An OpenAI-compatible API to open-weights models on our own hardware in Austria. Aliases that carry policy. Privacy in stages, up to an LLM-backed check. And every number on this page is measured nightly, not promised.
Your app talks to a name. What happens behind it, which model, which provider, how much privacy, you configure here and change any time, without touching a line of code. Try it:
Open-weights models on hardware we own, in Upper Austria. No sub-processors, no third countries, no reselling of third-party APIs.
An alias bundles model, provider and policy. It points at our own models, or, with your own key, outside. Switching models means: change it here, nothing else.
Per alias in stages: names, addresses, IBANs, and API keys, tokens and credentials that end up in prompts by accident. Only what's allowed to leave, leaves.
For companies that want to use AI without guessing at the hardware.
Which models you really need, and how much compute they take, nobody knows up front. So we reverse the order: you start on our infrastructure, your actual usage shows what you need, and only then do you buy matching hardware. Our software runs it, with the same API as here. We support the move.
More on whitelabel →Developers and agencies who need a stable EU endpoint for their own projects. Universities and schools for teaching and research, if the requirements fit.
High-risk applications such as recruitment, credit scoring or assessing learners.
Both are valid paths. The difference is invoicing, privacy, and who you email when something breaks.
The smart part was never the model. It's what surrounds it: routing, policy, privacy, proof.
An alias is more than a model name, it bundles model, provider and privacy policy. Swap the model centrally; your application never changes.
Off, pseudonymize, or go strict: an LLM on our own hardware double-checks for indirect identifiers before anything leaves the EU.
Our own hardware in Austria. No sub-processors, no third countries. External frontier models are reachable, but only through the masking layer.
Nightly benchmarks with published methodology. Detection recall for the privacy layer as a number, not an adjective.
Up to four targets per alias. If the primary provider flakes (429, 5xx, timeout), the request walks to the next one automatically. Across ALLSmartLLM and OpenRouter Free if you want.
Plug in your OpenAI, Anthropic or OpenRouter account. Keys land AES-256-GCM encrypted in a vault, master key from the OS keyring. Plaintext lives in request memory only. Never in logs, never in backups.
Rule-based policy checks per alias. „log_only" writes violations to the audit log, „enforce" breaks the request. For compliance proof and hard limits: no customer data into ext. models, no card numbers in the response stream, no PII into training data.
One key with us, every major upstream underneath: our own ALLSmartLLM hardware, OpenRouter (~300 models), OpenAI, Anthropic, Groq, DeepSeek. Switching provider means re-pointing an alias.
If your software speaks OpenAI, it already speaks ALLSmartLLM. Change one URL.
10M tokens free to test and set up. No card, no subscription.
One key, all models, all aliases. Usage-based, nothing else.
Call an alias, not a model version. Privacy is one parameter.
curl https://allsmartllm.com/v1/chat/completions \ -H "Authorization: Bearer $KEY" \ -d '{ "model": "frontier", "privacy": {"mode": "pseudonymize"}, "messages": [...] }'
Later, host it yourself: same API on your own hardware. When your box is running, you change one URL, done.
Vetted, benchmarked, and kept current. Aliases follow our recommendation, or pin a version and stay.
| Model | Alias | Can do | Context | License | Quality* | Measured | Status |
|---|---|---|---|---|---|---|---|
|
Qwen 3.8 27B
27B dense (NVFP4)
|
frontier
|
ToolsBildReasoning | 256K | Apache 2.0 | 28 | 38.0 t/s | loaded |
|
Qwen 3.8 Flash-Next
FP4 (NVIDIA FP4-Local, modelopt_mixed)
|
- | ToolsBildReasoning | 256K | Qwen Community 1.0 + NVIDIA Open Model Agreement |
40 | 36.5 t/s | internal use only |
|
Nemotron 3 Nano Omni
33B / 3B MoE (Q4_K_M)
|
- | ToolsBildReasoning | 256K | NVIDIA OML¹ | 10 | 41.8 t/s |
on-demand
First start 28 s
|
|
Nemotron 3.5 Bolt
74B / 9B MoE (IQ2_XXS (QAD))
|
judge
|
ToolsBildReasoning | 256K | OpenMDW-1.1 | 21 | 74.3 t/s | loaded |
|
Nemotron Cascade 2
32B dense (Q4_K_M)
|
- | ToolsReasoning | 256K | NVIDIA OML¹ | 12 | 44.2 t/s |
on-demand
First start 0,3 s
|
|
Phi 4
14B dense (Q4_K_M)
|
- | Text | 16K | MIT | 6 | 15.0 t/s |
on-demand
First start 3,5 s
|
|
Qwen 3.6 35B
35B / 3B MoE (FP8)
|
midvisionworker
|
ToolsBildReasoning | 256K | Apache 2.0 | 18 | 79.3 t/s | loaded |
|
Qwen 3.5 35B
35B / 3B MoE (Q4_K_M)
|
- | ToolsBildReasoning | 256K | Apache 2.0 | 19 | 33.2 t/s |
on-demand
First start 28 s
|
|
Granite 4.1
30B dense (Q4_K_M)
|
- | Tools | 256K | Apache 2.0 | 7 | 6.8 t/s |
on-demand
First start 42 s
|
|
Nemotron 3 Nano
30B / 3.5B MoE (Q4_K_M)
|
- | ToolsReasoning | 256K | NVIDIA OML¹ | 9 | 44.1 t/s |
on-demand
First start 6,3 s
|
|
Nemotron 3.5 Lightning
30B / 3B MoE (NVFP4 (ModelOpt, MIXED_PRECISION))
|
fastswift
|
ToolsReasoning | 256K | OpenMDW-1.1 | 13 | 110.1 t/s | loaded |
|
Qwen 3.6 27B
27B dense (Q4_K_M)
|
- | ToolsBildReasoning | 256K | Apache 2.0 | 21 | 7.4 t/s |
on-demand
First start 0,7 s
|
|
Gemma 4 26B
26B / 4B MoE (Q4_K_M)
|
- | ToolsBildReasoning | 256K | Apache 2.0 | 17 | 34.2 t/s |
on-demand
First start 33 s
|
|
Qwen3 Coder Next
80B dense (Q4_K_M)
|
coder
|
Tools | 256K | Apache 2.0 | 9 | 32.5 t/s |
on-demand
First start 0,4 s
|
|
Qwen3 Coder
31B dense (Q4_K_M)
|
- | Tools | 256K | Apache 2.0 | 10 | 42.0 t/s |
on-demand
First start 12 s
|
|
Qwen 3.5 9B
9B dense (Q4_K_M)
|
- | ToolsBildReasoning | 256K | Apache 2.0 | 13 | 18.4 t/s |
on-demand
First start 25 s
|
These two models produce no text. They take a state and typed questions and return a probability for every answer option, in a single pass, through a separate endpoint. That is why they are here and not above: they share hardly a column with the language models.
| Model | Alias | Context | Licence | Time per decision | Status |
|---|---|---|---|---|---|
|
Clef Flash
9B dense (Q8_0)
|
clef-flash
|
256K | - | 445 ms | loaded |
|
Clef
27B dense (Q4_K_M)
|
clef
|
256K | - | 1,1 s |
on-demand
First start 12 s
|
Input and output priced separately, tiered by model category. 10M tokens free. No base fee, no subscription, usage only.
| Category | Input | Output |
|---|---|---|
| Reasoning | € 0.75 / Mtok | € 3 / Mtok |
| Multimodal | € 0.75 / Mtok | € 3 / Mtok |
| Code | € 0.50 / Mtok | € 2 / Mtok |
| Standard | € 0.50 / Mtok | € 2 / Mtok |
| Schnell | € 0.25 / Mtok | € 1 / Mtok |
| Entscheidung | € 0.25 / Mtok | € 1 / Mtok |
| Privacy detector | € 0.10 / Mtok | only when active, on input tokens |
This list comes from the log schema, not from a marketing page. Nothing you send through the API is stored: neither your prompts nor the responses.
| Field | Via the API | What for |
|---|---|---|
| Your prompts | not stored | Never travel from the gateway to a database. |
| The model's responses | not stored | Delivered, then discarded. |
| Timestamp | stored | Billing and troubleshooting. |
| Model requested and used | stored | So you can see what actually answered. |
| Token counts (in and out) | stored | The basis for billing. |
| Request duration | stored | Shows when a backend slows down. |
| Status and error message | stored | Support: without it we cannot help you. |
| Endpoint called | stored | Billing and abuse detection. |
| IP address and user agent | stored | Abuse detection and rate limits. 30 days. |
| Privacy mode selected | stored | Evidence of which level applied to the request. |
/v1/. Storage is on our own hardware in Upper Austria.