EXP 001
NODE 00
Qwen3 14B
M1 Pro · 16 GB unified memory
Testing how far a 14B-class local model can be pushed
on a constrained Apple Silicon machine.
RESULT
Runs successfully, but pushes Node 00 close to its
practical memory ceiling. Memory pressure and swap
leave little headroom for larger contexts or
additional tools.
EXP 002
NODE 00
Ministral 14B · Agent Loop
Tool use · Python execution · self-correction
A local model generates Python, executes a task,
evaluates the result and corrects its own work.
RESULT
Successfully completed an autonomous tool-use loop:
generating Python, executing it, evaluating the output
and correcting its own work before producing the
final result.
EXP 003
NODE 00
Bonsai 27B
M1 Pro · 16 GB unified memory
Attempting to run a substantially larger model
at the target context size on Node 00.
RESULT
Unable to run at the target context size within
16 GB of unified memory. Node 00 reached a hard
hardware constraint rather than a model-quality
limitation.
EXP 004
CLOUD COMPUTE
Qwen3.5 35B-A3B
NVIDIA A100 · 80 GB VRAM
Comparing agentic research behavior against a
substantially larger dense model.
OBSERVATION
Produced stronger agentic research behavior than
Qwen2.5 72B in this specific workflow, despite the
larger model's parameter count. Model size alone
did not predict agent quality in this test.