
Dissecting LFM3-1.7B and Grafting Its Behavior Into Gemma-3-270M: Only 640 Neurons Edited, Zero Training
Vũ Khôi
Tác giả tại Wotech Electronics Blog

Dissecting LFM3-1.7B and Grafting Its Behavior Into Gemma-3-270M: Only 640 Neurons Edited, Zero Training
I just tried dissecting a 1.7B AI, pulling its behavior chain out and grafting it straight into the brain of another AI from a different family, 6 times smaller, without any training. That is the sharing of Nam Nam in the J2Team Community group (20/11/2026) about modding LLM weights, following a trend that has been hot lately. All figures, procedures, and views in this article follow the post and have not been independently verified by Wotech.
Quick answer: According to the post, normally to teach a small AI (like Gemma-3-270M) to answer concisely like a big one (LFM3-1.7B), you have to gather lots of data, rent GPUs, then build a fine-tune/distill pipeline. This time the author tried the "unorthodox" route of poking straight into the weight file, with no training at all. Unlike jailbreak or steering tricks on the same model, he extracted the behavior chain straight from LFM3-1.7B and grafted it onto Gemma-3-270M even though the two have completely different architectures, different tokenizers, and a 6x size difference. He did not add knowledge, only peeled off the "behavior" part of the 1.7B and hand-patched it into exactly 640 neurons in the 270M's weight file (only about 0.00024% of total parameters). Real results: Gemma-3-270M answers Vietnamese smoothly and follows LFM3-1.7B's behavior form much better than the original; the fix is baked into the weight file, so no runtime code intervention is needed. Caveat: this only transfers behavior, it does not pump in knowledge — where the 270M lacks knowledge it still lacks it, and it may glitch on some strange cases.
Knowledge and behavior (logic & personality) are two fairly separate parts in AI, per the post: knowledge is how many facts it knows and how many languages it can code; behavior is whether it stays in Vietnamese, answers long or short, and what tone it uses.
Why poke straight into the weight file instead of fine-tuning?
According to the post, normally to teach a small AI (like Gemma-3-270M) to answer concisely like a big one (LFM3-1.7B), you have to gather lots of data, rent GPUs, then build a fine-tune/distill pipeline. This time the author tried the "unorthodox" route of poking straight into the weight file, with no training at all. Unlike jailbreak or steering tricks on the same model, he extracted the behavior chain straight from LFM3-1.7B and grafted it onto Gemma-3-270M even though the two have completely different architectures, different tokenizers, and a 6x size difference.
Real results and limits
According to the post, the results are:
| Item | Result, per the post |
|---|---|
| Vietnamese | Gemma-3-270M answers Vietnamese smoothly and follows LFM3-1.7B's behavior form much better than the original |
| Deployment | The fix is baked into the weight file, so no runtime code intervention is needed |
| Limit | It only transfers behavior, it does not pump in knowledge — where the 270M lacks knowledge it still lacks it, and it may glitch on some strange cases |
Apply now: the full process, before-after images, and the modded model
According to the post, the full process of extracting/grafting the 640 numbers, before-after comparison images, and the Hugging Face link to the modded model are in the post for everyone to download and try: the blog at https://neeze.github.io/docs/practical-ai/weight-programming-lfm3-to-gemma/ and the model at https://huggingface.co/Neeze/gemma-3-270m-it-vi-steer. The author thanks Nguyễn Minh Khoa for supporting this experiment. Extra toys for modding LLMs he recommends: Weight-space programming at https://rikiyat.github.io/blog/weight-space-programming/ and Neuron-explorer at https://github.com/lyramakesmusic/neuron-explorer. Shared so that people who like tinkering with local LLMs can see that besides fine-tuning, "weight programming" still has many fun tricks.
Frequently asked questions
1. What did the author do with the two AI models? According to the post, the author dissected LFM3-1.7B, pulled its behavior chain out and grafted it straight into the brain of Gemma-3-270M — an AI from a different family, 6 times smaller — without any training, just poking straight into the weight file.
2. How is this different from fine-tuning or steering? According to the post, fine-tuning/distilling needs lots of data, rented GPUs, and a pipeline; steering or jailbreak tricks only work on the same model. This method extracts the behavior chain straight from LFM3-1.7B and grafts it onto Gemma-3-270M even though the two have completely different architectures, different tokenizers, and a 6x size difference.
3. How many neurons were edited, and what was the result? According to the post, the author hand-patched exactly 640 neurons in the 270M's weight file (only about 0.00024% of total parameters); the result is that Gemma-3-270M answers Vietnamese smoothly and follows LFM3-1.7B's behavior form much better than the original, with the fix baked into the weight file so no runtime code intervention is needed.
4. Does this pump extra knowledge into the small model? According to the post, no: this only transfers behavior, it does not pump in knowledge — where the 270M lacks knowledge it still lacks it, and it may glitch on some strange cases.
5. Where can I download the modded model and read more? According to the post: the process blog at https://neeze.github.io/docs/practical-ai/weight-programming-lfm3-to-gemma/, the model at https://huggingface.co/Neeze/gemma-3-270m-it-vi-steer; LLM modding toys: Weight-space programming at https://rikiyat.github.io/blog/weight-space-programming/ and Neuron-explorer at https://github.com/lyramakesmusic/neuron-explorer.
Source: post by Nam Nam in the J2Team Community group (20/11/2026) about weight programming from LFM3-1.7B to Gemma-3-270M. All figures, procedures, and views in this article follow the post and have not been independently verified by Wotech.
Bài viết liên quan

20 tháng 11, 2026 · 6 phut doc
Living Spec: A Controlled Experiment Shows the Spec File Is 54x Cheaper — But a Stale Spec Is a Single Point of Failure
I just made my own AI prove a hot method, and the result made me revise my own understanding. That is the sharing of Cáo Nhăn Nhó in the Muse AI Việt Nam group (19/11/2026) about the Living Spec meth…

20 tháng 11, 2026 · 7 phut doc
Muse Assistant Sues Meta Itself: Locked Out as a Suspected Hacker, Then It Drafted the $75,000 Lawsuit
A small business owner in South Carolina just used Meta's own Muse assistant to sue Meta, and this story is exactly the kind only the agent era could produce. That is the content of a post by Trần Mạ…

20 tháng 11, 2026 · 7 phut doc
Muse Connects to Google Calendar: A Calendar That Reports to AI, So Muse Knows About a Rescheduled Meeting Before You Do
One afternoon, a meeting with a supplier got rescheduled, and before you even opened your calendar to check, Muse already knew. Not because Muse guessed, but because your calendar is connected to it,…