Which size for which organisation?
All boxes share the same stack, the same behaviour, the same sovereignty guarantees. What changes: capacity, supported models, price.
LMbox Compact
For small firms of 5-15 people, no server room needed
Soundproofed desktop chassis, plugs into a standard wall outlet. Built for law firms, confidential SMBs, or branch offices without a dedicated server room.
LMbox Rack
For 30-80 person firms or 50-150 person mid-market with a server room
Standard 2U datacenter chassis. Same CPU/RAM as LMbox Compact, with doubled storage (RAID10), redundant hot-swap PSU and 10 GbE - for IT teams with a server room.
LMbox Rack Pro
For 100-300 person mid-market, multi-BU, heavy RAG - GPU-accelerated
2U rack chassis with doubled RAM, doubled storage, standard NVIDIA RTX A4000 GPU and 25 GbE networking. For mid-market multi-BU deployments with dozens of concurrent agents, massive RAG ingestion, and Gemma 4 MTP speedup.
Full specifications
| Attribute | LMbox Compact ★ | LMbox Rack | LMbox Rack Pro |
|---|---|---|---|
| Hardware | |||
| Form factor | Desktop soundproofed · 21 × 39 × 46 cm (Fractal Define 7 Mini) | 2U rack 19" · 8,7 × 43,7 × 64,8 cm (Supermicro CSE-825 class) | 2U rack 19" + GPU NVIDIA · 8,7 × 43,7 × 64,8 cm |
| Processor | AMD Ryzen 9 7950X · 16 cores / 32 threads · Zen 4 · AVX-512 | AMD Ryzen 9 7950X · 16 cores / 32 threads · Zen 4 · AVX-512 | AMD Ryzen 9 7950X · 16 cores / 32 threads · Zen 4 · AVX-512 |
| DDR5 ECC RAM | 128 Go | 128 Go | 256 Go |
| Storage | 2 To | 4 To | 8 To |
| AI accelerator | CPU inference (AVX-512 + AMX optimisations llama.cpp) | CPU inference (AVX-512 + AMX optimisations llama.cpp) | NVIDIA RTX A4000 · 16 GB VRAM · 140 W TDP · single-slot low-profile |
| Network | 2 × 2,5 GbE · BMC IPMI/Redfish dédié | 2 × 10 GbE + 2 × 1 GbE management · BMC IPMI/Redfish dédié | 2 × 25 GbE + 2 × 1 GbE management · BMC IPMI/Redfish dédié |
| Physical | |||
| Dimensions (h × w × d) | 21 × 39 × 46 cm (mid-tower compact micro-ATX soundproofed) | 8,7 × 43,7 × 64,8 cm (2U rack 19", alim redondante) | 8,7 × 43,7 × 64,8 cm (2U rack 19", alim redondante 2× 1000W) |
| Weight | 12.0 kg | 18.0 kg | 22.0 kg |
| Power draw | 180 W | 220 W | 380 W |
| Noise level | 25 dB | 55 dB | 58 dB |
| Performance | |||
| Concurrent users | 10 users | 25 users | 100 users |
| Avg throughput | ≈ 22 tok/s | ≈ 22 tok/s | ≈ 80 tok/s |
| Time-to-first-token | ≈ 800 ms | ≈ 800 ms | ≈ 320 ms |
| Models | |||
| Supported models | Mistral Small 4 MoE 119B/6B · Gemma 4 31B + MTP drafter · Qwen 3.6 27B · bge-m3 embedding · Whisper medium | Mistral Small 4 MoE 119B/6B · Gemma 4 31B + MTP drafter · Qwen 3.6 27B · bge-m3 embedding · Whisper medium | Mistral Small 4 MoE 119B/6B · Gemma 4 31B + MTP drafter (GPU-accelerated) · Qwen 3.6 27B · Codestral 22B · bge-m3 (GPU embeddings) · Whisper large-v3 |
| Pricing | |||
| CAPEX (one-shot) | 22 000 € | 25 000 € | 52 000 € |
| Annual support | 6 000 € / an | 6 000 € / an | 12 000 € / an |