llama.cpp – Local LLM Inference

llama.cpp enables fast, local LLM inference with minimal setup. It can be used directly via CLI, or through frontends like LM Studio or Ollama — both of which use llama.cpp as their backend. Running llama.cpp directly gives you full control over parameters and avoids GUI overhead.

Download

Download the latest prebuilt binaries from the releases page:

Platform	Package
Windows + NVIDIA GPU	llama-*-bin-win-cuda-cu12.x-x64.zip
Windows + AMD GPU	llama-*-bin-win-vulkan-x64.zip
Linux + AMD GPU (ROCm)	llama-*-bin-ubuntu-rocm-x64.zip
CPU only	llama-*-bin-win-cpu-x64.zip

AMD on Windows: Use Vulkan — ROCm on Windows is unstable. AMD on Linux: Use ROCm for best performance.

Commands

The examples below are based on running Qwen3.6-27B-Uncensored-HauhauCS-Balanced-IQ4_XS on an AMD Radeon RX 7900 XT (20 GB VRAM) with the Vulkan backend on Windows. Adjust -ngl, -c, and model path for your hardware.

Placeholders used:

• MODEL = full path to your .gguf file
• PORT = desired server port (e.g. 19999)

---

Start Server (OpenAI-compatible API)

Consider using https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates

Qwen3.6-27B sample

(ngl -1 we load all layers info VRAM)

llama-server.exe `
  -m MODEL `
    -ngl -1 `
  -c 50000 `
  -fa on `
  --cache-type-k q8_0 `
  --cache-type-v q8_0 `
  --kv-unified `
  --parallel 2 `
  --temp 0.9 `
  --top-p 0.95 `
  --top-k 20 `
  --host 0.0.0.0 `
  --batch-size 256 `
  --cache-prompt `
  --ctx-checkpoints 64 `
  --chat-template-kwargs '{"preserve_thinking": true}' `
  --cache-reuse 256 `
  --jinja `
  --port PORT

Qwen 3.6 with MTP

Here we have now much more settings most important we use:

• np parallel slots for e.g. search agents, compact agent etc.
• cache-reuse force a reuse of the cache
• cache-ram here we added Level-2 RAM of 16GB
• ctx-checkpoints more ceckpoings for better cache reuse
• c / np = context for each slot -- cache-reuse and np doesn't work with kv_unified - but kv_unified saves a lot of memory

llama-server.exe `
  -m "..\..\..\.lmstudio\models\unsloth\Qwen3.6-27B-MTP-GGUF\Qwen3.6-27B-UD-Q5_K_XL.gguf" `
  --alias "Qwen3.6-27B" `
  -ngl 999 `
  --prio 3 `
  -ts 1,1 `
  -sm layer `
  -c 150000 `
  -fa on `
  -ctk q8_0 `
  -ctv q8_0 `
  --kv-unified `
  -b 2048 `
  -ub 1024 `
  -np -1 `
  --threads-http 2 `
  --cache-reuse 256 `
  --cache-ram 16000 `
  --ctx-checkpoints 32 `
  --temp 0.6 `
  --top-p 0.95 `
  --top-k 20 `
  --min-p 0.0 `
  --reasoning-preserve `
  --chat-template-kwargs '{\"preserve_thinking\": true}' `
  --jinja `
  --chat-template-file qwen-fixed.jinja `
  --host 0.0.0.0 `
  --port 1234 `
  --spec-type draft-mtp `
  --spec-draft-n-max 4 `
  --spec-draft-p-min 0.8

Qwopus3.6-27B

IT IS CURRENTLY NOT STABLE!!!!

llama-server.exe `
  -m "..\..\..\.lmstudio\models\Jackrong\Qwopus3.6-27B-v2-MTP-GGUF\Qwopus3.6-27B-v2-MTP-Q6_K.gguf" `
  --alias "Qwen3.6-27B" `
  -ngl 999 `
  --prio 3 `
  -ts 1,1 `
  -sm layer `
  -c 150000 `
  -fa on `
  -ctk q8_0 `
  -ctv q8_0 `
  --kv-unified `
  -b 2048 `
  -ub 1024 `
  -np -1 `
  --cache-reuse 256 `
  --cache-ram 16000 `
  --ctx-checkpoints 32 `
  --temp 0.6 `
  --top-p 0.95 `
  --top-k 20 `
  --min-p 0.0 `
  --reasoning-preserve `
  --chat-template-kwargs '{\"preserve_thinking\": true}' `
  --jinja `
  --chat-template-file qwen-fixed.jinja ` 
  --host 0.0.0.0 `
  --port 1234 `
  --spec-type draft-mtp `
  --spec-draft-n-max 4 `
  --spec-draft-p-min 0.8

Ornith

we don't load the mmproj-F32.gguf which is used for vision to save some memory using an coding agent:

• ts parameter for multi gpu - remove if you system has only one GPU

llama-server.exe `
  -m "..\..\..\.lmstudio\models\unsloth\Ornith-1.0-35B-GGUF\Ornith-1.0-35B-UD-Q4_K_XL.gguf" `
  --alias "Ornith-1.0-35B" `
  -ngl 999 `
  --prio 3 `
  -ts 1,1 `
  -sm layer `
  -c 160000 `
  -fa on `
  -ctk q8_0 `
  -ctv q8_0 `
  --kv-unified `
  -b 2048 `
  -ub 1024 `
  -np 1 `
  --top-p 0.95 `
  --top-k 20 `
  --host 0.0.0.0 `
  --port 1234 `
  --cache-prompt `
  --ctx-checkpoints 64 `
  --reasoning-preserve `
  --chat-template-kwargs '{\"preserve_thinking\": true}' `
  --cache-reuse 256 `
  --jinja `
  --chat-template-file "qwen-fixed.jinja"

C:\Users\pauls.lmstudio\models\unsloth\Ornith-1.0-35B-GGUF

The server exposes an OpenAI-compatible API at http://localhost:PORT/v1. The built-in chat UI is available at http://localhost:PORT.

Not used in Qwen

  --context-shift `
  --swa-full `

On the RX 7900 XT with IQ4_XS and -c 95000, the model uses ~16.8 GB VRAM leaving ~2.7 GB free headroom.

---

Memory Check (VRAM usage before server start)

Loads the model, prints memory breakdown, and exits immediately: (ngl to play with the layers -- usually we should offload all.)

llama-cli.exe `
  -m "..\..\..\.lmstudio\models\unsloth\Ornith-1.0-35B-GGUF\Ornith-1.0-35B-UD-Q4_K_XL.gguf" `
  -ngl 64 `
  -c 150000 `
  -fa on `
  --cache-type-k q8_0 `
  --cache-type-v q8_0 `
  -n 10 `
  -p "test" `
  -no-cnv `
  --verbose `
  -v 2>&1 | findstr /i "_memory_ load_tensors done_getting_tensors llama_model_loader common"

Look for the line:

common_params_fit_impl: projected to use XXXXX MiB of device memory vs. XXXXX MiB of free device memory

Run this before changing -c to verify the new context size still fits in VRAM.

---

Benchmark (PP + TG at various context depths)

• https://github.com/ggml-org/llama.cpp/blob/master/tools/llama-bench/README.md
• ts parameter for multi gpu

llama-bench.exe `
  -m "..\..\..\.lmstudio\models\unsloth\Ornith-1.0-35B-GGUF\Ornith-1.0-35B-UD-Q4_K_XL.gguf" `
  -ngl 99 `
  -ts 1,1 `
  -fa 1 `
  -ctk q8_0 `
  -ctv q8_0 `
  -t 6 `
  -b 512 `
  -ub 512 `
  -p 1024 `
  -n 50 `
  -d 0,16000,32000,48000,65000 `
  -r 3 `
  --progress `
  -o md

Outputs a Markdown table with PP (prompt processing) and TG (token generation) speeds in tokens/s at each context depth.

Example results — Qwen3.6 27B IQ4_XS on RX 7900 XT (Vulkan):

Context depth	PP (t/s)	TG (t/s)
d=0	611	29.9
d=16k	481	30.2
d=32k	386	27.6
d=48k	325	26.7
d=65k	289	26.5

---

Key Parameters

Parameter	Description
-ngl 64	Offload all layers to GPU (set lower if VRAM is insufficient)
-c 95000	Context size in tokens
-fa on	Enable Flash Attention (required for KV cache quantization)
--cache-type-k/v q8_0	Quantize KV cache to Q8 — reduces VRAM significantly
--parallel 1	Single-user mode — maximizes VRAM for one context slot
--temp / --top-p / --top-k	Sampling defaults (can be overridden per request via API)

Frontends

LM Studio and Ollama are convenient alternatives if you prefer a GUI or want automatic model management — they both run llama.cpp under the hood. Direct llama.cpp gives you the latest builds and full parameter control without any overhead.

Option	Setup	Control	Overhead
llama.cpp direct	Manual	Full	None
LM Studio	GUI	Limited	Medium
Ollama	CLI/GUI	Medium	Low

See also

• https://github.com/eugr/llama-benchy