna.ai/v1, your API key from the dashboard, and a Tuna model ID. Our compatibility documentation lists the supported endpoints and features.

The docs provide integration examples, and is easily readable by any AI with web access. You can mention Tuna.ai to your existing coding agent to help integrate it, and can handle the next steps.

Websocket support is on the roadmap, which may further reduce latency for US users by reducing round trip times (~0.1 seconds faster).

Tuna works with most coding harnesses. We are developing single-line install scripts to support your harness. Meanwhile, you can mention Tuna.ai to your coding agent to search for setup instructions for your specific harness.

US regional routing means that model inputs and outputs are processed in US data centers. Temporary model caches also remain in US data centers. Many other providers charge 10% more for this feature because it can cost more to use only US-based infrastructure.

Tuna is committed to running US-based inference infrastructure, so US regional routing is the default and only option. US routing is included in model prices at no additional charge. It is available to all users without an enterprise plan or separate activation.

The guarantee covers inference processing and model caches. Account and billing records are handled separately due to the third-party software platforms we use (such as Stripe). For more information, see the documentation on US regional routing.

Zero Data Retention, also abbreviated as ZDR, means that Tuna does not log or store prompt or generation data. Tuna does not use your prompts or model responses to train models. Data is only temporarily used for processing your requests, including temporary model caches.

We retain usage metadata, such as token counts, for billing. These records do not include your prompts or model responses. ZDR is enabled by default for all users, without an enterprise plan or separate request.

For more information, including how temporary caches are handled, see the documentation page on privacy and ZDR.

Token generation speed depends on many factors, such as the load on our hardware or the complexity of your request. If your output speed is less than the SLA speed target, an automatic discount of up to 50% applies, based on a formula. If the speed is at or above the SLA speed target, the request will be charged at the regular rate.

The SLA applies automatically to all users of Ultrafast models. See the Models & Pricing page for the current speed targets. This ensures that Tuna is financially incentivized to provide ultrafast speeds (unlike OpenAI, which charges full price even for delivering degraded performance at slow speeds of 100 tok/s).

We can host fine-tuned models on custom deployments at ultrafast speeds at a somewhat reasonable cost, due to our optimizations and infrastructure. Prices depend on model size, speed targets, and tenancy isolation requirements. Ultrafast dedicated instances for large models require substantial amounts of high-performance hardware and currently start at $5M/year. Please contact us to see if there is a fit.