Open-weight models have closed most of the gap with closed frontier labs in 2026, but "open-source" doesn't tell you which license you're under, how much VRAM you need, or whether you can run it on infrastructure outside the CLOUD Act's direct reach. This is a working list of the models we actually route traffic to, what they're good at, and — where it applies — what it costs per million tokens on EU sovereign infrastructure.
If you need one takeaway: GLM-5.2 remains the strongest open-weight generalist served on EU sovereign infrastructure and the one we'd default to for most business use cases; Qwen3.5 397B is the largest multimodal model we serve on EU sovereign infrastructure; and Mistral Small 3.2 is worth a specific look if the fact that the lab itself is European, not just the hosting, is part of your evaluation. Below is the full picture, model by model, with real pricing and licensing — not marketing claims. For most generative AI workloads, the ai inference cost per token is the number that matters at scale.
GLM-5.2 and GLM-5.3 — the all-rounders
Full spec sheet, licence and price →
Built by Zhipu AI / Z.ai, GLM-5.2 is a 753-billion-parameter mixture-of-experts model released under the MIT license — about as permissive as open-source licensing gets, with no restrictions on commercial use or fine-tuning. It's widely regarded as the strongest fully open-weight generalist model available as of mid-2026, competitive with closed frontier models on long-horizon coding and reasoning benchmarks. On Frontière AI, it's served through Scaleway's Generative APIs at €3.024 per million input tokens and €9.24 per million output tokens (every Frontière AI price in this guide is a client price per million tokens, margin and VAT included — the amount debited from your prepaid balance) — a premium price for a premium model, and the one we'd point most teams to first if budget isn't the binding constraint.
Zhipu has since shipped the GLM-5.3 line: GLM-5.3 (under Zhipu’s own GLM-5.3 License) and the lighter, multimodal GLM-5.3 Flash (MIT), both with a 1M-token context window (1,048,576 tokens). They extend the family's long-horizon coding and reasoning profile, but neither is on the sovereign lane: today they are routed as fast access through non-EU-controlled providers — so GLM-5.2 on Scaleway remains the sovereign-lane GLM choice. GLM-5.3 spec sheet → GLM-5.3 Flash spec sheet →
The Qwen family — from Qwen3 to Qwen3.8
Full spec sheet, licence and price →
Alibaba's Qwen family is the deepest open-weight lineup available, and Frontière AI's catalog spans six of its Apache-licensed Qwen3-generation models across three price tiers — plus the newer Qwen3.8 line below, mostly under Qwen's own license, whose 27B has just joined the sovereign lane:
- Qwen3.5 397B (397B total, 17B active — a mixture-of-experts design) is the flagship: multimodal and Apache-2.0 licensed. Both OVHcloud and Scaleway offer it; we route it to one provider only — Scaleway, the cheaper of the two — so there is no automatic failover between them. €1.008 / €6.048 per million tokens (input/output).
- Qwen3 235B is the previous generation's flagship, still fully capable and noticeably cheaper on output: €1.26 / €3.78 per million tokens (input/output).
- Qwen3.6 27B is a newer, smaller multimodal model — Apache-2.0, served by OVHcloud alone and temporarily paused as of 2026-10-20 (OVHcloud is currently refusing requests). Qwen3.6 35B A3B, its sibling from the same generation (a mixture-of-experts with about 3B active parameters, also multimodal), is served by Scaleway at €0.42 / €2.52 per million tokens (input/output).
- Qwen3 Coder 30B is code-specialized — the model to reach for if the workload is CI pipelines, code review, or developer tooling rather than general chat.
- Qwen3.5 9B is the budget tier: small, fast, and cheap enough to run at high volume for simple classification or lightweight chat. It is temporarily paused as of 2026-10-20 (OVHcloud is currently refusing requests).
- Qwen3.8 Max is the newest flagship of the family: about 2.4T total parameters with 95B active, a 1M-token context, and strong reasoning — but its weights carry Qwen’s own Qwen3.8-Max License, so check the terms before commercial redistribution. It entered the catalog as fast access only, and is temporarily paused as of 2026-09-23 (Modal is currently refusing requests).
- Qwen3.8 27B is the surprise of the line: its weights turned out to be Apache-2.0 (verified on its Hugging Face repository, unlike the rest of the 3.8 family — Qwen3.8 Flash ships under the Qwen Community License 1.0, not a permissive one). Better still, it just crossed to EU sovereign infrastructure: OVHcloud picked it up on 2026-09-08 and Scaleway serves it too (its main route on Frontière since 2026-10-19, OVHcloud as fallback); it is a reasoning model with a 256k context, and it costs €0.252 / €1.092 per million tokens (input/output).
Qwen3.6 27B and Qwen3.5 9B have no second sovereign provider: OVHcloud is the only European operator serving them, which is why both are paused today. Of the newer Qwen3.8 entries, only the 27B is routed through a sovereign provider today — the Flash is served as fast access, and the Max is paused.
Mistral Small 3.2 — the European option
Full spec sheet, licence and price →
Mistral AI is a Paris-based lab, which makes Mistral Small 3.2 (24B parameters, Apache-2.0) the one model in this list where the sovereignty argument extends to the model's origin, not just where it's hosted. It supports a wide range of languages — its model card names English, French, German, Spanish, Arabic, Chinese and Hindi among dozens of others, the same list our model page cites — which matters if your business genuinely operates across multiple European markets rather than English-first with translation layered on top.
gpt-oss 120B/20B — OpenAI's own open weights
Full spec sheet, licence and price →
OpenAI released two open-weight models, gpt-oss-120b and gpt-oss-20b, in August 2025 under Apache-2.0. They're not the newest models on this list, but they're inexpensive, well-documented, and carry brand recognition that matters in procurement conversations where "which company built this" is part of the internal sign-off. The 120B is served by Scaleway, with OVHcloud as fallback, at €0.252 / €1.008 per million tokens (input/output); the 20B, served by OVHcloud alone, costs €0.0672 / €0.252 and is temporarily paused as of 2026-10-20 (OVHcloud is currently refusing requests).
DeepSeek V4, Kimi K3 and R1 — the frontier tier
Full spec sheet, licence and price →
The frontier tier has moved quickly since this guide was first published. DeepSeek V4 Flash 0731 (284B total, 13B active, MIT, 1M-token context in its model card, 256k on Scaleway during the preview) has crossed the line: it is now served on Scaleway, making it an EU sovereign option rather than a fast-access compromise — the first V4 variant with that status. DeepSeek V4 Pro (1.6T total, 49B active, MIT) remains on the fast-access tier. Kimi K3, from Moonshot AI, is the largest model in our catalog at 2.8 trillion parameters, with native vision support and a 1M-token context window; a dedicated EU sovereign variant of K3 is planned on Frontière AI's own infrastructure but is not callable yet.
The pace of these releases is the practical lesson: a model can move from fast access to sovereign within a single month once a managed sovereign cloud picks it up. Check the label on the model page before you call — sovereignty is per model and per provider, not a property of the family.
Beyond the original families
Several capable open-weight families sit outside the GLM/Qwen/DeepSeek/Mistral grouping of the first edition of this guide. Llama 3.3 70B from Meta is served on Scaleway, with OVHcloud as fallback, making it one of the few large generalists available on EU sovereign infrastructure from a non-Chinese lab. Meta’s Muse Spark 1.1 and 1.2 (1M-token context, multimodal) are reachable only as fast access today, and they are not open-weight: their weights are not published, so they sit in this catalog as API-only models. Gemma 3 27B from Google is inexpensive and also fast access today, under Google’s Gemma Terms of Use rather than a permissive licence. Two newer releases have joined the catalog since the first edition, both EU sovereign: Mistral Small 4 119B, served directly by Mistral AI in its EU region, and Gemma 4 26B, served by AKI.IO in Germany. Llama 4 Maverick exists publicly but is not yet in the Frontière AI catalog, so no pricing or hosting claim is made for it here.
Licenses at a glance
| Model | License | Notably |
|---|---|---|
| GLM-5.2 | MIT | No restrictions on commercial use |
| Qwen3.5 397B / Qwen3 235B / Qwen3.6 27B / Qwen3 Coder 30B / Qwen3.5 9B | Apache-2.0 | Permissive, patent grant included |
| Mistral Small 3.2 | Apache-2.0 | Dozens of languages documented |
| gpt-oss 120B / 20B | Apache-2.0 | OpenAI's first open-weight release since GPT-2 |
| DeepSeek V4 (Pro/Flash) | MIT | 1M-token context in the V4 Flash 0731 model card (256k served on Scaleway) |
| Qwen2.5-VL 72B | Qwen's own license ("other") | Not Apache/MIT — check terms before commercial redistribution |
| GLM-5.3 | GLM-5.3 License | Successor to GLM-5.2 — not a permissive licence, check the terms |
| GLM-5.3 Flash | MIT | Lighter multimodal sibling, served as fast access today |
| Qwen3.8 Max | Qwen3.8-Max License | Not Apache/MIT — check terms before commercial redistribution |
| Qwen3.8 Flash | Qwen Community License 1.0 | Not Apache/MIT — check terms before commercial redistribution |
| Qwen3.8 27B | Apache-2.0 | Verified on the Hugging Face repo (2026-09-08); served by AKI.IO in Germany at €0.252 / €1.092 per million tokens (input/output) |
| Gemma 3 27B | Gemma Terms of Use | Google’s open-weight line under its own terms, fast access today |
FAQ
Which of these models should I use if I just want the best general-purpose option?
GLM-5.2 for capability, if budget allows it. Qwen3.5 397B if you want the largest multimodal model on the sovereign lane. Mistral Small 3.2 24B if you want a solid, inexpensive default for everyday workloads.
Are DeepSeek V4 and Kimi K3 available today through Frontière AI?
They are. DeepSeek V4 Flash 0731 is served on Scaleway and labeled EU sovereign. DeepSeek V4 Pro (the April 2026 preview) is fast access today. Kimi K3 remains fast access while a dedicated EU sovereign variant is prepared on our own infrastructure.
Which models in this guide are EU sovereign today?
Everything served by OVHcloud or Scaleway: Qwen3 235B, Qwen3.5 397B, GLM-5.2, DeepSeek V4 Flash 0731, Llama 3.3 70B, Qwen3.6 35B A3B, Qwen3 Coder 30B, Pixtral 12B, GPT-OSS 120B, Mistral Small 3.2 24B, Mistral Large 4, Apertus v1.5 70B, Qwen3.8 27B, GLM-5.3, DeepSeek V4.1 Flash, Gemma 4 26B, Kolibri 1 and Mistral Small 4 119B. Fast access today, served by Modal or OpenRouter (US companies): Kimi K3, Muse Spark 1.1, Muse Spark 1.2, GLM-5.3 Flash, Qwen3.8 Flash, DeepSeek V4 Pro and Gemma 3 27B. Temporarily paused, their provider currently refusing requests: Qwen3.6 27B, Qwen3.5 9B, Qwen2.5-VL 72B, GPT-OSS 20B and Qwen3.8 Max — they come back unchanged once the route is restored. The sovereign Kimi K3 variant on our own EU servers is planned, not callable yet.
Why does the same model sometimes cost different amounts depending on the provider?
OVHcloud and Scaleway each publish their own price grid and negotiate their own hosting economics — so the same open-weight model can carry different prices depending on which cloud serves it. Some models have a single sovereign route (Qwen3.5 397B, on Scaleway only); others have two, a main route and a fallback on the other cloud (Llama 3.3 70B and gpt-oss 120B: Scaleway first, OVHcloud as fallback). A model with two routes is priced on the more expensive one, so the price you see does not change when a call falls back.
Is Apache-2.0 or MIT better for a commercial product?
Both are highly permissive and allow commercial use, modification, and redistribution with minimal obligations (attribution, typically). Apache-2.0 additionally includes an explicit patent grant, which some legal teams prefer for that reason. Neither imposes a copyleft requirement to open-source your own product.
What does 'mixture-of-experts' (MoE) mean for models like GLM-5.2 or Qwen3.5 397B?
It means the model has many more total parameters than it actually uses per token — Qwen3.5 397B holds 397B parameters but activates only about 17B per forward pass. That keeps inference cost and latency closer to a much smaller dense model while retaining the capacity of a much larger one.
Ready to try it?
Create an account and call any model in the catalog in minutes.