actually a longer context does not take up more embedding params. though it sometimes can depending on the positional embeddings used; if you are using RoPE more or less context doesnt affect param count. Out of the most popular types only learned absolute embeddings will always create more params whereas relative learned embeddings will sometimes create more. most others dont.
Hoglet (Ash) PRO
Hoglet-33
P(doom) <0.1%
AI & ML interests
Open source AI, datasets, parameter efficiency, SLMs, AI for the betterment of humanity. Contact at ash@basicallyai.co
Recent Activity
new activity about 1 hour ago
AxiomicLabs/GPT-X3-150M:WOW! new activity about 3 hours ago
basically-ai/Pebble-10M-Chat:Holy sh, a small language model that is mamab! repliedto Banaxi-Tech's post about 5 hours ago
Well GPT X3 takes the lead. As of right now.
We are now announcing BananaMind 3 🍌! (not ai for those emoji guys)
All models will use BGA (which is almost just NSA) and our BM3X architecture.
Its sizes will be:
BananaMind 3 Flash Lite, 3M parameters at a context of 8K context.
BananaMind 3 Lite, 10M parameters with 16K context.
BananaMind 3 Flash, 25M Parameters with 16K context.
BananaMind 3 Pro, 50M parameters with 24K context.
BananaMind 3 Ultra, 100M parameters with 32K context.
And lastly, BananaMind 3 Max with 150M parameters and 64K CONTEXT.
I can assure you BananaMind 3 Max WILL beat GPT X3 or match it, we won't release it otherwise. We hope for a 40+ INTELLIGENCE INDEX!
BananaMind 3 may also be partnered with dot labs.
We will cancel BananaMind 2.1 and BananaMind 2 Ultra.
As of the BETU SLM Leaderboard we may need to release it after October 11, im very busy right now (even though we said We will release BETU leaderboard before Oct 11 😟)
Organizations
replied to Banaxi-Tech's post about 5 hours ago
reacted to Banaxi-Tech's post with 👍 9 days ago
reacted to DedeProGames's post with 🔥 12 days ago
Post
2888
🧱 SLM Tetris Arena: can a small language model play Tetris without ever being trained on it?
I built an arena where tiny decoder-only LMs (50K–250M params) play Tetris zero-shot. There is no fine-tuning and no game data. They only use what they picked up from pre-training on text.
How it works:
- For every piece, the engine simulates each legal placement and describes the result in plain English ("clears one line, creates no new holes, keeps the stack low…").
- The model never sees the grid. It reads each description, and the arena compares log P(" good move") with log P(" bad move"). The best-rated placement is played.
- Every player gets the same piece sequence, so it's a fair race.
- There are two protocols: Guided (the rules are in the prompt) and Blind (no rules, only pre-training knowledge).
Two ways to play:
- Match: pick any models (even your own, custom architectures welcome) and watch them play side by side on retro 8-bit boards.
- Ranked: press Play and the arena picks up to 4 models at random from a curated pool of 29. Nobody chooses their opponents, so Elo can't be farmed. Matches run on the server and count even if you close the tab.
First results (~225 ranked matches):
- gpt2 (124M) leads with 1283 Elo, but SupraNeo-4M (4M) is right behind at 1239. Next come LowOnMind-5M and BananaMind-2.1-Pico (1.5M!).
- Model size barely predicts Elo (r ≈ 0.06). Survival does (r ≈ 0.9): the models that avoid holes and keep the stack low are the ones that win.
Every ranked match (seed, model commit SHAs, scores, Elo before/after) is logged in a public dataset.
▶ Play: DedeProGames/SLM-Tetris-Arena
📊 Results: DedeProGames/lm-tetris-arena-results
Want your model in the Ranked pool? Drop it in the comments!
I built an arena where tiny decoder-only LMs (50K–250M params) play Tetris zero-shot. There is no fine-tuning and no game data. They only use what they picked up from pre-training on text.
How it works:
- For every piece, the engine simulates each legal placement and describes the result in plain English ("clears one line, creates no new holes, keeps the stack low…").
- The model never sees the grid. It reads each description, and the arena compares log P(" good move") with log P(" bad move"). The best-rated placement is played.
- Every player gets the same piece sequence, so it's a fair race.
- There are two protocols: Guided (the rules are in the prompt) and Blind (no rules, only pre-training knowledge).
Two ways to play:
- Match: pick any models (even your own, custom architectures welcome) and watch them play side by side on retro 8-bit boards.
- Ranked: press Play and the arena picks up to 4 models at random from a curated pool of 29. Nobody chooses their opponents, so Elo can't be farmed. Matches run on the server and count even if you close the tab.
First results (~225 ranked matches):
- gpt2 (124M) leads with 1283 Elo, but SupraNeo-4M (4M) is right behind at 1239. Next come LowOnMind-5M and BananaMind-2.1-Pico (1.5M!).
- Model size barely predicts Elo (r ≈ 0.06). Survival does (r ≈ 0.9): the models that avoid holes and keep the stack low are the ones that win.
Every ranked match (seed, model commit SHAs, scores, Elo before/after) is logged in a public dataset.
▶ Play: DedeProGames/SLM-Tetris-Arena
📊 Results: DedeProGames/lm-tetris-arena-results
Want your model in the Ranked pool? Drop it in the comments!
- 1 reply
replied to DedeProGames's post 12 days ago
Why do you think that GPT-2 is in the top followed by sub 10M models?
replied to Banaxi-Tech's post 12 days ago
like instead of being locked into gpt6 luna i can use deepseek or claude, or another openai model
replied to their post 12 days ago
oh sorry, i dont understand anything sometimes
replied to Banaxi-Tech's post 13 days ago
I know you have been working a lot on this, and its great, but could you maybe add support for other models so i can choose or use an API key and model from another platform if the platform is OpenAI compatible?
replied to their post 13 days ago
the Orion in Refract Labs? I let HF have you and Refract Labs stuff show up in feed plus notifications if thats what your asking, plus pinging me sends a notification. i havnt tested it yet though because of this sidetrack, and i accidentally broke a few things in my code base so once i get all of that sorted i will def test it
replied to their post 13 days ago
yeah i know, the base model doesnt even support those tokens, but it was a dataset i could remember easily so thats what i did
reacted to their post with 🔥 13 days ago
Post
6208
Hey everyone! I got sidetracked from my main projects and decided to test out the BananaAll app and see if I could make a small model not regress too much during SFT. Here is what happened:
The base model I chose was BananaMind/BananaMind-2.1-Pico-Preview, and the dataset I used was SupraLabs/SupraThink-Dataset-500x
I trained for 5 whole steps using a LoRA adapter.
Results:
A model that scores better on some benchmarks and worse on others, and still lacks most general capabilities.
You can find the model here: Hoglet-33/Hogleto
Credits:
- Thank you to @Banaxi-Tech for the BananaAll app (works perfectly on Windows and CPU)
- GPT-6 Sol for knowing how to merge some confusing files created by the app
- Myself for the idea
- Someone else somewhere who might have contributed to some of my ideas and might in the future
- And readers like you!
The base model I chose was BananaMind/BananaMind-2.1-Pico-Preview, and the dataset I used was SupraLabs/SupraThink-Dataset-500x
I trained for 5 whole steps using a LoRA adapter.
Results:
A model that scores better on some benchmarks and worse on others, and still lacks most general capabilities.
You can find the model here: Hoglet-33/Hogleto
Credits:
- Thank you to @Banaxi-Tech for the BananaAll app (works perfectly on Windows and CPU)
- GPT-6 Sol for knowing how to merge some confusing files created by the app
- Myself for the idea
- Someone else somewhere who might have contributed to some of my ideas and might in the future
- And readers like you!
- 6 replies
posted an update 13 days ago
Post
6208
Hey everyone! I got sidetracked from my main projects and decided to test out the BananaAll app and see if I could make a small model not regress too much during SFT. Here is what happened:
The base model I chose was BananaMind/BananaMind-2.1-Pico-Preview, and the dataset I used was SupraLabs/SupraThink-Dataset-500x
I trained for 5 whole steps using a LoRA adapter.
Results:
A model that scores better on some benchmarks and worse on others, and still lacks most general capabilities.
You can find the model here: Hoglet-33/Hogleto
Credits:
- Thank you to @Banaxi-Tech for the BananaAll app (works perfectly on Windows and CPU)
- GPT-6 Sol for knowing how to merge some confusing files created by the app
- Myself for the idea
- Someone else somewhere who might have contributed to some of my ideas and might in the future
- And readers like you!
The base model I chose was BananaMind/BananaMind-2.1-Pico-Preview, and the dataset I used was SupraLabs/SupraThink-Dataset-500x
I trained for 5 whole steps using a LoRA adapter.
Results:
A model that scores better on some benchmarks and worse on others, and still lacks most general capabilities.
You can find the model here: Hoglet-33/Hogleto
Credits:
- Thank you to @Banaxi-Tech for the BananaAll app (works perfectly on Windows and CPU)
- GPT-6 Sol for knowing how to merge some confusing files created by the app
- Myself for the idea
- Someone else somewhere who might have contributed to some of my ideas and might in the future
- And readers like you!
- 6 replies
replied to Banaxi-Tech's post 13 days ago
This comment has been hidden
replied to harshitkgupta's post 13 days ago
Why not have used Qwen 3 or 3.5?
reacted to Banaxi-Tech's post with 🔥 13 days ago
Post
7615
We're releasing a MAJOR update to the BananaAll SLM Super App.
If you want to use a custom architecture, previously you had to go trough reviewing the code yourself, now add an Openrouter API key and review it with GPT 6 Luna in one button. A review cost be half a cent so anyone can try it. This is one of the main features.
Now ROCm, AMD and Windows, Mac support.
Colab and Molab support.
Detailed list of features:
Get improved Windows Python detection and support paths for compatible AMD ROCm, Intel XPU, and Apple MPS setups.
Choose local training or export a self-contained Python script for Colab or Molab. Notebook runs produce a downloadable model ZIP.
Start pretraining with an existing model’s tokenizer, or train a new one from your datasets.
Try experimental 1.58-bit Ternary fake-quantized training on NVIDIA GPUs.
Watch live tokens per second. Model compilation is on by default and falls back automatically if it fails.
Build custom architectures with separate configuration and modeling files, then review the training code manually or with optional OpenRouter AI Review.
Install from source with the new coding-agent instructions.
This release also fixes inflated loss reporting for custom models.
And for those users who didn't want to try it out just because installation would be so hard, it isnt now.
Go to any coding agent (Pi, Claude Code, Codex, OpenCode, basically all work), and just paste "Install BananaAll for me. Fetch and follow https://raw.githubusercontent.com/BananaMind/BananaAll/main/agent_install.txt."
That's it.
Check it out at https://github.com/BananaMind/BananaAll/
Also on SAICR, we're currently training a new major model (NACR v2) and ACR 1.0 is in the finishing.
If you want to use a custom architecture, previously you had to go trough reviewing the code yourself, now add an Openrouter API key and review it with GPT 6 Luna in one button. A review cost be half a cent so anyone can try it. This is one of the main features.
Now ROCm, AMD and Windows, Mac support.
Colab and Molab support.
Detailed list of features:
Get improved Windows Python detection and support paths for compatible AMD ROCm, Intel XPU, and Apple MPS setups.
Choose local training or export a self-contained Python script for Colab or Molab. Notebook runs produce a downloadable model ZIP.
Start pretraining with an existing model’s tokenizer, or train a new one from your datasets.
Try experimental 1.58-bit Ternary fake-quantized training on NVIDIA GPUs.
Watch live tokens per second. Model compilation is on by default and falls back automatically if it fails.
Build custom architectures with separate configuration and modeling files, then review the training code manually or with optional OpenRouter AI Review.
Install from source with the new coding-agent instructions.
This release also fixes inflated loss reporting for custom models.
And for those users who didn't want to try it out just because installation would be so hard, it isnt now.
Go to any coding agent (Pi, Claude Code, Codex, OpenCode, basically all work), and just paste "Install BananaAll for me. Fetch and follow https://raw.githubusercontent.com/BananaMind/BananaAll/main/agent_install.txt."
That's it.
Check it out at https://github.com/BananaMind/BananaAll/
Also on SAICR, we're currently training a new major model (NACR v2) and ACR 1.0 is in the finishing.
- 11 replies
replied to Banaxi-Tech's post 13 days ago
so basically the update for today is:
Rocm, Apple, Intel, AI Support for writing scripts via GPT 6 Luna, better UI, Molab and Colab support, Ternary model training, Windows Support
YES!
replied to their post 14 days ago
not much actually, most things stay the same, some things are a little weird though
replied to Banaxi-Tech's post 15 days ago
replied to Banaxi-Tech's post 15 days ago
also i rent vast.ai GPUs for training, could i set it up in jupyter easily?
replied to Banaxi-Tech's post 15 days ago
Even my RNN/SSM based models?
reacted to Enderchef's post with 🔥 16 days ago
Post
3044
AxiomicLabs released new benchmark, Tiny Theory of Mind, to test your SLM models' Theory of Mind Intuition!
Check it out and like it!
AxiomicLabs/Tiny_Theory_of_Mind
Check it out and like it!
AxiomicLabs/Tiny_Theory_of_Mind