Mac Studio M3 Ultra 256GBBudget pick
256 GB Unified (+128)
A
Adds memory headroom for longer context windows and future model growth.261.7 tok/s decode
Adds memory headroom for longer context windows and future model growth.
~$6,999 MSRP
LFM2.5 8B A1B needs ~19.1 GB VRAM. Gaudi 3 128GB has 128.0 GB. With Q4_K_M quantization, expect ~1217 tok/s.
Operating mode
Interactive favors responsiveness, while light API and scale-out lean harder on serving readiness. The fit stays the same, but the recommendation lens changes.
Current mode
Balanced
Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.
Select quantization to explore
Fit status
Runs well
Decode
1217.2 tok/s
TTFT
350 ms
Safe context
128K
Memory
19.1 GB / 128.0 GB
The raw memory story may look fine, but the software ecosystem is still a constraint here.
Runtime ecosystem is narrower than CUDA
Intel GPUs can look attractive on memory per dollar, but local AI tooling, kernels, and model coverage are still broader and easier on CUDA today.
Prefer CUDA if you want the path of least resistance
If your goal is maximum runtime coverage, easier troubleshooting, and better support for new local AI releases, CUDA is usually still the safer upgrade path.
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | B | Runs well | 1217.2 tok/s | 350 ms | 128K |
| Coding | B | Runs well | 1217.2 tok/s | 350 ms | 128K |
| Agentic Coding | B | Runs well | 1217.2 tok/s | 350 ms | 128K |
| Reasoning | B | Runs well | 1217.2 tok/s | 350 ms | 128K |
| RAG | B | Runs well | 1217.2 tok/s | 350 ms | 128K |
Inference speed
Estimated decode speed (tokens/sec) for LFM2.5 8B A1B at Q4_K_M across popular GPUs and Apple Silicon, using the fastest local runtime per device. Fastest is RTX 5090 32GB at ~508 tok/s. Speed is memory-bandwidth bound, so cards that fit the whole model in VRAM run far faster than ones that offload to system RAM.
| GPU / Mac | Memory | Quant | Speed (tok/s) | Fits? |
|---|---|---|---|---|
 RTX 5090 32GB | 32 GB | Q4_K_M | 507.8 | Fits |
RX 7900 XTX 24GB | 24 GB | Q4_K_M | 324.8 | Fits |
| RTX 4090 24GB | 24 GB | Q4_K_M | 324.0 | Fits |
| RTX 3090 24GB | 24 GB | Q4_K_M | 277.1 | Fits |
Mac Studio M3 Ultra 256GB | 256 GB | Q4_K_M | 261.7 | Fits |
| RTX 4080 Super 16GB | 16 GB | Q4_K_M | 258.4 | Fits |
Mac Studio M2 Ultra 128GB | 128 GB | Q4_K_M | 218.1 | Fits |
Mac Studio M1 Ultra 128GB | 128 GB | Q4_K_M | 206.8 | Fits |
MacBook Pro M4 Max 128GB | 128 GB | Q4_K_M | 161.7 | Fits |
MacBook Pro M4 Max 64GB | 64 GB | Q4_K_M | 161.7 | Fits |
| RTX 4070 12GB | 12 GB | Q4_K_M | 159.9 | Fits |
MacBook Pro M3 Max 64GB | 64 GB | Q4_K_M | 112.8 | Fits |
MacBook Pro M1 Max 64GB | 64 GB | Q4_K_M | 103.4 | Fits |
| RTX 3060 12GB | 12 GB | Q4_K_M | 100.5 | Fits |
MacBook Pro M4 Pro 48GB | 48 GB | Q4_K_M | 98.8 | Fits |
| RTX 4060 8GB | 8 GB | Q4_K_M | 84.0 | Tight |
Estimates for single-stream decoding at Q4_K_M; real tokens/sec varies with prompt length, context, batch size, and runtime build. Prompt processing (prefill) is faster than the decode figures shown here.
How LFM2.5 8B A1B (8.5B params) fits at each quantization level on Gaudi 3 128GB (128.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 3.3 GB | Low | B62 |
Q3_K_S | 3 | 4.2 GB | Low | B62 |
NVFP4 | 4 | 4.8 GB | Medium | B62 |
Q4_K_M | 4 | 5.2 GB | Medium | B62 |
Q5_K_M | 5 | 6.1 GB | High | B62 |
Q6_K | 6 | 7.0 GB | High | B62 |
Q8_0 | 8 | 9.1 GB | Very High | B63 |
F16Best for your GPU | 16 | 17.4 GB | Maximum | B63 |
Copy-paste commands to run LFM2.5 8B A1B on your machine.
Run
lms load LFM2.5-8B-A1B && lms server startUpgrade options
Yes, Gaudi 3 128GB can run LFM2.5 8B A1B with a B grade (Runs well). Expected decode speed: 1217.2 tok/s.
LFM2.5 8B A1B (8.5B parameters) requires approximately 19.1 GB of memory with Q4_K_M quantization.
The recommended quantization for LFM2.5 8B A1B is Q4_K_M, which balances quality and memory efficiency.
On Gaudi 3 128GB, LFM2.5 8B A1B achieves approximately 1217.2 tokens per second decode speed with a time-to-first-token of 350ms using Q4_K_M quantization.
For coding workloads, LFM2.5 8B A1B on Gaudi 3 128GB receives a B grade with 1217.2 tok/s and 128K context.
On Gaudi 3 128GB, LFM2.5 8B A1B can safely use up to 128K tokens of context. The model's official context limit is 128K, but available memory constrains the safe maximum.
Prefer CUDA if you want the path of least resistance. If your goal is maximum runtime coverage, easier troubleshooting, and better support for new local AI releases, CUDA is usually still the safer upgrade path.
Often yes, if your goal is the easiest setup and the widest runtime support. Intel can offer attractive memory capacity, but CUDA still tends to win on tooling maturity, guides, kernels, and model coverage for local AI.
Paste this snippet into any page to show a live fit card.
<iframe src="https://willitrunai.com/embed/lfm2.5-8b-a1b-on-gaudi-3-128gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>
Preview: