What question did this study set out to answer?

The aim is to develop a compact language model for reliable Python code synthesis on edge hardware.

March 24, 2026Open Access

MicroPy-5M: Reliable Code Synthesis via Recursive Error-Feedback (REF) in 5.6M Parameter Models for Edge Intelligence

Key Points

The aim is to develop a compact language model for reliable Python code synthesis on edge hardware.
Introduced MicroPy-5M, a decoder-only transformer with 5.6M parameters.
Implemented a Recursive Error-Feedback (REF) mechanism for error correction.
Performed self-correction of syntax faults through feedback iterations.
Evaluated performance on a curated suite of programming tasks.
Achieved 100% convergence on axiomatic programming tasks within 1-2 feedback iterations.
Model occupies an 11.2MB on-disk footprint and approximately 42MB of RAM.
Established a new standard for deploying AI on previously incompatible edge devices.

Abstract

Abstract—Deployinglanguagemodelsonedgehardwareisfundamentallyconstrainedbymemorybudgetsthatprecludebillion-parameterarchitectures.ThispaperintroducesMicroPy-5M,adecoder-onlyTransformercomprisingonly5.6Mparam-eters,designedexclusivelyforlocalisedPythonscriptsynthesisonmicrocontroller-classdevices.Ourcentralfindingisthat,atthisextremesub-10M-parameterscale,conventionalprobabilisticone-shotdecodingproducesunacceptablyhighratesofstructuralerrors;reliabilityisinsteadachievablethroughaclosed-loopRecursiveError-Feedback(REF)mechanismthattreatsthePythoncompile()bytecodecompilerasanobjective,externalverificationoracle.Within1–2feedbackiterationsthemodelself-correctssyntaxfaultsandachieves100%convergenceonacuratedsuiteofaxiomaticprogrammingtasks,whileoccupyingan11.2MBon-diskfootprint(FP16)andapproximately42MBofRAMatinferencetime.TheseresultsestablishanewoperatingpointforwhatwetermLogic-GatedMicro-Intelligence:compact,formallyverified,andsuitablefordeploymentondevicespreviouslyconsideredincompatiblewithon-deviceAI.IndexTerms—EdgeAI,TinyML,codegeneration,Transformer,formalverification,recursiveself-correction,IoT,microcontroller

MicroPy-5M: Reliable Code Synthesis via Recursive Error-Feedback (REF) in 5.6M Parameter Models for Edge Intelligence

Key Points

Abstract

Cite This Study