llama.cpp VS koboldcpp

> The benchmark prompt was:

> Write a compact Python function that parses a unified diff and returns the changed file paths. Then explain two edge cases.

> Each benchmark generated about 128 tokens.

Generating 128 tokens is probably not enough for good benchmark results. MTP speedup depends on how often the predicted tokens are accepted. In my experience, the very early output has a higher acceptance rate, so short testing can give false positive speedups.

Also llama.cpp includes a tool specifically for benchmarking:

https://github.com/ggml-org/llama.cpp/blob/master/tools/llam...

In my build, MTP came from mainline llama.cpp, not ik_llama. ik_llama got me to ~47 (engine + quant), but I couldn't get MTP running there — my build rejected the -mtp flags and ignored the model's nextn tensors. Mainline llama.cpp added MTP fairly recently (PR #22673, merged 2026-05-16), and that's where it worked for me. (There may well be an ik_llama path I missed — this is just what got it going on my box.)

That script grew up. Today I'm releasing LlamaStash, the first public release of a fast, cross-platform, terminal-native launcher for llama.cpp with zero overhead.

LlamaStash spawns the unmodified upstream llama-server. So three different questions follow from that, and there is a benchmark suite for each.

llama.cpp includes a benchmarking tool called llama-bench https://github.com/ggml-org/llama.cpp/blob/master/tools/llam...

ik_llama includes llama-sweep-bench https://github.com/ikawrakow/ik_llama.cpp/blob/main/examples...

When comparing hardware, the output of these tools is very helpful to let others put it into context. The post says the output is "reading speed" but knowing the prefill and token generation speeds would be a lot more helpful.

lol the same way we implement all of the reduced precision fp8, fp4 types today: by storing them in the corresponding uint:

https://github.com/ggml-org/llama.cpp/discussions/15095

# Build llama.cpp with Metal backend git clone https://github.com/ggml-org/llama.cpp cd llama.cpp && cmake -B build -DGGML_METAL=ON && cmake --build build -j # Community-quantized GGUFs (Google ships safetensors; unsloth ships GGUF) huggingface-cli download unsloth/gemma-4-31B-it-GGUF \ gemma-4-31B-it-Q4_K_M.gguf --local-dir . huggingface-cli download unsloth/gemma-4-26B-A4B-it-GGUF \ gemma-4-26B-A4B-it-Q4_K_M.gguf --local-dir . # Benchmark: 200 generations of 512 tokens, log per-call timing ./build/bin/llama-bench -m gemma-4-31B-it-Q4_K_M.gguf -n 512 -r 200 -o json > dense.json ./build/bin/llama-bench -m gemma-4-26B-A4B-it-Q4_K_M.gguf -n 512 -r 200 -o json > moe.json

https://github.com/LostRuins/koboldcpp Download models at HuggingFace and run them locally. No logins, no spying, no hidden data harvesting.

Oh, that horrible Electron UI. Under Windows it pegs a core on my CPU at all times!

If you're just working as a single user via the OpenAI protocol, you might want to consider koboldcpp. It bundles a GUI launcher, then starts in text-only mode. You can also tell it to just run a saved configuration, bypassing the GUI; I've successfully run it as a system service on Windows using nssm.

https://github.com/LostRuins/koboldcpp/releases

Though there are a lot of roleplay-centric gimmicks in its feature set, its context-shifting feature is singular. It caches the intermediate state used by your last query, extending it to build the next one. As a result you save on generation time with large contexts, and also any conversation that has been pushed out of the context window still indirectly influences the current exchange.

KoboldCpp is a popular text generation software for GGML and GGUF models. It also comes with an OpenAI-compatible API endpoint when serving a model, which makes it easy to use with LibreChat and other software that can connect to OpenAI-compatible endpoints.

FYI this is the easiest way to host on the horde: https://github.com/LostRuins/koboldcpp