You’re right about the coding, but not about the rest. I use this 27B for everyday tasks that aren’t actually coding. When it comes to coding, GLM5.2 is, in my view, a marvel; it’s inexpensive and works very well when used properly – but for everything else, from sorting emails to pre-generation and analysis of varying quality, it does the job that GPT and Claude can’t do for free, and sometimes not quite as well.
I use Claude for coding because I’ve got used to the way it works, which suits me. But the 27B does the opposite: the ‘local’ mode is there to help you get used to the way it works, and that might be the point.
Unconvincing, I use this model for agentic work, and while it's obviously not as good, this person is using their personal setup and experience as a justification.
1. Mac is slow at prompt processing, the Spark is much better for coding. Most of the time in coding is spent in prompt processing.
3. Harness matters, they gave no indication of what they are using. (likely an entire post just to describe a good setup)
The low, double-digit billion param models have improved vastly, notably earlier this year. I'm personally excited for the next iteration later this year because everything keeps getting better. The DFlash addons are a pure computational performance boost, there will be capability improvements later this year in this size range.
You’re right about the coding, but not about the rest. I use this 27B for everyday tasks that aren’t actually coding. When it comes to coding, GLM5.2 is, in my view, a marvel; it’s inexpensive and works very well when used properly – but for everything else, from sorting emails to pre-generation and analysis of varying quality, it does the job that GPT and Claude can’t do for free, and sometimes not quite as well. I use Claude for coding because I’ve got used to the way it works, which suits me. But the 27B does the opposite: the ‘local’ mode is there to help you get used to the way it works, and that might be the point.
Unconvincing, I use this model for agentic work, and while it's obviously not as good, this person is using their personal setup and experience as a justification.
1. Mac is slow at prompt processing, the Spark is much better for coding. Most of the time in coding is spent in prompt processing.
2. DFlash addon improves tgen speed 2-3x. (https://huggingface.co/collections/z-lab/dflash)
3. Harness matters, they gave no indication of what they are using. (likely an entire post just to describe a good setup)
The low, double-digit billion param models have improved vastly, notably earlier this year. I'm personally excited for the next iteration later this year because everything keeps getting better. The DFlash addons are a pure computational performance boost, there will be capability improvements later this year in this size range.