We're delaying BananaMind 2.1! When BananaMind 2.1 Lite was almost done, we benchmarked it and the results we're worse than BananaMind 2 Mini.
We're going to spend alot more time in research on tiny models and then scaling up our techniques to the actual BananaMind 2.1 models!
We're also announcing these new models: BananaMind 2.1 Coder: A 149M instruction tuned coder model trained on 75B tokens + 10B tokens of stack-v3-train. BananaMind 2.1 Pico: A 1M parameter model trained on 22B tokens of data. We also may release BananaMind 2.1 Large with around 100M parameters depending on how much compute we have.
We have updated the BananaMind Base Bench leaderboard! We now have these benchmark cards, they make it way easier to see which models are actually good! We've also added the model advisor. It asks you what you want to use the model for and the parameter range and gives you the best model for your task!
Weโre excited to release BananaMind OS 2.0, a major update to our portable operating system for running AI models locally.
BananaMind OS runs directly from an ISO without Linux, installation, a cloud connection, or modifying your disks.
The new Version 2.0 adds a graphical interface with mouse support, a model library, multi-turn chat, configurable KV cache, maximum tokens and temperature controls, automatic x87/SSE/SSE2 CPU detection, BIOS support, native UEFI support and a dedicated 486 compatibility mode.
Model weights are not loaded during GRUB or startup. Only the model catalog is read, and the selected model is loaded after BananaMind OS starts.
We now have a new .litemodel format supports multiple architectures, including BananaMind models, SmolLM, SmolLM2, GPT-X2.5, min-spark 1.1 and Rose-Mini. Check it out at: https://github.com/BananaMind/BananaMindOS
You can build an ISO yourself using the .sh or .bat script. Select the models and quantizations you want, and the builder will download, quantize and package them locally. Prebuilt 10MB, 25MB, 100MB and 250MB model presets are available here: https://github.com/BananaMind/BananaMindOS/releases/tag/v2.0.0 Use the regular preset ISOs for BIOS and GRUB, including the 486 compatibility mode.
Warning: Im currently uploading the ISOs, all up to 100MB is present, 250MB is getting uploaded Use the files ending in -uefi.iso for the native x86-64 UEFI graphical frontend and improved firmware mouse support.
The preset name describes the RAM class of the individual included models. Models remain on the ISO until selected, so including multiple models does not load all of them into RAM. 486DX with an x87 FPU, Pentium and newer x86 processors are supported. Modern x86-64 computers are supported through UEFI. It currently doesent support processors without an FPU.
Comment if you want me to run it on 0.04MHz (pls dont)
Video Credit: Song: Matzan - Redesigned Music provided by NoCopyrightSounds
We're releasing Overfitter 1.0. Its a completly useless overfitted model trained on 202 epochs of SWE Bench Verified, SWE Bench Pro, Terminal Bench 2.1, DeepSWE. It gets 100% on SWE Bench Verified, 98.6% on SWE bench Pro, 100% on Terminal Bench 2.1 and 100% on DeepSWE!
We're announcing our BananaMind 2.1 model series! The models will include: - BananaMind 2.1 Nano: 10M parameters with 60B tokens. - BananaMind 2.1 Lite: 25M parameters with 40B tokens. - BananaMind 2.1 Flash: 50M parameters with 55B tokens. - BananaMind 2.1 Pro: 135M-145M parameters (still deciding) with 100B tokens. These model will use a multi tower architecture (like BananaMind/BananaMind-2.1-Unified) with some more architectural changes.
BananaMind 2.1 Pro will probrably use 2 no output towers, instead of one!
We're currently training some experimental models based on this architecture to see its scaling!
We're releasing BananaMind 2.1 Unified, a 35M three-tower relay model where the two output towers can only talk to each other through a silent middle tower that has no output head and no loss term.
Tower B trains entirely on indirect gradient. It was never told what to predict. It woke up anyway. Jacobian lens shows it carrying the correct answer ("Paris", "oxygen", "blue") at its deepest layer. Its bridge gates grew 4-47x from init. Feed it from only one side and the representations collapse to junk โ it needs both outer towers to become semantic.
The PIQA result is the cleanest demonstration: Tower A alone scores 50.11 (chance). Tower C alone 52.12. Full system 61.75. All physical reasoning lives in the integration. The 35M three-tower beats the 50M single-tower BananaMind 2 Medium on PIQA.
Trained on 38B tokens in ~7.5 hours on 8x RTX PRO 6000. Ships with 7 ablation modes so you can surgically cut the model apart without retraining. Full training logs, J-lens fits, and eval outputs for every mode included.
We're excited to open source BananaMind OS, our OS for running BananaMind models. It allows you to run local AI on hardware such as a 486 or 386 (untested) Check it out at: https://github.com/BananaMind/BananaMindOS
We're excited to release BananaMind 2 Pro, our final version of the Pro model. Trained on 100B tokens it performs extremely good for its token and size class. The training took 22 days on one RTX 5070 Ti. Check it out at BananaMind/BananaMind-2-Pro We did not release a Chat version yet because it regressed. Release Later. Follow us to know when BananaMind 2 Ultra releases and support us at
Today we wanted to release BananaMind 2 Pico, our smallest model yet at ~0.9M parameters. Instead, we accidentally ran a very expensive experiment on what happens when you push a tiny model way past its useful token budget.
Short version: we trained on 200B tokens (~222K:1 tokens-per-parameter). The model peaked at 20B tokens with an INT Index of 4.55, then degraded monotonically over the next 160B to 3.31 โ a 27% regression. Three of four Open SLM benchmarks were worse at the end of training than they were at 10% through.
The useful compute-optimal range for Pico-tier models looks like ~22Kโ30K tokens per parameter. Ratios like 7K:1, 15K:1, and 22K:1 all work fine โ TinyStories and most sub-3M community models sit in this range. Push much further and benchmarks start rotting.
We're excited to release BananaMind 2 SLMoE, an experimental sequence-level mixture-of-experts model. It uses only 8M parameters per message but has 25M total parameters, 13 experts (out of 64) are selected based on the message prefix and reused for the entire response. We're testing with this sequence-level architecture to find out how big the capability loss actually is and how much of it can be fixed. The long-term idea is that this could make very large sparse models usable on machines that can't fit them in RAM by putting the entire model (which is big) on disk and only loading the active parts into VRAM. This architecture is still in research and shouldn't be used for production models.
We trained it on 60B tokens (of FineWeb-HQ, FineWeb-Edu, DCLM ,Cosmopedia v2, FineMath and NPSet-2) on 8 RTX Pro 6000s.
We're excited to release BananaMind 2 Micro, our smallest model yet. It fits a compact architecture in only 2.9M parameters achieving the highest parameter efficiency on BananaMind Base Bench against comparable models. It achieves comparable performance to GPT S2 5M and GPT S 5M at almost half the size while beating CMA 1M Mini. BananaMind 2 Micro achieved the #1 spot on the Open SLM Leaderboard for the sub 3M category (not added yet but it achieves #1) For the training we used Muon + the XSA Refresh Gate with a 5e-2 lr for Muon and 4e-3 for the 1D weights. Its score on our efficiency measure is 0.326 getting the first place with Syn 2.6M on the second place scoring 0.291 and GPT S 5M at 0.235* Check it out at BananaMind/BananaMind-2-Micro and follow us at: @vovaRL @DedeProGames @Banaxi-Tech
We're exited to announce BananaMind OS, our OS specically for running BananaMind models! Its able to run BananaMind 2 Nano at 4 bit on only 7-8MB of ram, the 2 bit on 6MB of ram and the 8 bit version on 14MB of RAM! It runs on a 486 or newer! Check out this video and image running BananaMind 2 Nano 4 Bit on 9 MB of RAM and a emulated 486 in QEMU at ~1TPS! We asked it: "What is the first letter of the alphabet?" The response is: "The first letter of the alphabet is: - A. " And if you're asking because of the video, yes I am a arch btw. Comment and like this post for a GitHub link and comment for adding other models!
BananaMind 2 Pro got better! Release on August 13-14 Follow our team for BananaMind 2 Ultra: @vovaRL for training code @Banaxi-Tech @DedeProGames for providing compute
We did an experiment, we wanted to see if AI is good enough to train models. We used GPT 5.6 Sol Max for this because its one of the most powerful ones right now. Our instructions were, it should write the training code, and start the training process and monitor it by itself. We also gave it a link to BananaMind 2 Mini to get our architecture right. The result: It worked, it made the working BananaMind 2 Nano, and even beat our previous MiniBananaMind v4 9M. Its getting way easier to develop your own models now!
We're excited to release BananaMind 2 Pro Preview, our best model yet. Trained on ~52B tokens it performs extremely good for its token and size class. We trained it on a single 5070 Ti in about 11 days. Check it out at BananaMind/BananaMind-2-Pro-Preview. Sadly we need to delay BananaMind 2 Micro until the launch of the final BananaMind 2 Pro. We will release the final checkpoint with 100B tokens in ~11 days. Go and fine-tune it! We've also released BananaMind 2 Pro Preview Chat which is the instruct version of it! BananaMind/BananaMind-2-Pro-Preview-Chat
Follow us to know when the final releases and support us at
BananaMind 
