This is the seventh episode in my series of tests of a small model on a 16GB MacBook Pro. As usual the recording is raw — not sped up and not edited — so the timing and behavior are what you actually get.
Today we’re testing Spark-X2.6–4B (also written Spark X2.6 4B) from SparkLLM / XHToken. It is roughly 4.11 billion parameters, released under Apache 2.0. The architecture is the same as the previous Spark-X2.5–4B; the gains should come from improved post-training. The model card highlights better coding, agentic, and long-context performance and claims results that match or exceed some ~9B models on selected benchmarks (those are the team’s own numbers). It supports a 1 million token context window and three thinking levels (none, low, high). English and Chinese are the main languages. In this session I kept the context at 32k because of the memory limit on the MacBook.
I used the same two checks as always, running through MLX Serve.
First, the short question “What is a dolphin?” The answer was pretty good and complete. For a 4B model it ranks near the top of everything I’ve tested so far on this machine — top two or three.
Second, the coding task: generate a small utility that counts words in files inside a folder. The model asked for tools — first to write the file, then later to run or test it. It produced a bash script (Wordcount.sh) that works and is portable. Generation didn’t take especially long and the overall behavior was solid.
Taken together, Spark-X2.6–4B is a strong small model for text generation and light coding. It sits at roughly the same practical level as the Google Gemma and Nemotron Dory tests I did recently. Throughput is decent and you still have a bit of room left on a 16GB machine if you want to keep other small apps open.
If you have other models that fit comfortably in 16GB and you’d like them tested the same way, leave a comment. I read them. And if you like these tests, like and share and subscribe to my YouTube channel!