Can Surface Laptop Ultra · RTX Spark N1X (18-core) · 24 GB run LFM2.5 24B-A2B?
24 GB unified · 300 GB/s memory bandwidth · LFM2.5 24B-A2B has 24B parameters (2B active per token).
Short answer
It loads, but it's borderline even for chat.
Best quantization for chat: Q4_K_M (14 GB of weights) on Ollama. Generation: 85–150 tok/s, leaving about 0.9 GB free for your other apps.
How it rates for each workload
Same model and machine, different jobs. Agents and long documents send far bigger prompts than chat, so they need much more speed and memory.
| Workload | Verdict | Quant | Generation | First reply | Context | |
|---|---|---|---|---|---|---|
| Chatwith Open WebUI | Borderline | Q4_K_M | 85–150 tok/s | under 1 s | 8K | Details |
| Coding questionswith Continue | Does Not Fit | — | — | — | — | Details |
| Coding in a repositorywith Aider | Does Not Fit | — | — | — | — | Details |
| Agentic codingwith OpenCode | Does Not Fit | — | — | — | — | Details |
| Long documents (64K)with Open WebUI | Borderline | Q4_K_M | 50–80 tok/s | ≈5 s | 64K | Details |
Speeds are bandwidth-based estimates; no direct benchmark exists for this pair. “First reply” is the time to read the workload’s opening prompt. How we calculate
Other models for Surface Laptop Ultra · RTX Spark N1X (18-core) · 24 GB
The best-rated models for chat on this machine.
- gpt-oss-20b50–85 tok/sExcellent
- MiniCPM5 2B55–90 tok/sExcellent
- Gemma 4 26B-A4B60–95 tok/sExcellent
- Qwen3.5 4B35–55 tok/sExcellent
- Qwen3.5 9B18–30 tok/sExcellent
- Gemma 4 12B20–35 tok/sExcellent
Cheapest machines for LFM2.5 24B-A2B
Hardware rated comfortable or better for chat, lowest price first.
- Mac mini M4 24 GB$999 · 45–60 tok/sComfortable
- Mac mini M6 24 GB$1,299 · 50–85 tok/sComfortable
- Laptop · Ryzen AI 9 HX 370 (Radeon 890M) · 32 GB$1,300 · 35–60 tok/sExcellent
- Laptop · Core Ultra 7 258V (Arc 140V) · 32 GB$1,300 · 30–50 tok/sExcellent
- Mac mini M4 32 GB$1,399 · 45–60 tok/sExcellent
- MacBook Air M4 24 GB$1,399 · 45–60 tok/sComfortable
Same chip, other memory: 128 GB
Different tool, context size or workload?
These pages use typical settings. The full check models your exact setup.