Full-Stack Physical AI
Sovereign AI on your own infrastructure — open-weight models you own, run and govern, adapted to your domain.
Select & size
Deploy
Adapt
Govern & evaluate
A four-stage pipeline: select and size the model, then deploy, then adapt by fine-tuning, then govern and evaluate.
Many industrial settings can't send data to a third-party API — for sovereignty, latency, or regulation. Open-weight models run on your own infrastructure, adapted to your domain, fully under your control. We help you choose, size, deploy and govern them — so capability doesn't cost you ownership of your data. Mistral is our first choice, but the recommendation is workload-selected and vendor-independent: depending on quality, latency, cost, sovereignty and operational constraints, the right answer may be Mistral, another open-weight model, a managed frontier model, retrieval without fine-tuning, fine-tuning, or no LLM at all.
Model selection and sizing for the task; a deployment topology (on-premise, edge, or private cloud); domain adaptation through fine-tuning or retrieval; an inference-serving layer; and a governance and evaluation loop, so quality is measured rather than assumed.
Mistral and other open-weight models; quantization (GGUF, AWQ, GPTQ) for the target hardware; serving with vLLM, llama.cpp or Ollama; LoRA/QLoRA fine-tuning; retrieval integration; evaluation harnesses for quality and regression; on-premise GPU and edge deployment.
A dated snapshot of the Mistral model line — Hyperion's first-choice runtime, workload-selected and vendor-independent. Each entry is vendor-documented; verify a model at its official source before relying on it.
Mistral Large 3 25.12
Flagship multimodal LLM · Released 2025-12-01 · 128K+ tokens · text + vision
Licence: Open-weight (commercial licence for production — confirm at source)
Mistral Medium 3.5 26.04
Frontier multimodal / agentic + coding LLM · Released 2026-04-01 · 128K+ tokens · text + vision
Licence: Commercial / Mistral platform
Mistral Small 4 26.03
Efficient hybrid LLM (instruct + reasoning + coding) · Released 2026-03-01 · 128K+ tokens · text + vision
Licence: Apache-2.0
Ministral 3 (14B / 8B / 3B) 25.12
Edge-class LLM family (14B / 8B / 3B) · Released 2025-12-01 · 128K tokens · text
Licence: Mixed (open-weight + platform — confirm per size at source)
Codestral 25.08
Code-specialised LLM · Released 2025-08-01 · 256K tokens · text (code)
Licence: Mistral AI Non-Production License (commercial licence for production)
Devstral 2 / Devstral Small 2 25.12
Agentic coding LLM · Released 2025-12-10 · 128K+ tokens · text (code)
Licence: Mixed (Devstral Small 2 open; Devstral 2 platform — confirm at source)
Pixtral Large 24.11
Multimodal (vision) model · Released 2024-11-18 · 128K tokens · text + vision
Licence: Mistral Research License (commercial licence for production)
Registry verified 2026-06-28
Sovereignty, cost and longevity usually favour models you host yourself.
| Dimension | Open-weight (you run) | Proprietary API |
|---|---|---|
| Ownership | You host the weights | Vendor-hosted |
| Data residency | Stays in your perimeter | Sent to the vendor |
| Cost model | Fixed compute you control | Per-token, unbounded |
| Adaptation | Full fine-tune / LoRA | Limited to what's exposed |
| Longevity | Pin a version you trust | Deprecation risk |