Libra by GreenBitAI — a fully offline, local AI Agent OS.
Powerful AI agents running 100% on your device. No cloud. No data leaving your hardware.
See it live at #VivaTech2026 👇 @VivaTech
#LocalAI #AIAgents #OnDeviceAI
1
250
A 113 GB model on a MacBook Pro with 48 GB of unified memory.
In a GreenBitAI benchmark, gbx loaded parts of the model from SSD as needed and generated text at ~11 tokens/s.
See how it works in our #electronica2026 recap.
#GreenBitAI #LocalAI #EdgeAI #electronica2026
“Two fingers.” “Close hand.” “Rock sign.” 🤘 At electronica 2026 in Munich, this robotic hand responded to spoken commands using GreenBitAI’s voice model and inference engine. #electronica2026 #GreenBitAI #Robotics #RoboticHand #EdgeAI #VoiceControl
7
75
Behind the demos: GTA1000 voice AI.
- ASR, language understanding & TTS fully on-chip
- <700 ms response; data stays local
- GreenBitAI 1–4 bit models; full pipeline in a few hundred MB
- ~530 mW typical voice-interaction power
- No internet or subscription fees
web: gruenbit.cn
38
Day 1 at #electronica2026: multilingual TTS, a voice-controlled robotic hand and face-to-face exchanges at Giantec’s booth.
GreenBitAI provides model customization and the inference engine for GTA1000’s on-device voice AI.
Thanks to everyone who stopped by!
📅 Nov 10–13 | Hall C6 · Booth 421
1
8
119
How can large AI models run on devices with limited memory and power?
Meet GreenBitAI at #electronica2026 and explore the GTA1000, featuring fully offline voice AI powered by 1–4 bit model compression.
📅 Nov 10–13 | Hall C6 · Booth 421
#EdgeAI #OnDeviceAI
Fish Audio’s Drama 3 TTS now runs locally on Apple silicon with gbx-lm 🎙️
• 2.6x real time on 64 GB+ Macs, still runs on a 48 GB MacBook
• Zero-shot voice cloning, across languages
• OpenAI-compatible streaming API
huggingface.co/GreenBitAI/Dr…
Built with Fish Audio
78
Ternary Bonsai 2 27B on gbx v0.8.2 ⚡
• 114 tok/s on M3 Ultra, 2.7× the official MLX runtime
• 62 tok/s on DGX Spark, 2.1× llama.cpp
• 27B in 5.9 GB of memory
GreenBit Ternary Models, coming soon.
GreenBitAI retweeted
⚡ Inside DeepInfer
🔹 High-performance kernels & optimized inference engine
🔹 Distributed inference & speculative decoding
🔹 SWA Bounded Replay & structured outputs
Inference code for future DeepSeek models and ongoing optimizations will land here too.
2/4
It works with Claude Code and Codex out of the box; setup in the model card.
Model: huggingface.co/GreenBitAI/De…
Engine: github.com/GreenBitAI/gbx-lm…
@deepseek_ai
1
39
Why it's worth running locally: DeepSeek reports 77.0 on DeepSWE v1.1 (from 74.2) and 42.3 on Terminal-Bench 4.0 (from 31.2), with shorter outputs.
An agentic coding model that now stays on your own machine.
1
53
DeepSeek's new V4.1-Flash-1017, running locally on a Mac.
Our 4-bit build on one M3 Ultra Mac Studio: • ~22 tok/s decode • ~44 tok/s on code with DeepSeek's draft head (1.99×) • ~33 tok/s chat
Replying to @moelkholy95
Good math, and you're right that it's read-bound. Two corrections though. 32 GB is the process's peak footprint, not the expert cache. The cache itself is 15.7 GiB: 114 of 512 experts per layer (22%), sized automatically from the machine's memory. And ~11 tok/s is with the MTP draft head on, where one verify pass reads each expert once for several tokens. Plain paged decode at the same budget is 6–11 tok/s, depending on the prompt.
Cache hit rate is 61–76% without MTP and 52–70% with it (the draft head takes ~11 slots per layer). Measured on a Studio at the same 25 GiB budget, cold reads. That leaves ~270 MB per token from SSD: about 70 ms waiting on disk and 60 ms of GPU work.
On topic shifts: the cache is a per-layer LRU, so it follows routing within a few tokens. Our runs mix code, prose and chat on one server, and decode stays in a 9.6–12 tok/s band. We haven't measured a switch in the middle of an answer on its own. Good idea, we'll add it.
2
31
Full write-up with methodology: medium.com/@semamediadata/ru…
Download: github.com/GreenBitAI/gbx-lm…
Models: huggingface.co/collections/G…
Which model would you run locally next?
26
How it works: • cost-aware adaptive speculative decoding • mixed-precision draft head • elastic expert memory that streams MoE experts from the SSD • fused kernels for MLA, linear attention, Hyper-Connections
One shared layer: Qwen3.5/3.6/3.8, GLM-5.3, DeepSeek-V4.1-Flash.
1
49
vs Strata (SSD-streaming engine on an RTX 5070 PC, its published numbers), same model at ~4-bit, 16K prompt, cold SSD:
• prompt: 591 vs 160 tok/s (3.7×) • decode: 12.6 vs 7.9 tok/s
Different hardware, so not like-for-like. Full conditions in the write-up.
1
65
Not enough memory? Same 105 GB model, 25 GB budget, experts read cold from the SSD, vs omlx:
• 7k prompt: 362 vs 123 tok/s (2.9×) • decode: 12.7–14.5 vs 8.3–9.7 tok/s • later chat turns: 13 s vs 27 s to first token
mlx-lm and mlx-vlm can't load it at this budget.
1
62
On Qwen3.8-Flash-Next (4-bit, fully in memory) the lead grows: 53 vs 35 tok/s without speculative decoding, and +22–32% with it across code, prose and chat.
Prompt processing is on par.
1
41























