batisay(STT, 무μμ λ§νλ) + batispeak(νμλΆλ¦¬, λκ° λ§νλ) = ν΅νΒ·νμ νμλ³ μ μ¬. 16GB Mac on-device.
AI & ML interests
On-device AI, GGUF quantization, Apple Silicon, macOS automation
Recent Activity
Organization Card
BatiAI β The Frontier, On Your Mac π
Kimi K2.6 β 1T MoE, SWE-Bench Pro 58.6 (beats GPT-5.4 xhigh, Claude Opus 4.6) β running locally on M3 Ultra. Gemma 4 E4B β 57 tokens/sec on a 16GB Mac mini M4. No API costs. No rate limits. No cloud.
We quantize popular open-weight models for every Mac β direct from official publisher weights, calibrated with imatrix, verified on real hardware, signed with provenance metadata.
β‘ Quick Start
# 16GB Mac mini M4 β entry, 57 t/s
ollama pull batiai/gemma4-e4b:q4
# 512GB M3 Ultra β 1T MoE frontier
ollama pull batiai/kimi-k2.6:iq4
Pick Your Mac β Real Hardware Benchmarks
Every speed measured on actual hardware. Full reports in each model card. We continuously add new hardware as it ships.
| Your Mac | Best Pick | Size | Speed | Use Case |
|---|---|---|---|---|
| Mac mini M4 16GB | batiai/gemma4-e4b:q4 |
5.0 GB | 57 t/s | Daily chat, balanced |
| MacBook Air 16GB | batiai/qwen3.5-9b:q4 |
5.2 GB | 12.5 t/s | Tool calling, JSON |
| Mac mini M4 Pro 24GB | batiai/gemma4-26b:iq4 |
15 GB | 85 t/s | MoE, larger context |
| MacBook Pro 32GB | batiai/nemotron3-nano:iq4 |
17 GB | β | NVIDIA NemotronH MoE, hybrid Mamba+Attention |
| MacBook Pro 48GB | batiai/qwen3.6-35b:iq4 |
22 GB | ~30 t/s | Tools + thinking, MoE |
| MacBook Pro 128GB | batiai/nemotron3-super:iq4 |
62 GB | β | NemotronH 120B-A12B, agentic |
| MacBook Pro 96GB | batiai/qwen3.6-35b:q6 |
29 GB | ~27 t/s | Top quality chat |
| MacBook Pro M4 Max 128GB | batiai/minimax-m2.7:iq3 |
82 GB | 36.7 t/s | 229B Dense β frontier class |
| MacBook Pro M4 Max 128GB | batiai/DeepSeek-V4-Flash-GGUF (Q3) |
127 GB | β | 284B-A13B MoE β DeepSeek V4 early access via bati.cpp |
| Mac Studio M3 Ultra 512GB | batiai/kimi-k2.6:iq4 |
509 GB | β | 1T MoE, SWE-Bench Pro 58.6 |
| Mac Studio M3 Ultra 512GB | batiai/DeepSeek-V4-Flash-GGUF (Q8) |
282 GB | β | DeepSeek V4-Flash near-lossless, frontier coding |
Real measurements, not estimates. Numbers expand as benchmarks come in.
Browse the Collections below β for series-by-series lineup.
π°π· BatiAI Perception Suite β our own Korean, on-device models
Beyond quantizing frontier LLMs, BatiAI builds and fine-tunes its own Korean-first perception models β speech, speaker, and document understanding that run entirely on a Mac. Named for what they do: say (speech) Β· speak (speaker) Β· see (document/vision).
| Model | Task | Highlights |
|---|---|---|
| batisay Β· speech-to-text | Korean STT | batisay-ko-turbo (GGUF) β multi-domain (call Β· meeting Β· conversation); also batisay-ko-base Β· batisay-ko-large |
| batispeak Β· speaker diarization | Korean who-spoke-when | batispeak-diarize β Korean speaker segmentation (preview, accuracy still improving) |
| batisee Β· document OCR | Korean document/handwriting | batisee v1 + v2 β text Β· tablesβHTML Β· handwriting; DeepSeek-OCR arch, MIT, built from original weights |
Korean-verified, BatiAI-signed, on-device. These are our own models β not third-party re-hosts.
π― Why BatiAI?
Direct from SourceQuantized from the publisher's official FP8/BF16 weights β never re-quantized from third-party GGUFs. Every file signed |
Verified on Real MacsTested on Mac mini M4 16GB + MacBook Pro M4 Max 128GB. Korean validation, tool-call JSON, 200-token throughput β measured, reproducible, documented in each model card. |
imatrix-CalibratedEvery model uses importance-matrix calibration with wikitext-2. Aggressive IQ3_XXS keeps quality where plain Q3_K_M visibly degrades. |
Frontier-CapableWe handle 1T MoE (Kimi K2.6), 229B Dense (MiniMax M2.7), NVIDIA NemotronH (Mamba-hybrid 30Bβ120B), and now DeepSeek V4-Flash 284B via our own |
π§° Built for Real Mac Users
| BatiFlow | 5MB native macOS app. Connects BatiAI models to 60+ tools β KakaoTalk, iMessage, Slack, Calendar, Notes, Chrome, file system, browser. Free. Unlimited. 100% private. Even drives your Mac via Telegram / Discord / Slack bots. |
| Bati CIS | K-Beauty Commerce Intelligence β settlement processing 3 days β 3 hours, 42+ marketplaces. Trusted by COSRX, Pharma Research (Rejuran). Revenue Anomaly Β· Repurchase Cohort Β· Marketing Budget Optimizer. |
π§ jk@bati.ai (enterprise) Β· π¬ GitHub Issues Β· π bati.ai
Private by default Β· On-device first Β· Verified every step.
models 41
batiai/batisay-ko-large
Automatic Speech Recognition β’ Updated β’ 1
batiai/batisay-ko-turbo
Automatic Speech Recognition β’ 0.8B β’ Updated β’ 124 β’ 2
batiai/batisay-ko-base
Automatic Speech Recognition β’ 0.8B β’ Updated β’ 368 β’ 1
batiai/batisay-ko-turbo-mlx
Automatic Speech Recognition β’ Updated β’ 43
batiai/batisee
Image-Text-to-Text β’ 3B β’ Updated β’ 194
batiai/batispeak-diarize
Updated β’ 22
batiai/batisay-ko-turbo-GGUF
Automatic Speech Recognition β’ Updated
batiai/GLM-5.2-GGUF
Text Generation β’ 753B β’ Updated β’ 520
batiai/Mistral-Small-4-119B-GGUF
Text Generation β’ 119B β’ Updated β’ 177
batiai/Mistral-Medium-3.5-128B-GGUF
Text Generation β’ 125B β’ Updated β’ 798 β’ 1
datasets 0
None public yet