Source context: This short field note first appeared on Eric's Advisory Hour Substack. View the original note.
The field note
Watching the new qwen3.5:0.8b model one shot out an entire plan for a c# app from my dusty old 3X series nvidia hardware is nothing short of spectacle-the generation took under a second.
Why it matters
The useful signal is not just that a small model answered. It is that a sub-billion-parameter model could sketch a complete application plan locally, quickly, and cheaply enough to change how often the loop can be run.
When planning latency drops under the threshold where it feels like waiting, the role of the model shifts. It becomes less like a scarce cloud call and more like a disposable planning primitive: ask, compare, throw away, retry, and let the harness decide what is worth keeping.
What this note points toward
- Local first passes: small models can draft project structure, task lists, and implementation routes before a larger model enters the loop.
- Cheap breadth: under-a-second planning makes it practical to sample several approaches instead of trusting the first plan.
- Harness pressure: the next constraint becomes orchestration, scoring, memory, and evals rather than raw model access.
Connection to the SLM series
This sits in the same notebook as the NPU, Gemma, nanochat, and pico-model experiments: the emerging question is not whether a small model can imitate a bigger one in every domain. It is where a narrow, fast, locally available model is already sufficient for a useful slice of work.
For app planning, that useful slice may be early decomposition: create the outline, name the parts, expose assumptions, and hand a cleaner brief to the rest of the system.
Related notes
For the local-hardware thread, see Tinkering with a NPU, NPU and Gemma4, and Local Inference Puzzle. For the broader small-model argument, start with Why Tiny Specialists Matter.
Original source
This local copy preserves the note text and source link from Eric's Advisory Hour. Original Substack URL: https://substack.com/@advisoryhour/note/c-222349081. You can also subscribe to Advisory Hour.