Hermes 4.3 36B needs ~26.2 GB VRAM. MacBook Pro M4 32GB has 23.0 GB. With Q3_K_S quantization, expect ~7 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
7.5 GB over capacity — needs offload or smaller quantization
Fit status
Too heavy
Decode
4.8 tok/s
TTFT
40503 ms
Safe context
4K
Memory
30.5 GB / 23.0 GB
Offload
20%
It fits through host-memory offload, and offload is the main reason performance drops.
CPU or host-memory offload is active
About 10% of the working set spills out of accelerator memory, which usually hurts latency and sustained decode throughput.
Very little memory headroom
You can run the model, but there is not much room left for longer context, bigger batches, extra apps, or future model updates.
Shared-memory contention still exists
The OS, browser, and inference runtime all compete for the same physical memory pool, so real-world headroom is less forgiving than raw capacity suggests.
Remove offload with more accelerator memory
Prioritize a GPU or unified-memory tier that fits the whole model natively. Removing offload usually helps more than small compute gains.
Buy headroom, not only minimum fit
A slightly larger memory tier gives you safer context growth and makes the recommendation more future-proof.
Increase host RAM if you keep offloading
This setup may need roughly 2.1 GB of extra host RAM just for the offloaded portion, before OS and other tools.
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | F | Too heavy | 5.2 tok/s | 20266 ms | 4K |
| Coding | F | Too heavy | 4.8 tok/s | 40503 ms | 4K |
| Agentic Coding | F | Too heavy | 4.1 tok/s | 68432 ms | 4K |
| Reasoning | F | Too heavy | 4.8 tok/s | 47867 ms | 4K |
| RAG | F | Too heavy | 4.1 tok/s | 85540 ms | 4K |
Inference speed
Estimated decode speed (tokens/sec) for Hermes 4.3 36B at Q4_K_M across popular GPUs and Apple Silicon, including multi-GPU rigs, using the fastest local runtime per device. Fastest is RTX 5090 32GB at ~55 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 | 54.7 | Offloads |
2× RX 7900 XTX 24GB | 48 GB | Q4_K_M | 53.5 | Fits |
| 2× RTX 4090 24GB | 48 GB | Q4_K_M | 28.1 | Fits |
MacBook Pro M4 Max 128GB | 128 GB | Q4_K_M | 27.3 | Fits |
MacBook Pro M4 Max 64GB | 64 GB | Q4_K_M | 27.3 | Fits |
| 2× RTX 3090 24GB | 48 GB | Q4_K_M | 25.8 | Fits |
Mac Studio M3 Ultra 256GB | 256 GB | Q4_K_M | 25.4 | Fits |
| 4× RTX 3060 12GB | 48 GB | Q4_K_M | 22.7 | Fits |
Mac Studio M2 Ultra 128GB | 128 GB | Q4_K_M | 21.1 | Fits |
Mac Studio M1 Ultra 128GB | 128 GB | Q4_K_M | 20.0 | Fits |
MacBook Pro M4 Pro 48GB | 48 GB | Q4_K_M | 17.2 | Tight |
| RTX 4090 24GB | 24 GB | Q4_K_M | 16.6 | Too big |
RX 7900 XTX 24GB | 24 GB | Q4_K_M | 15.3 | Too big |
| RTX 3090 24GB | 24 GB | Q4_K_M | 14.2 | Too big |
MacBook Pro M3 Max 64GB | 64 GB | Q4_K_M | 10.9 | Fits |
MacBook Pro M1 Max 64GB | 64 GB | Q4_K_M | 10.0 | Fits |
| RTX 4080 Super 16GB | 16 GB | Q4_K_M | 6.0 | Too big |
| RTX 4070 12GB | 12 GB | Q4_K_M | 2.6 | Too big |
| RTX 3060 12GB | 12 GB | Q4_K_M | 2.0 | Too big |
| RTX 4060 8GB | 8 GB | Q4_K_M | 2.0 | Too big |
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 Hermes 4.3 36B (36B params) fits at each quantization level on MacBook Pro M4 32GB (23.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_KBest for your GPU | 2 | 14.0 GB | Low | C50 |
Q3_K_S | 3 | 17.6 GB | Low | F0 |
NVFP4 | 4 | 20.2 GB | Medium | F0 |
Q4_K_M | 4 | 22.0 GB | Medium | F0 |
Q5_K_M | 5 | 25.9 GB | High | F0 |
Q6_K | 6 | 29.5 GB | High | F0 |
Q8_0 | 8 | 38.5 GB | Very High | F0 |
F16 | 16 | 73.8 GB | Maximum | F0 |
Copy-paste commands to run Hermes 4.3 36B on your machine.
Run
lms load hf-nousresearch--hermes-4-3-36b-gguf && lms server startOpciones de mejora
Hace que el modelo quepa en el acelerador en lugar de seguir fuera de alcance.
Elimina el offload a memoria del sistema, que suele ser la mayor mejora individual en latencia y throughput.
~$1,099 MSRP
Hace que el modelo quepa en el acelerador en lugar de seguir fuera de alcance.
Elimina el offload a memoria del sistema, que suele ser la mayor mejora individual en latencia y throughput.
~$1,599 MSRP
Hace que el modelo quepa en el acelerador en lugar de seguir fuera de alcance.
~$1,999 MSRP
Yes, MacBook Pro M4 32GB can run Hermes 4.3 36B at Q3_K_S quantization (Very compromised (needs ~2.1 GB host RAM)). The recommended Q4_K_M requires 30.5 GB which exceeds available memory, but at Q3_K_S it needs only 26.2 GB. Expected decode speed: 6.7 tok/s.
Hermes 4.3 36B (36B parameters) requires approximately 30.5 GB at Q4_K_M quantization. On MacBook Pro M4 32GB, it fits at Q3_K_S using 26.2 GB.
The recommended quantization is Q4_K_M, but on MacBook Pro M4 32GB the best fitting quantization is Q3_K_S, which uses 26.2 GB.
On MacBook Pro M4 32GB, Hermes 4.3 36B achieves approximately 6.7 tokens per second decode speed with a time-to-first-token of 28962ms using Q3_K_S quantization.
For coding workloads, Hermes 4.3 36B on MacBook Pro M4 32GB receives a F grade with 4.8 tok/s and 4K context.
On MacBook Pro M4 32GB, Hermes 4.3 36B can safely use up to 4K tokens of context at Q3_K_S quantization. The model's official context limit is —, but available memory constrains the safe maximum.
Remove offload with more accelerator memory. Prioritize a GPU or unified-memory tier that fits the whole model natively. Removing offload usually helps more than small compute gains.
Not always. MacBook Pro M4 32GB can often fit larger models thanks to unified memory, but a discrete GPU with dedicated high-bandwidth VRAM may still decode faster once the model fits. For this combination, the important distinction is capacity versus sustained throughput.
Paste this snippet into any page to show a live fit card.
<iframe src="https://willitrunai.com/embed/hf-nousresearch--hermes-4-3-36b-gguf-on-m4-32gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>
Preview: