← Back to the AI Box
Compare S, M, L and XL
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 S
For teams of 20-40 people
Compact, silent, plugs into a standard wall outlet.
Concurrent users
up to 40
Avg throughput
≈ 28 tok/s
Time-to-first-token
≈ 800 ms
Power · noise
65 W · 22 dB
★ Most popular
LMbox M
For mid-market organisations of 50-150
Reference build: enough horsepower to run Gemma 4 31B comfortably, no exotic hardware required.
Concurrent users
up to 150
Avg throughput
≈ 52 tok/s
Time-to-first-token
≈ 450 ms
Power · noise
180 W · 32 dB
LMbox L
For 150+ users or multi-site
Maximalist Apple Silicon: 192 GB unified RAM for heavy models or hundreds of concurrent users.
Concurrent users
up to 400
Avg throughput
≈ 85 tok/s
Time-to-first-token
≈ 280 ms
Power · noise
350 W · 40 dB
LMbox XL
Datacenter build — frontier MoE models, locally
4U Linux rack with 4–8× NVIDIA H100/H200 — the only way to run open-weights frontier models (Kimi K2.6, DeepSeek V4 Pro) fully on-prem.
Concurrent users
up to 1500
Avg throughput
≈ 220 tok/s
Time-to-first-token
≈ 150 ms
Power · noise
3500 W · 65 dB
Full specifications
| Attribute | LMbox S | LMbox M ★ | LMbox L | LMbox XL |
|---|---|---|---|---|
| Hardware | ||||
| Form factor | Compact desktop · 5 × 19,7 × 19,7 cm | Mac Studio compact · 9,5 × 19,7 × 19,7 cm | Mac Studio M3 Ultra MAX ou Linux 1U rackable | Linux 4U rackable 19" — multi-GPU NVIDIA |
| Processor | Apple M4 Pro · 14 cœurs CPU · 20 cœurs GPU | Apple M3 Ultra · 32 cœurs CPU · 80 cœurs GPU | Apple M3 Ultra max ou AMD EPYC 7763 (variante x86) | AMD EPYC 9754 128-core ou Intel Xeon Platinum 8592+ |
| Unified RAM | 64 Go | 96 Go | 192 Go | 1024 Go |
| Storage | 2 To | 4 To | 8 To | 16 To |
| AI accelerator | Apple Neural Engine 16 cœurs (38 TOPS) | Apple Neural Engine 32 cœurs (60 TOPS) · 80-core GPU | Apple Neural Engine 32 cœurs (Mac) ou 2 × NVIDIA L40S (variante x86) | 4 × NVIDIA H100 80 Go (config base) ou 8 × H200 141 Go (config max) |
| Network | 1 × 10 GbE · Wi-Fi 6E · Thunderbolt 4 | 2 × 10 GbE · Wi-Fi 6E · 6 × Thunderbolt 4 | 2 × 10 GbE · 1 × 25 GbE optionnel · IPMI (variante x86) | 2 × 25 GbE · 1 × 100 GbE InfiniBand optionnel · IPMI 2.0 · BMC |
| Physical | ||||
| Dimensions (h × w × d) | 5 × 19,7 × 19,7 cm | 9,5 × 19,7 × 19,7 cm | Mac Studio compact · ou 1U rack 19" | 4U rack 19" · 175 × 482 × 712 mm |
| Weight | 1.4 kg | 3.6 kg | 3.6 kg | 32.0 kg |
| Power draw | 65 W | 180 W | 350 W | 3500 W |
| Noise level | 22 dB | 32 dB | 40 dB | 65 dB |
| Performance | ||||
| Concurrent users | 40 users | 150 users | 400 users | 1500 users |
| Avg throughput | ≈ 28 tok/s | ≈ 52 tok/s | ≈ 85 tok/s | ≈ 220 tok/s |
| Time-to-first-token | ≈ 800 ms | ≈ 450 ms | ≈ 280 ms | ≈ 150 ms |
| Models | ||||
| Supported models | Gemma 4 9B · Mistral Small 3 · Llama 3 8B · Qwen 2.5 7B | Gemma 4 31B · Mistral Large 2 · Llama 3 70B (q4) · Qwen 2.5 32B · Codestral 22B | Gemma 4 31B fp16 · Llama 3 70B · Qwen 3.6 32B · GLM 5.1 · Mistral Large 2 · Mixtral 8x22B | Kimi K2.6 · DeepSeek V4 Pro · MiMo V2.5 Pro (1M ctx) · Qwen 3.6 Max · Llama 3 70B fp16 · Mistral Large 2 fp16 |
| Pricing | ||||
| CAPEX (one-shot) | 12 000 € | 25 000 € | 38 000 € | 95 000 € |
| Annual support | 4 800 € / an | 9 600 € / an | 14 400 € / an | 28 800 € / an |
Decision helper
How to choose?
LMbox S
Starting or piloting
6-month PoC, pilot team, 30-person subsidiary, or consulting firm with occasional needs. The S checks all boxes at a reasonable price.
LMbox M
Deploying at scale
Our default recommendation: mid-market 50–150 people, 30B models at full resolution, low latency, no exotic hardware needed.
LMbox L
You have hard constraints
Multi-site, 70B+ models, > 80 tok/s, 1U datacenter rack (Linux x86), or heavy concurrent loads.
LMbox XL
Frontier models, fully on-prem
Datacenter with multi-GPU NVIDIA H100. To run Kimi K2.6 / DeepSeek V4 Pro / MiMo 1M with zero cloud calls. 200+ dev mid-market or critical-infra with maximum sovereignty requirements.