Qwen3.6 35B A3B model. 55+ tokens/sec. $300 GPU.

No, this isn't a server card. It's an RTX 4060 Ti 8GB.

Previously I posted that I 41 t/s on this gpu and that post blew up and went viral. I went back and made it 34% faster.

And now the speed doesn't drop with context depth at all.

New benchmarks + what changed 🧵

Before (mainline llama.cpp, Q4_K_S, q8_0 KV):
• 41 t/s at 16k context
• 24 t/s at 200k context ← −42% over depth

After (ik_llama.cpp, IQ3_K_R4, q4_0 KV):
• 55 t/s at 16k context
• 55 t/s at 163k context ← completely flat

Same GPU. Same model. Same CPU offload trick.
| depth | mainline | ik_llama.cpp | gain |
|------:|---------:|-------------:|-----:|
| 16k | 41.6 t/s | 57.6 t/s | +38% |
| 65k | ~34 t/s | 54.3 t/s | +60% |
| 131k | ~28 t/s | 55.1 t/s | +97% |
| 163k | ~26 t/s | 55.6 t/s | +114% |

Results are slightly noise, as I was running lots of tabs on Chrome in parallel and was having a few windows of Claude Code open as well!

Two things:

**1. IQ3_K_R4 quant (ik_llama.cpp format)**
The R4 format reorders expert FFN weights specifically for the CPU+GPU split — better L3 cache locality per token.
3.4 bpw. Smaller than Q4_K_S. Yet faster.

**2. q4_0 KV cache** (was q8_0)
Halves KV memory per token. The original bottleneck at 200k was GPU scanning a huge q8_0 KV cache. Switch to q4_0 and that bottleneck disappears.

With -ncmoe 99, every token streams ~11 GB of expert weights through DDR5.

That's constant. Context depth doesn't change it.

Previously, q8_0 KV scanning was eating ~40% throughput at 200k. It was competing for the same GPU pipeline.

q4_0 KV is half the data. FlashAttention handles it in microseconds. CPU experts become the only bottleneck — and they don't care how long your context is.

There's a knob: how many of the 41 expert layers to keep on GPU.
| config | avg TG | max context | peak VRAM |
|--------|-------:|------------:|----------:|
| ncmoe=99 (all on CPU) | ~55 t/s | **~196k tokens ✓** | 4.4 GB |
| ncmoe=30 (11 on GPU) | ~60 t/s | ~163k tokens | 7.5 GB |
ncmoe=99 is the safe default — 196k context fits with 3+ GB VRAM to spare. You can fit full 262k context here!
ncmoe=30 gets you +8% speed but eats 7.5 GB at 163k — no room for more context.
8GB is tight. Pick your priority: speed or context.

32Gb of RAM is enough!

Expert weights (~11 GB) live in system RAM via mmap.
VRAM holds attention layers + KV cache only (~4.5 GB at 196k).
| component | RAM usage |
|-----------|----------:|
| Expert weights (paged in) | ~11 GB |
| OS + desktop + process | ~4–5 GB |
| **Total** | **~15–16 GB** |
Fits 32GB machines comfortably.

Recipe:
```bash
# Engine: ik_llama.cpp (ikawrakow/ik_llama.cpp)
# Model: IQ3_K_R4 — re-quantized from Q8_0

llama-server \
--model Qwen3.6-35B-A3B-IQ3_K_R4.gguf \
-ngl 99 --n-cpu-moe 99 -fa 1 \
-ctk q4_0 -ctv q4_0 \
-c 131072 -t 12
```
Model on HuggingFace: huggingface.co/abovespec/Qwen…

IQ3_K_R4 only works with ik_llama.cpp — not mainline llama.cpp.

Hardware used: RTX 4060 Ti 8GB · Ryzen 9 7900X · 96GB DDR5

There's a knob: how many of the 41 expert layers to keep on GPU.
| config | avg TG | max context | peak VRAM |
|--------|-------:|------------:|----------:|
| ncmoe=99 (all on CPU) | ~55 t/s | ~196k tokens | 4.5 GB |
| ncmoe=30 (11 on GPU) | ~60 t/s | ~163k tokens | 7.5 GB |
More experts on GPU = faster but less context headroom.
8GB is tight. ncmoe=99 is the safe default.

What GPU are you running local LLMs on?

8GB cards are way more capable than people think in 2026.

Drop your setup below 👇

My original thread where I got 41 tps is here:

View Tweet