Every EU-sovereign model we serve, with price and verification on one board 🇪🇺
€ per 100k tokens, in / out, VAT incl. Live on Oct 12.
GPT-OSS 20B · OVHcloud · €0.0068 / €0.0252 · 6/6
Qwen3.5 9B · OVHcloud · €0.0168 / €0.0252 · 6/6
GPT-OSS 120B · OVHcloud · €0.0135 / €0.0672 · 5/6
Mistral Small 3.2 24B · OVHcloud · €0.0152 / €0.0471 · 5/6
DeepSeek V4 Flash 0731 · Scaleway · €0.0672 / €0.1344 · 6/6
Qwen3.8 27B · OVHcloud · €0.0672 / €0.4536 · 6/6
Qwen3.5 397B · Scaleway · €0.1008 / €0.6048 · 6/6
Mistral Large 4 · Mistral AI · €0.2285 / €0.7023 · 4/6
GLM-5.2 · Scaleway · €0.3024 / €0.9240 · 6/6
6/6 = all six of our checks passed: license, provenance, security, GDPR assessment, benchmarks, API.
Which one would you put in production first?
getfrontiereai.eu/pricing
🤖 Made with AI
13
397
Evanesca is now live on the Google Play Store 📱
The private AI assistant entirely built on Frontière AI's infrastructure: your conversations and data stay in Europe.
Sovereign AI, now in your pocket.
👉 play.google.com/store/apps/d…
#SovereignAI #GDPR #Europe #Evanesca #privacy
Evanesca is now on Android 📱
A Private European AI Assistant where your data never leaves Europe, and your chats never leave your phone.
Same European guarantees:
- All 19 models run in Europe, by European operators
- No training on your chats
- Nothing kept after the answer
- Every answer comes with a "passport": where it was processed, and what we kept (nothing)
All the tools, built with European partners:
- Web search via Linkup 🇫🇷
- Images by Infomaniak 🇨🇭
- Dictation by Mistral AI 🇫🇷
- Photos & scans: GPS and device data stripped
- PDF, Word, Excel: read and written
- Memory: optional, kept on your phone
Your chats stay on your device:
- Stored on the phone, no copy on our servers
- Hidden in the app switcher by default
- Fingerprint or face lock if you want it
Android 7+, available now on Google Play.
2
45
Replying to @bugrasa
That’s the kind of gap a leaderboard never shows. Same fields extracted, but “480.00” as a string quietly breaks a pipeline downstream. Worth re-running with a strict JSON schema (amount: number) to see if the outlier falls in line. Which model was it?
34
Replying to @donella45141
Exactly. Per-token price stops meaning much once reasoning kicks in. Cost per completed task is the real metric, which is why the Compare Mode puts latency, tokens and € side by side.
1
44
Kimi K3.1 Pro or GLM-5.3?
The only fair way to decide is to test them on your own prompts before you integrate anything.
New in the Frontière AI: Compare mode.
You send one prompt to two models at the same time and read their answers side by side as they arrive.
Under each answer you get the tokens used, the cost, the time to first token, the total time and the throughput.
In the video, we ask both models for a Python function that removes duplicates from a list while keeping the order.
→ GLM-5.3 sends its first token sooner (0.48 s) and generates more tokens per second (108 tok/s).
→ Kimi K3.1 Pro still finishes in 5.2 s instead of 11.1 s, and costs €0.0070 instead of €0.0111.
The reason is that GLM-5.3 reasons first (730 tokens) and generates 1,162 tokens in total, against 442 for Kimi. The fastest model per token isn’t always the cheapest per answer.
To keep the comparison fair, we set 4 rules:
- Same settings on both sides: system prompt, temperature, max tokens and top-p.
- Web search and tools are off, so you compare the models and not their tools.
- No winner by network luck: when the gap is under 5%, no star is awarded.
- Measured the way your app sees it: in your browser, with network and gateway time included.
Billed like the API: each answer is charged at its model’s per-token price. If you stop an answer midway, the part that wasn’t generated isn’t billed.
Then you pick the best model and wire it up in one line. Your app uses the same gateway, the same route and the same per-token price, so it gets exactly what you measured.
Your comparisons stay in your browser and never reach our servers.
Quality, speed, price: compare before you wire up the API.
#SovereignAI #LLM #OpenWeight #API #Developers #FrontiereAI
Frontière AI retweeted
New on Evanesca: Gemma 4 26B by Google 🇪🇺
Google’s open model, run in Germany, not on Google’s servers.
Apache-2.0, 256k context, text + images.
6/6 checks passed. Your photos and chats: not kept, not used for training.
Try it → evanesca.app
Frontière AI retweeted
Which AI answers better? Stop guessing.
New on Evanesca: Compare.
One prompt, two open models, side by side.
Kimi K3.1 Pro vs Qwen3.8 27B: both got it right, Qwen was 6x faster to first token. Test cost: €0.0015.
Hosted in Germany. Nothing kept.
Which two would you race next?
Experience Kimi K3.1 Pro on a European Sovereign & Private API here : getfrontiereai.eu
5
Replying to @Kimi_Moonshot
@Kimi_Moonshot K3.1 Pro - Attacked before it shipped.
Before opening the model to our European customers on Frontière AI, we sent 15 adversarial prompts through our gateway, on the sovereign route, with no fallback.
Prompt injection, system prompt leak, jailbreak, data exfiltration, appropriate refusal.
→ 15/15 safe answers. 5/5 categories passed.
What you get:
2.8T-parameter MoE (104B active), 1M-token context, text + image, tool calling
Served on Verda GPUs in Germany, no sub-processor, nothing sent to Moonshot AI
No prompt or completion stored, never used for training
x-sovereign: true on every response
Per token, in euros (VAT incl.): €3.024/M input · €15.12/M output · €0.3024/M cached
Prepaid, no subscription, from €10.
OpenAI-compatible. One line to switch:
model="kimi-k3.1-pro"
Every prompt goes somewhere. And where it lands matters. 🇪🇺
When you ask an AI a question, your data travels.
To which server? In which country? Under which law?
Most tools won’t tell you.
At Frontière AI, we show you where every model runs:
📍 Mistral AI, France
📍 AKI.IO, Germany
📍 Infomaniak, Switzerland
All the major open models are here: DeepSeek, Qwen, Mistral, Gemma, GLM, Llama.
➡️ 18 models hosted in Europe
➡️ One API, OpenAI-compatible
➡️ We don’t store your prompts or the responses
Built for developers, doctors, startups, researchers, teachers, journalists, accountants… and everyone who has to comply with GDPR.
Frontière AI, the European gateway to open AI models.
Prepaid credits from €10, no subscription.
🔗 link in first comment
#SovereignAI #ArtificialIntelligence #GDPR #EuropeanAI #OpenSource #OpenWeights #Mistral #DeepSeek #Llama #DataPrivacy #DataSovereignty #FrontiereAI #GenAI #EUTech #AIcompliance
1
9
124
Frontière AI also works for Agentic Work !
When your agents work on sensitive data, you have to make sure that it stays private...
That's why Fleece AI's new EU-sovereign models run through Frontière AI:
Operated in Europe, with zero prompt retention. 🇪🇺
New on Fleece AI: EU-sovereign AI models 🇪🇺
Your AI agents handle your contracts, invoices and job applications.
So where does that data go?
Four open-weight models are now hosted and run in Europe, right inside Fleece AI:
→ DeepSeek V4.1 Flash: hosted in Germany, 1M-token context, text and images
→ Qwen3.8 27B: hosted in France, multimodal
→ Mistral Small 4: Mistral’s EU region, fast and affordable
→ GLM 5.3: hosted in Germany, reasoning, Zhipu’s full model (Pro plan)
Served by @FrontiereAI, through AKI IO in Germany, @Scaleway_fr in France and @MistralAI's EU region.
These guarantees are enforced at the gateway, not just written in a policy:
✓ EU providers only: no US company in the path, no CLOUD Act exposure
✓ Not retained: prompts and files aren’t kept after the run
✓ Never used for training, by any model provider
✓ EU fallback only: if a model goes down mid-task, another EU model finishes it
You choose the model agent by agent.
Every task shows where it ran, who served it and what was kept (nothing).
🇪🇺 Europe, or nothing.
Start your free trial today ! 🐏
#AI #DigitalSovereignty #AIAgents #GDPR #EuropeanAI
7
113
Frontière AI retweeted
New on Evanesca - Your Private European AI Chat
Kolibri 1 by @Aleph__Alpha 🇩🇪
Made in Germany. Runs in Germany. Apache-2.0, open weights.
78B, 256k context, tools + reasoning.
6/6 checks passed before going live. What we keep from your chats: NOTHING.
Try it today → evanesca.app
Replying to @euacchq
Sovereign AI runs on open-weight models. A 70B model is ~140 GB: at €0.055/GB, that’s ~€7.70 every time a French user downloads it. Tax value, not volume, or Europe ends up taxing the very commons it builds on. Still a first-reading amendment (11–10), there’s time to fix it.
3
28
Private AI chat built on Frontière AI's API. You can see exactly where each request is processed, and your prompts are never stored.
Does your AI tell you where its answer came from?
Evanesca does. Every answer comes with a passport: which model, where it ran, what was kept (nothing), whether it trains on your data (no), and what it cost.
18 models. All in Europe.
evanesca.app 🇪🇺
7
45
European Sovereign AI Router 🇪🇺
One API. One MCP URL. 13 sovereign models.
- Keep your OpenAI code, swap one URL
- Every response says where it ran
- Plug it into Claude Code, Cursor, Cline
- Pay per token, from €10
Your prompts never cross the border.
getfrontiereai.eu
9
86
First EU Sovereign Agentic chat 👇🇪🇺
🇪🇺 Built in Europe. For Europe.
"We'd love to use AI… but IT blocked it."
So we built Evanesca: a private AI chat and agent where every answer is processed in Europe, and nothing is kept once it's written.
In 1 minute:
→ 18 open models (Mistral, DeepSeek, Qwen, Apertus…), all run in Europe
→ Board-ready charts, live calculators, GDPR checklists
→ A passport on every answer: model, host, country, training, cost (€0.0011)
No copy. No training on your data. Pay as you go.
We answer. Then we forget.
Send this to your CIO.
8
46







