# KFlash: Benchmarks of newly released versions of Apple Silicon optimized engines: 1.4.0 Splash (DFlash2 ~ 50 tok/s) vs 0.2 Kapitsa (KFlash ~65 tok/s), MLX(MTP) & more on Qwen 3.8 27b/4b/35b 4Q on Macbook Pro M4 Pro 48 GB 2024, 48GB & Macbook Pro M1 Pro 16GB 2021 — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/b26fc2effe2a2494

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-21 04:47 ingested 2026-10-22 02:43

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xbdxg3/kflash_benchmarks_of_newly_released_versions_of/)

## retained text



```
hey folks, its Ye, I work on an open source project called [Kapitsa](https://github.com/nace-ai/kapitsa): its a fast Apple Silicon optimized inference engine built for Qwen and Nemotron models. this week two new releases landed in local inference: Inco shipped Splash 1.4.0 yesterday, and we shipped Kapitsa 0.2 today with a new speculative decoding method we call KFlash. so we put the newest versions next to the other engines people here run, on a couple of macs people actuall yuse not some 128gb M5 Ultra xd, same opencode like harness. 

[https://github.com/nace-ai/kapitsa](https://github.com/nace-ai/kapitsa)   
  
**whats KFlash:**

most fast engines now use a small draft model that guesses the next \~7 tokens, and the big model checks them in one pass. DFlash2 as published checks one chain of guesses and stops at the first wrong one. KFlash checks a best-first *tree* of the drafter's candidates in that same pass, so when the drafter hesitates its runnerup guesses get checked. more details here [https://github.com/nace-ai/kapitsa/blob/main/docs/kflash.md](https://github.com/nace-ai/kapitsa/blob/main/docs/kflash.md)

**the numbers (i** dont trust benchmarks myself, neither should you, run `kapitsa bench` on your own Mac, it will allow you to submit your bench to our repo so we all could see the performance of kapitsa from across different machines.

decode tok/s of a 512-token coding answer, T 0.6, median of 5 runs, 4-bit weights, each engine in its fastest setup:

MacBook Pro M4 Pro (2024), 48GB:

|engine|version|Qwen3.8 27B|Qwen3.6 35B-A3B|Qwen3.5 4B|
|:-|:-|:-|:-|:-|
|**Kapitsa**|**0.2**|**63 tok/s** KFlash|**161** **tok/s** KFlash|**201** **tok/s** KFlash|
|Splash|1.4.0|53 DFlash2|135 DFlash2|✗|
|mlx-serve|26.10.2|54 DFlash2|124 MTP|137 MTP|
|oMLX|0.7.1rc1|45 DFlash2|127 DFlash2|102 MTP|
|Ollama|0.40.5 / 0.40.6|32 MTP|89 MTP|109 MTP|
|MTPLX|2.12.2|36 MTP|80 MTP|104 MTP|
|MLX (mlx-vlm 0.7.8)||28 MTP|86 DFlash2|88 MTP|
|llama.cpp|b11669|18 MTP|69 MTP|70 MTP|
|LM Studio|MLX runtime 1.13.1|15 no spec|86 no spec|81 no spec|

MacBook Pro M1 Pro (2021), 16GB: **Qwen3.5 4B** (the 27B is \~15GB of weights and the 35B \~19GB, they don't fit 16GB)

|engine|version|Qwen3.5 4B|
|:-|:-|:-|
|**Kapitsa**|**0.2**|**95** KFlash|
|mlx-serve|26.10.2|62 MTP|
|oMLX|0.7.1rc1|55 MTP|
|Ollama|0.40.6|55 MTP|
|LM Studio|MLX runtime 1.13.1|47 no spec|
|MLX (mlx-lm)|0.32.2|46 no spec|
|MTPLX|2.12.2|42 MTP|
|llama.cpp|b11669|38 MTP|
|Splash|1.4.0|✗ (needs an M3 or newer)|

  
**quickstart.** 

Kapitsa serves OpenAI Anthropic compatable API along with a really simple CLI that allows it to run Kapitsa w Opencode/Pi/Claude etc.

    brew install nace-ai/tap/kapitsa
    # or: curl -fsSL https://raw.githubusercontent.com/nace-ai/kapitsa/main/install.sh | bash
    
    kapitsa claude --model 27b   # or kapitsa claude, kapitsa codex, etc

`kapitsa bench` runs this same benchmark on your Mac, and `kapitsa bench --submit` adds it to the community results (shows the JSON and asks first). M2/M3 Macs and 16-36GB ones are the results we want most

**things to mention**

* Splash runs the 27B and 35B (and Flash-Next GGUFs) and needs an M3 or newer, Kapitsa works on all M chips M1 to M6
* vibecode, expect bugs. issues/PRs welcome ^(\^\^)

**credits**

DeepSeek (DSpark), Inco(DFlash2) (a really solid engine, go compare it), z-lab (DFlash), NVIDIA (LiLiCorr and the Nemotron checkpoints), Qwen (MTP heads), MLX, mlx-serve, oMLX, MTPLX, Ollama, llama.cpp and LM Studio folks. Fable 5.1, Opus 5.5 and [Nace.AI](http://Nace.AI) coding models

**asks**

try it, bench it, break it, upload your benches please ^(\^\^) 
```



document fetched 2026-10-22 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1xbdxg3/kflash_benchmarks_of_newly_released_versions_of
