Evanesca vs Fireworks AI: fine-tuning and speed tiers, or models run in Europe
Fireworks AI and Evanesca agree on one point: neither keeps your prompts — on open models, unless you opt in to logging, Fireworks AI holds them only in volatile memory while the request runs. They differ on the rest. Fireworks AI, a US company, adds fine-tuning, speed tiers and dedicated deployments, and serves requests from any region unless an Enterprise account restricts it. On GPT-OSS 120B, before VAT on both sides, Fireworks AI costs 37% less on both input and output. Evanesca runs its 17 models in Europe, most of them in France with French operators, on every request.
Facts about Fireworks AI read on its own pages on 6 Oct 2026
At a glance
Each Fireworks AI cell links to the Fireworks AI page it comes from, read on 6 October 2026; each Evanesca cell links to the page that describes it.
| Question | Fireworks AI | Evanesca |
|---|---|---|
| Who runs it | Fireworks.ai, Inc., San Mateo, California; terms under Delaware law source | FLEECE AI SASU, a French company legal notice |
| What it sells | Serverless inference in Standard, Priority and Fast modes, dedicated deployments, fine-tuning and training source | A chat and an OpenAI-compatible API, inference only docs |
| Prompts kept | None on open models unless you opt in; the Responses API stores conversations for 30 days unless you set store=False source | None: held in memory while the answer is written; billing data is kept privacy |
| Training on your data | Not without your explicit opt-in source | Never by Evanesca; Mistral AI lets its customers object to training, and Evanesca's account objects, except for what is sent to Mistral Large 4: it is a preview, and Mistral's commercial terms let it use the data of preview models for training, whatever the objection source source |
| Where requests run | Any region by default; a region lock is an Enterprise setting, with US-only serverless at a premium and EU-only on request source | Always in Europe — OVHcloud or Scaleway in France, Mistral AI in its EU region, AKI.IO in Germany, Infomaniak in Switzerland sovereignty |
| What the price includes | US dollars, “exclusive of any taxes”, value-added tax among them source | Euros, margin and 20% VAT included; no subscription pricing |
Prices, model by model
GPT-OSS 120B is on both price lists, run on Evanesca by Scaleway in Paris, with OVHcloud in Gravelines as a fallback. Fireworks AI prices it in two modes, Standard and Priority.
| Model | Fireworks AI, before tax | Evanesca, VAT included | Evanesca, before VAT | Cheaper, before tax |
|---|---|---|---|---|
| GPT-OSS 120B | $0.15 / $0.60 (Standard) source | €0.252 / €1.008 | €0.21 / €0.84 | Input: Fireworks AI, −37% · Output: Fireworks AI, −37% |
| GPT-OSS 120B | $0.18 / $0.72 (Priority) source | €0.252 / €1.008 | €0.21 / €0.84 | Input: Fireworks AI, −24% · Output: Fireworks AI, −24% |
Fireworks AI publishes its prices in US dollars, before tax; Evanesca's are in euros, margin and 20% VAT included. So the verdict compares both before tax: Evanesca's price divided by 1.2, and the dollar converted at the ECB reference rate of 6 October 2026, 1 EUR = 1.1269 USD (ECB). A gap under 2% names no winner. On GPT-OSS 120B in Standard mode, Fireworks AI costs 37% less on both input and output; against Priority, Fireworks AI costs 24% less on both input and output.
DeepSeek V4.1 Flash and GLM-5.3 are on both lists too: on Evanesca, AKI.IO runs them in Germany (DeepSeek V4.1 Flash, GLM-5.3). Fireworks AI also serves Kimi K3, which Evanesca does not: no European operator runs it for Evanesca.
What happens to your prompts
« Fireworks has Zero Data Retention by default. »
In detail: for open models, prompt and generation data “exist only in volatile memory for the duration of the request”; with prompt caching, some may stay in memory for several minutes; nothing is written to persistent storage unless you opt in, and Fireworks AI logs metadata such as token counts source. One exception is documented: the Responses API stores conversations by default, for 30 days, unless the request sets store=False.
Evanesca works the same way on its side: prompts and answers are held in memory while the answer is written, never stored, and your history stays in your browser; billing data — tokens and cost — is kept. What differs is the company running the model. At Evanesca it is named and European: OVHcloud keeps “only the data required for billing purposes” source; Scaleway may keep and read requests that cause errors or look abusive, for up to 14 days source; Mistral AI keeps its models' inputs and outputs 30 days to monitor abuse source.
Region and jurisdiction
Fireworks AI's privacy notice says it has “servers for the Service in the US” source. Its data residency setting, for Enterprise accounts, restricts inference to one region; “None is the default”, which means any region source. The listed region is the US — US-only serverless models cost 1.5 times the base price from 1 September 2026 — and “for EU-only Serverless”, Fireworks AI asks you to contact sales source.
Placement aside, the company is in the US: Fireworks.ai, Inc., with terms governed by Delaware law source. The US CLOUD Act lets US authorities require a provider under US jurisdiction to disclose data in its “possession, custody, or control”, wherever it is stored source. Zero retention reduces what there is to disclose; it does not change who can be asked.
Evanesca's models need no setting: they run in Europe — in France with OVHcloud or Scaleway, in Mistral AI's EU region, in Germany with AKI.IO, or in Switzerland with Infomaniak. OVHcloud, Scaleway, Mistral AI and Infomaniak are European companies with no US parent, which puts them out of the CLOUD Act's direct reach; AKI.IO is a German company whose shareholders are not published. The sovereignty page details each operator.
Where Fireworks AI is the better choice
- Fine-tuning and training: supervised fine-tuning, preference optimisation and reinforcement learning, with deployment of the result, available today; on Evanesca, fine-tuning is planned source.
- Speed tiers: Standard, Priority and Fast modes on serverless, so you can pay for latency where it matters source.
- Dedicated deployments and batch inference at half the serverless price, for large, steady workloads source.
- Zero retention on open models by default, documented down to memory and caching source.
- A broader catalogue, with any model it hosts priced by size when not listed by name, and a service that is running today; Evanesca has not opened yet source.
If you need to train or fine-tune today, or to buy guaranteed throughput, Fireworks AI does it now.
Where Evanesca fits better
- Every model in Europe on every request, with no Enterprise contract or region setting needed sovereignty.
- A French company, and a European operator for every model legal notice.
- Euro prices with VAT included, no subscription, credit from €2 that never expires pricing.
- A chat and an Agent mode as well as the API, with a passport under each answer naming the model, operator, country and law Agent mode · passport.
- For teams: a signed Art. 28 DPA, SSO and roles, spending limits business.
Embedding models run in France too, behind the same OpenAI-compatible API embeddings.
How to choose
- Training, fine-tuning, priority latency or dedicated capacity: Fireworks AI.
- Personal, client or regulated data processed in Europe by a European operator, with nothing to configure: Evanesca's models.
- A mix: fine-tune on Fireworks AI, and send sensitive inference to Evanesca's models through the same OpenAI client.
Whichever you pick, check the endpoint as well as the provider: on Fireworks AI, the Responses API keeps conversations for 30 days unless store=False source.
How this page was checked
Fireworks AI's pages were read on 6 October 2026. Evanesca's prices come from its catalogue and every verdict above is computed from them, with the dollar at the ECB reference rate of that day. If a line no longer holds, write to us.
Frequently asked questions
Does Fireworks AI keep my prompts?
Not on open models by default: prompts and outputs stay in volatile memory for the duration of the request, and only metadata such as token counts is logged. The Responses API is the exception: it stores conversations for 30 days unless you set store=False.
Can Fireworks AI keep my requests in the EU?
Not by default: its data residency setting is for Enterprise accounts, the listed region is the US, and EU-only serverless is on request through sales. Evanesca's models always run in Europe, with no setting to change.
Is Fireworks AI cheaper than Evanesca?
On GPT-OSS 120B, before VAT on both sides and in Standard mode, Fireworks AI costs 37% less on both input and output. Prices read on 6 October 2026.
Can I fine-tune a model on Evanesca?
Not yet: fine-tuning is planned on Evanesca, which serves inference for now. Fireworks AI offers fine-tuning and training today; you can use both, and keep sensitive inference on Evanesca's models.