Post History

Current version by Nick Antonaccio

Current VersionAug 17, 2026 at 13:39

Here are some output examples, by a variety of the Bonsai models explained above. Notice how fast some of the tiny models can generate tokens, and even the 4B model can produce a significant percentage of actually valid code.

A human, or a large model would really need to work the code from the smaller models, to make it usable, but even the smaller model output could potentially help save time when working on short coding tasks. Pay particular attention to how quickly the small models performed on a tiny netbook with no GPU:

That 4th generation, by the big Qwen 3.6 27b model (which is one of the most capable open source LLMs available) was completed on an extremely low power netbook with an N5030 CPU!

I expect that the most viable way to use these models, to generate actually usable production code, would likely be to use smaller models to generate the bulk of code, and then have progressively larger models fix/refine the output of each successively better model. Notice that input tokens are processed something in the order of 2x-3x faster than output tokens, so using small models to provide some basic code structure, and then molding it into better working code, seems like a reasonable approach to building large code bases quickly. Build some basic code outlines, and then work out the details - that should work more quickly than using a large model to generate everything at a snail's pace.

Or you could always just run a long process overnight on Qwen 3.6 27b - the output from the super compressed Bonsai versions of that model is very good, for the size and hardware requirements needed to run it.

Here's a session in which Gemini 3.7 Flash evaluated all the code examples above:

https://com-pute.com/nick/gemini37-evaluating-flask-code-by-bonsai-models--pi-session-2026-08-17T00-25-55-986Z_01a00d1c-5612-759f-9da8-96c79895faa4.html

Previous Versions
Version 5Aug 17, 2026 at 13:39

Here are some output examples, by a variety of the Bonsai models explained above. Notice how fast some of the tiny models can generate tokens, and even the 4B model can produce some actually valid code.

A human would really need to work this code to make it usable, but the model could potentially help save some time when working on short coding tasks. It's very quick on a tiny netbook with no GPU:

That 4th generation, by the big Qwen 3.6 27b model (which is one of the most capable open source LLMs available) was completed on an extremely low power netbook with an N5030 CPU!

I expect that the most viable way to use these models, to generate actually usable production code, would likely be to use smaller models to generate a bulk of code, and then have progressively larger models fix/refine the output of each successively better previous model. Notice that input tokens are processed something in the order of 2x-3x faster than output tokens, so using small models to provide some basic code structure, and then molding it into better working code, seems like it may be a reasonable way to build large code bases quickly, and then work out the details - more quickly than using a large model to generate everthing at a snail's pace.

Or you could always just run a long process overnight on Qwen 3.6 27b - the output from this super compressed version is still very good.

Here's a session in which Gemini 3.7 evaluates all the code examples above:

https://com-pute.com/nick/gemini37-evaluating-flask-code-by-bonsai-models--pi-session-2026-08-17T00-25-55-986Z_01a00d1c-5612-759f-9da8-96c79895faa4.html

Version 4Aug 17, 2026 at 13:34

Here are some output examples by the smallest Bonsai models. Notice how fast some of the tiny models can generate tokens, and even the 4B model can produce some valid code. A human would really need to work this code to make it usable, but the model could potentially help save some time when working on short coding tasks. It's very quick on a tiny netbook with no GPU:

That 4th generation, by the big Qwen 3.6 27b model (which is one of the most capable open source LLMs available) was completed on an extremely low power netbook with an N5030 CPU!

Here's a session in which Gemini 3.7 evaluates all the code output above:

https://com-pute.com/nick/gemini37-evaluating-flask-code-by-bonsai-models--pi-session-2026-08-17T00-25-55-986Z_01a00d1c-5612-759f-9da8-96c79895faa4.html

Version 3Aug 17, 2026 at 12:40

Here are some output examples by the smallest Bonsai models. Notice how fast some of the tiny models can generate tokens, and even the 4B model can produce some valid code. A human would really need to work this code to make it usable, but the model could potentially help save some time when working on short coding tasks. It's very quick on a tiny netbook with no GPU:

That last generation, by the big Qwen 3.6 27b model (which is one of the most capable open source LLMs available) was completed on an extremely low power netbook with an N5030 CPU!

Here's a session in which Gemini 3.7 evaluates all the code output above:

https://com-pute.com/nick/gemini37-evaluating-flask-code-by-bonsai-models--pi-session-2026-08-17T00-25-55-986Z_01a00d1c-5612-759f-9da8-96c79895faa4.html

Version 2Aug 17, 2026 at 12:40

Here are some output examples by the smallest Bonsai models. Notice how fast some of the tiny models can generate tokens, and even the 4B model can produce some valid code. A human would really need to work this code to make it usable, but the model could potentially help save some time when working on short coding tasks. It's very quick on a tiny netbook with no GPU:

https://com-pute.com/nick/bonsai1.7b-output.txt https://com-pute.com/nick/bonsai4b-output.txt https://com-pute.com/nick/bonsai8b-output.txt https://com-pute.com/nick/bonsai27b-output.txt

Here's a session in which Gemini 3.7 evaluates all the code output above:

https://com-pute.com/nick/gemini37-evaluating-flask-code-by-bonsai-models--pi-session-2026-08-17T00-25-55-986Z_01a00d1c-5612-759f-9da8-96c79895faa4.html

Version 1Aug 17, 2026 at 12:33

Here are some output examples by the smallest Bonsai models. Notice how fast some of the tiny models can generate tokens, and even the 4B model can produce some valid code (a human would really need to work this code to make it usable - but the model would certainly save some work), and it's very quick on a tiny netbook with no GPU:

https://com-pute.com/nick/bonsai1.7b-output.txt https://com-pute.com/nick/bonsai4b-output.txt https://com-pute.com/nick/bonsai8b-output.txt

Here's a session in which Gemini 3.7 evaluates all the code output above:

https://com-pute.com/nick/gemini37-evaluating-flask-code-by-bonsai-models--pi-session-2026-08-17T00-25-55-986Z_01a00d1c-5612-759f-9da8-96c79895faa4.html

Back to Thread