MOEs:

Strata: Qwen 3.8 Flash Next: solid. First time I’ve felt I’m actually working with Sonnet 5 / 4.8 intelligence.

Maya: https://github.com/mw00/project-maya

Lemonade for the rest (I use Zest):

Pick between: GLM 4.7 Flash: confident little bullshitter. Runs faster than Flash-Next, but I do not trust this thing to code. It hallucinates so much. You can make it good with skills, but I would only pick it if you can’t run Qwen 3.6.

Qwen 3.6-35B-MTP: I really like it. It’s very fast and competent at basic tasks - I wouldn’t let it handle a codebase, because it does tend to mangle even small projects, but it seems to do okay at researching stuff and tweaking system configs and scripts, does agentic workflows better than GLM. Great coding assistant for in-terminal use.

Give it a few skills - especially obra/superpowers (https://github.com/obra/superpowers) and it becomes a very good daily driver.

Dense: Gemma4 31B-MTP Qwen 3.8 27B

Small model:

LFM2-24B-A2B: massive respect for this model. Runs on my GTX 1080. However, tool calls failed all over the place, it threw itself into thinking looops . . . can’t recommend it.

LFM2.5-8B-A1B: wonderful model. Blazing fast, runs on damn near anything, tool calls well, and above all, its answers are short and to the point. Probably the smartest small model I’ve used when paired with a harness like OpenChamber: give it web search and it’s off to the races. I honestly think of it as a drop-in replacement for whatever the Gemini Flash model is.


Back to top

To get in touch, reach out to hello [at] yudhanjaya [dot] com, or contact finegan [at] zenoagency [dot] com to talk to my agent (Stevie Finegan). Responses from my end may take time.

/ search
rss · substack