BananaMind AI

company

https://bananamind-web.web.app

BananaMind

Activity Feed Request to join this org

AI & ML interests

Pretraining, QAT, SLM, Overtraining, AI Interpretability.

Recent Activity

Banaxi-Tech new activity about 3 hours ago

BananaMind/BananaMindBench-Leaderboard:models to add

Banaxi-Tech updated a Space about 4 hours ago

BananaMind/BananaMindBench-Leaderboard

Banaxi-Tech new activity about 4 hours ago

BananaMind/BananaMindBench-Leaderboard:models to add

View all activity

Organization Card

Community About org cards

BananaMind AI

BananaMind AI builds practical AI models, tools, and experiments focused on local inference, mobile deployment, and efficient open-source AI.

Our work includes compact language models, speech models, vision models, safety tooling, and developer utilities designed to run on real hardware instead of only large cloud systems.

If you're interested, give us a follow.

Models

All BananaMind 2 models are decoder-only Transformers trained from scratch on a single RTX 5070 Ti, using digit-aware byte-level BPE tokenizers (digits are isolated so numbers never merge into single tokens), grouped-query attention with QK norm, SwiGLU, RoPE, RMSNorm and tied embeddings. They ship custom architecture code, so load them with trust_remote_code=True.

Try them in the browser

BananaMind 2 — Base

Model	Params	Context	Vocab	Tokens	Notes
BananaMind-2-Pro-Preview (demo)	139M	3,072	32,768	51.9B / 100B	First public checkpoint of the ongoing Pro run (96K steps). INT Index 23.04, Base Bench 1.1 Elo 1106
BananaMind-2-Medium (demo)	49.6M	3,072	12,288	50B	39.27 benchmark average — HellaSwag 32.43, ARC-E 43.81, PIQA 61.86
BananaMind-2-Mini	25.2M	4,096	8,192	30B	First model of the BananaMind 2 series
BananaMind-2-Nano	10.0M	4,096	8,192	30B	Smallest base model, 35.77 average

BananaMind 2 — Chat

Instruction-tuned on smol-smoltalk with full-parameter SFT and assistant-only loss masking. System prompts and multi-turn conversations supported.

Model	Params	Base model	Instruct Bench 1.1 Elo
BananaMind-2-Pro-Preview-Chat	139M	Pro Preview	888
BananaMind-2-Medium-Chat	49.6M	Medium	787
BananaMind-2-Mini-Chat	25.2M	Mini	654
BananaMind-2-Nano-Chat	10.0M	Nano	643

Earlier generations

Model	Params	Context	Tokens	Notes
BananaMind-1.5-Base	75.1M	4,096	~27B	Llama-style architecture, our first fully pretrained medium model

Licensing: BananaMind 2 Pro Preview and Pro Preview Chat are released under the BananaMind Community License 1.0. All other models above are Apache 2.0.

Focus Areas

Local and on-device AI
Small and efficient language models
Text-to-speech and audio generation
Computer vision models
Model quantization and optimization
Open-source AI tooling
Experiments with mobile AI apps

Naming Scheme

0.x > 1.0 > 1.1 (sometimes) > 1.5 > 1.7 (sometimes) > 2.0 > 2.1...

Partners

If you like our models, check out our partners:

FromZero

SupraLabs

AxiomicLabs

Projects

BananaMind AI hosts experimental and production-ready models for:

Mobile-friendly AI assistants
Local LLM inference
Speech and TTS systems
Face and vision detection
AI safety and evaluation
Developer tools for model testing

We believe AI shouldn't need a datacenter to be smart.

Upcoming Models

BananaMind 2 Ultra!
BananaMind 2.1, 2026 Q4
BananaMind 3, Q1-Q2 2027

Join us

If you're into any of the following requirements and known in the community, press join button on the organization! Requirements: Pretraining, QAT, SLM, AI safety, AI Interpretability, benchmarks, datasets.