BananaMind AI

We build small, efficient AI models that run on real hardware — phones, laptops, a single consumer GPU — instead of only in datacenters. Our releases cover compact language models, speech, vision, safety tooling and developer utilities.

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

---

What we work on

• Small language models, pretrained from scratch
• Quantization and QAT for local / on-device inference
• Text-to-speech and audio generation
• Computer vision and detection models
• AI safety and interpretability
• Benchmarks, datasets and evaluation tooling
• Mobile AI apps and developer tools

---

Models

All BananaMind 2 models are decoder-only Transformers trained from scratch on a single RTX 5070 Ti.

Shared design across the series:

• Digit-aware byte-level BPE tokenizer — digits are isolated, so numbers never merge into a single token
• Grouped-query attention with QK norm
• SwiGLU, RoPE, RMSNorm, tied embeddings

They ship custom architecture code, so load them with trust_remote_code=True.

BananaMind 2 — Base

Try them in the browser: Pro Preview demo · Medium demo

BananaMind 2 — Chat

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

Earlier generations

Licensing

• BananaMind Community License 1.0 — BananaMind 2 Pro Preview and Pro Preview Chat
• Apache 2.0 — all other models listed above

---

Benchmarks

The Elo numbers above come from our own evaluation suites:

• Base Bench 1.1 — base model comparison
• Instruct Bench 1.1 — chat / instruct comparison
• Safety Bench 1.1 — safety evaluation

Live rankings: BananaMindBench Leaderboard · SLM Leaderboard

---

Also in the org

• Safety and detection — Content-Safety-Mini-1.5, Qwen3.5-4B-Activation-Safety-Classifier, 2-AI-Detect
• Task-specific — MathBananaMind-1.1, Title-1.0
• Speech — TTS-Neon
• Mixture of experts — 2-MoE, 2-SLMoE
• Experiments — MiniBananaMind and MicroBananaMind series, StoryBananaMind, the KV1-8M quantization tests (1-bit / 2-bit / 16-bit), Overfitter-1.0

---

Versioning

0.x → 1.0 → 1.5 → 2.0 → 2.1 → ...

Intermediate versions (1.1, 1.7) only ship when there is something worth releasing between two major steps.

Roadmap

• BananaMind 2 Ultra — planned
• BananaMind 2.1 — Q4 2026 (preview checkpoints already public: 2.1-Unified, 2.1-Pico-Preview)
• BananaMind 3 — Q1–Q2 2027

---

Join us

We're looking for people already active in the community working on pretraining, QAT, small language models, AI safety, interpretability, benchmarks or datasets. If that's you, press the Join button on the organization page.

Who we are

Partners

---

If our work is useful to you, follow the org to get new releases in your feed.