Catalog
Open-source AI models, and the jurisdiction each one runs under
Frontière AI serves 26 models live out of 32 in the catalog, behind a single OpenAI-compatible API. 19 of them run on EU-sovereign infrastructure — providers with no non-EU capital control, designed to stay outside US jurisdiction. The others are labeled “fast access” because today they are only reachable through US-controlled infrastructure, and that label appears on the card and in the API response before you call. Every price below is what you actually pay, per 100,000 tokens, margin and VAT included — the entry price is €0.0168 on Gemma 4 26B (A4B MoE). No subscription: you top up a prepaid balance from €10, and calls are refused once it reaches zero (a call already running can take it slightly below; the difference comes off your next top-up).
26
models live, 32 in catalog
19
run on EU sovereign infrastructure
€0.0168
entry price /100k input tokens
Best for…
Which model for which job
Eight categories, each derived from the registry with a stated rule — not editorial picks. The rule is printed under every card so you can recompute the answer yourself.
Pixtral 12B (vision)
Fastest response — smallest live model
Pixtral 12B (vision)
Cheapest — lowest combined input + output price
Kimi K3 (2.8T parameters, 1M context)
Best reasoning — largest live model flagged “Reasoning”
Qwen3 Coder 30B
Best coding — code-specialised model
Pixtral 12B (vision)
Best vision — dedicated vision model
Pixtral 12B (vision)
Best agents — measured tool calling, sovereign first
Kimi K3 (2.8T parameters, 1M context)
Best long context — largest verified context window
Kimi K3.1 Pro (2.8T parameters, 1M context)
Best sovereign — most capable EU-sovereign model
Which model should I use?
Pick what matters; Frontière AI narrows the catalog to three. Two answers are enough — the rules are the ones printed above.
Budget
Sovereignty requirement
Context needs
Main use
EU sovereign
19No non-EU capital or jurisdictional control over the company operating the infrastructure — the SecNumCloud principle. No US parent, no US-controlled subsidiary: a US court order has no direct line to these providers.
| Model | Publisher | Size | Licence | Hosted via | Price /100k in | Price /100k out |
|---|---|---|---|---|---|---|
| Qwen3 235B (instruct)qwen3-235bcontext 250kOpen weights · GPAI5/6Model card ↗ | Qwen | 235B (22B active) | Apache-2.0 | 🇪🇺 scaleway | €0.126 | €0.378 |
| Qwen3.5 397B (multimodal)qwen3.5-397bmultimodalcontext 250kOpen weights · GPAI✓ VerifiedModel card ↗ | Qwen | 397B (17B active) | Apache-2.0 | 🇪🇺 scaleway | €0.1008 | €0.6048 |
| GLM-5.2 (Zhipu/Z.ai)glm-5.2context 256kOpen weights · GPAI✓ VerifiedModel card ↗Training docs ↗ | Zhipu / Z.ai | 753B | MIT | 🇪🇺 scaleway | €0.3024 | €0.924 |
| DeepSeek V4 Flash 0731deepseek-v4-flash-0731context 256kOpen weights · GPAI✓ VerifiedModel card ↗Training docs ↗ | DeepSeek | 284B (13B active) | MIT | 🇪🇺 scaleway | €0.0672 | €0.1344 |
| Llama 3.3 70Bllama-3.3-70bOpen weights · GPAI5/6Model card ↗ | Meta | 70B | Llama 3.3 Community LicenseNot a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇪🇺 scaleway+1 EU fallback | €0.1512 | €0.1512 |
| Qwen3.6 35B A3B (multimodal)qwen3.6-35bmultimodalcontext 256kOpen weights · GPAI✓ VerifiedModel card ↗ | Qwen | 36B (3B active) | Apache-2.0 | 🇪🇺 scaleway | €0.042 | €0.252 |
| Qwen3 Coder 30Bqwen3-coder-30bcodecontext 256kOpen weights · GPAI✓ VerifiedModel card ↗Training docs ↗ | Qwen | 30.5B (3.3B active) | Apache-2.0 | 🇪🇺 scaleway | €0.0336 | €0.1344 |
| Pixtral 12B (vision)pixtral-12bvisioncontext 128kOpen weights · GPAI5/6Model card ↗ | Mistral AI (France) | 12B | Apache-2.0 | 🇪🇺 scaleway | €0.0336 | €0.0336 |
| GPT-OSS 120B (OpenAI)gpt-oss-120bOpen weights · GPAI✓ VerifiedModel card ↗Training docs ↗ | OpenAI | 117B (5.1B active) | Apache-2.0 | 🇪🇺 scaleway+1 EU fallback | €0.0252 | €0.1008 |
| Mistral Small 3.2 24Bmistral-small-3.2-24bmultimodalOpen weights · GPAI5/6Model card ↗ | Mistral AI (France) | 24B | Apache-2.0 | 🇪🇺 scaleway+1 EU fallback | €0.0252 | €0.0588 |
| Mistral Large 4 (preview, 512K context)mistral-large-4multimodalcontext 512kGPAI4/6Model card ↗ | Mistral AI (France) | 1T (49B active) | — | 🇪🇺 mistral | €0.2514 | €0.7725 |
| Apertus v1.5 70B (Swiss open model)apertus-v1.5-70bmultimodalcontext 100kOpen weights · GPAI5/6Model card ↗ | Swiss AI Initiative (EPFL, ETH Zurich, CSCS) | 72B | Apache-2.0 | 🇪🇺 infomaniak | €0.1294 | €0.462 |
| Kimi K3.1 Pro (2.8T parameters, 1M context)kimi-k3.1-promultimodalcontext 1MOpen weights · GPAI✓ VerifiedModel card ↗ | Moonshot AI | 2.8T (104B active) | Kimi K3.1-Pro LicenseNot a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇪🇺 verda | €0.3024 | €1.512 |
| Qwen3.8 27B (multimodal)qwen3.8-27bmultimodalcontext 256kOpen weights · GPAI✓ VerifiedModel card ↗ | Qwen | 27.8B | Apache-2.0 | 🇪🇺 aki | €0.0252 | €0.1092 |
| GLM-5.3 (Zhipu/Z.ai)glm-5.3context 512KOpen weights · GPAI✓ VerifiedModel card ↗Training docs ↗ | Zhipu / Z.ai | 753B | GLM-5.3 LicenseNot a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇪🇺 aki+1 EU fallback+1 US fallback (opt-in) | €0.3024 | €0.924 |
| DeepSeek V4.1 Flash (552B MoE, 1M context)deepseek-v4.1-flashmultimodalcontext 1MOpen weights · GPAI✓ VerifiedModel card ↗ | DeepSeek | 763B (8B prefill / 16B decode active) | MIT | 🇪🇺 aki+1 EU fallback+2 US fallback (opt-in) | €0.1008 | €0.3024 |
| Gemma 4 26B (A4B MoE)gemma-4-26bmultimodalcontext 256KOpen weights · GPAI✓ VerifiedModel card ↗ | 26B (4B active) | Apache-2.0 | 🇪🇺 aki | €0.0168 | €0.084 | |
| Kolibri 1 (Aleph Alpha)kolibri-1context 256KOpen weights · GPAI✓ VerifiedModel card ↗ | Aleph Alpha | 78B (3.5B active) | Apache-2.0 | 🇪🇺 aki | €0.0252 | €0.0924 |
| Mistral Small 4 119Bmistral-small-4-119bmultimodalcontext 256KOpen weights · GPAI5/6Model card ↗ | Mistral AI (France) | 119B (6.5B active) | Apache-2.0 | 🇪🇺 mistral | €0.0278 | €0.1109 |
Fast access (US)
7Served through US companies (Modal, OpenRouter) — subject to the CLOUD Act wherever they run, and often served from US regions. We list these models because they are the newest frontier weights and you may want them anyway; we never call them sovereign, and a sovereign variant is tracked separately below when one is planned.
| Model | Publisher | Size | Licence | Hosted via | Price /100k in | Price /100k out |
|---|---|---|---|---|---|---|
| Kimi K3 (2.8T parameters, 1M context)kimi-k3context 1MOpen weights · GPAI3/6Model card ↗ | Moonshot AI | 2.8T (104B active) | Kimi K3 LicenseNot a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇺🇸 modal+1 US fallback | €0.4536 | €2.268 |
| Muse Spark 1.1 (1M context)muse-spark-1.1multimodalcontext 1MGPAI3/6Model card ↗ | Meta | — | Proprietary (API only)Not a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇺🇸 openrouter | €0.205 | €0.6956 |
| Muse Spark 1.2 (1M context)muse-spark-1.2multimodalcontext 1MGPAI3/6Model card ↗ | Meta | — | Proprietary (API only)Not a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇺🇸 openrouter | €0.205 | €0.6956 |
| GLM-5.3 Flash (Zhipu/Z.ai)glm-5.3-flashmultimodalcontext 1MOpen weights · GPAIModel card ↗Training docs ↗ | Zhipu / Z.ai | 320B (18B active) | MIT | 🇺🇸 modal+1 US fallback | €0.0689 | €0.2268 |
| Qwen3.8 Flash (Qwen)qwen3.8-flashmultimodalcontext 1MOpen weights · GPAI3/6Model card ↗ | Qwen | — | Qwen Community License 1.0Not a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇺🇸 openrouter | €0.0252 | €0.0773 |
| DeepSeek V4 Pro (1.6T MoE)deepseek-v4-procontext 1MOpen weights · GPAI5/6Model card ↗ | DeepSeek | 1.6T (49B active) | MIT | 🇺🇸 openrouter | €0.2453 | €0.5124 |
| Gemma 3 27B (Google)gemma-3-27bmultimodalcontext 128kOpen weights · GPAI4/6Model card ↗ | 27B | Gemma Terms of UseNot a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇺🇸 openrouter | €0.0135 | €0.074 |
Temporarily paused
5These models were live and will come back unchanged — same id, same price — as soon as their provider route is restored. Until then they cannot be called and nothing is billed; each fiche says why.
| Model | Publisher | Size | Licence | Hosted via | Price /100k in | Price /100k out |
|---|---|---|---|---|---|---|
| Qwen3.6 27B (multimodal)qwen3.6-27bmultimodalcontext 256kOpen weights · GPAI✓ VerifiedModel card ↗ | Qwen | 27.8B | Apache-2.0 | 🇪🇺 ovhcloud | €0.0672 | €0.4536 |
| Qwen3.5 9B (fast, budget)qwen3.5-9bmultimodalcontext 256kOpen weights · GPAI✓ VerifiedModel card ↗ | Qwen | 9.65B | Apache-2.0 | 🇪🇺 ovhcloud | €0.0168 | €0.0252 |
| Qwen2.5-VL 72B (vision)qwen2.5-vl-72bvisioncontext 32kOpen weights · GPAI5/6Model card ↗ | Qwen | 73.4B | Qwen License AgreementNot a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇪🇺 ovhcloud | €0.1529 | €0.1529 |
| GPT-OSS 20B (OpenAI)gpt-oss-20bOpen weights · GPAI✓ VerifiedModel card ↗Training docs ↗ | OpenAI | 21B (3.6B active) | Apache-2.0 | 🇪🇺 ovhcloud | €0.0068 | €0.0252 |
| Qwen3.8 Max (2.4T, 1M context)qwen3.8-maxcontext 1MOpen weights · GPAIModel card ↗ | Qwen | 2.4T (95B active) | Qwen3.8-Max LicenseNot a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇺🇸 modal+1 US fallback | €0.3024 | €0.9072 |
Coming soon
1Entries already validated (weights public, provider identified) but not callable yet. They carry no price until a provider actually serves them.
| Model | Publisher | Size | Licence | Hosted via |
|---|---|---|---|---|
| Kimi K3 — EU sovereign variantkimi-k3-souveraincontext 1MOpen weights · GPAIModel card ↗ | Moonshot AI | 2.8T (104B active) | Kimi K3 LicenseNot a permissive licence (neither Apache nor MIT) — flag it if you run a legal review. | 🇪🇺 self-hosted |
A dash means we have not verified that value ourselves — not that it does not exist. We would rather leave a gap than publish a licence or a parameter count we have not checked against the publisher.
A curated selection, not every open-source model
An entry is added only once we have verified who serves it and under which licence, cross-checked against the Hugging Face Hub. Prices are based on public per-token tariffs — the provider's own, or for a model on a dedicated endpoint the public market tariff of the same model — not on a cost we measured ourselves; margin and VAT included, the numbers on this page are the ones you are billed, with nothing added at checkout. That is why this catalog is short.
How we draw the sovereignty line: GDPR vs. the CLOUD Act →FAQ
Which models on Frontière AI are GDPR-compliant?
The 19 models labeled EU sovereign. They run on OVHcloud or Scaleway — companies with no non-EU capital control, so a US court order has no direct line to them — and never leave that lane unless you explicitly opt in to a fallback. Models labeled “fast access” are served through US-controlled infrastructure and are marked as such before you call.
What is the cheapest model on Frontière AI?
Gemma 4 26B (A4B MoE), at €0.0168 per 100,000 input tokens — margin and VAT included, that is the amount actually debited from your prepaid balance. Output tokens are billed separately at each model's own rate, listed in the tables above.
Can I switch models without changing my code?
Yes. Every model is reachable through the same OpenAI-compatible endpoint — you change the “model” field in the request body, nothing else. One API key covers the whole catalog, instead of one account per model vendor.
Why is Kimi K3 not labeled EU sovereign?
Because it is served through Modal, a US company, and a US company is subject to the CLOUD Act even when it runs in an EU region — physical location is not what decides jurisdiction. An EU sovereign variant of Kimi K3, planned on dedicated EU servers, is listed separately as coming soon.
Call any of these models in minutes
Create an account, top up €10, change two lines. One key for the whole catalog.