Upload README.md with huggingface_hub

README.md

@@ -75,7 +75,7 @@ cmake --build build --target llama-cli llama-server
 
 ```bash
 llama-server -m Qwen3.6-27B-Q5_K_M-mtp.gguf \
-  --spec-type mtp --spec-draft-n-max 5 \
+  --spec-type mtp --spec-draft-n-max 3 \
   --cache-type-k q4_0 --cache-type-v q4_0 \
   -np 1 -c 262144 --temp 0.7 --top-k 20 -ngl 99 --port 8081
 ```
@@ -84,7 +84,7 @@ That's it. Three optimizations in one command:
 
 | Flag | What it does | Impact |
 |---|---|---|
-| `--spec-type mtp --spec-draft-n-max 5` | Multi-Token Prediction (built into the model) | **2.5x faster** generation |
+| `--spec-type mtp --spec-draft-n-max 3` | Multi-Token Prediction (built into the model) | **2.5x faster** generation |
 | `--cache-type-k q4_0 --cache-type-v q4_0` | 4-bit KV cache (instead of 16-bit) | **Quarter the KV memory** |
 | `-c 262144` | 262K context window | Full native context on **48 GB Mac** with q4_0 KV |
 
@@ -225,7 +225,7 @@ curl http://localhost:8081/v1/messages \
 ```bash
 # Text generation
 llama-cli -m Qwen3.6-27B-Q5_K_M-mtp.gguf \
-  --spec-type mtp --spec-draft-n-max 5 \
+  --spec-type mtp --spec-draft-n-max 3 \
   --cache-type-k q8_0 --cache-type-v q8_0 \
   -np 1 -c 4096 -n 2048 --temp 0.7 -ngl 99 \
   -p "Your prompt here"
@@ -270,12 +270,12 @@ Effect on hardware requirements (Q5_K_M, 80K context):
 The model predicts 5 extra tokens per step using its own MTP heads, then verifies them in one pass. No extra model needed.
 
 ```bash
---spec-type mtp --spec-draft-n-max 5 -np 1
+--spec-type mtp --spec-draft-n-max 3 -np 1
 ```
 
 MTP currently requires `-np 1` (single-sequence mode). Without it, you'll get: `MTP currently supports only n_parallel=1`.
 
-Tune `--spec-draft-n-max`: **5** for general use, **8–10** for code, **3** for creative writing.
+Tune `--spec-draft-n-max`: **3** is optimal for general use (83% acceptance rate). Values of 1–2 are more conservative; 4–5 waste compute on rejected tokens.
 
 ### Draft model (~2.3x faster)