White chocolate @TheStageAI Co-founder, CEO, ex Huawei P50 AI cameras
Joined October 2022
- Tweets341
- Following539
- Followers698
- Likes752
Kirill Solodskikh retweeted
Very good update on our vLLM-omni version. Fastest TTS with latency and cost analysis!
Blog link 👇
Kirill Solodskikh retweeted
Whisper v3 turbo fine-tune from @TheStageAI improves over the original in both English and multilingual, and with significant speedup 🔥
Huge thanks to @EricBezzam bringing our open model to @huggingface OpenASR Leader board.
Its a great result from really small team. It beats @MistralAI 24B model and @Meta Muse on multilingual setup!
Get model card in the comment 🔽
Whisper v3 turbo fine-tune from @TheStageAI improves over the original in both English and multilingual, and with significant speedup 🔥
Good speculative decoding in @TheStageAI Apple on-device SDK.
~185 tok/s on @Apple iPhone 17 Pro NPU for @googlegemma Gemma4 E2B. Github👇
Kirill Solodskikh retweeted
We helped @JensenHuang to catch all GPUs using @liquidai LFM2.5-1.2B and our new @Apple NPU friendly speculative decoding.
4x acceleration on average. Blog and Github 👇
Base model is not fine-tuned, but still funny.
Want to test already with a new model. Fully on-device on Apple NPU.
Kirill Solodskikh retweeted
Speaker diarization + transcription running real time on my iphone 13. On the ANE. whole thing is ~300mb
TheWhisper by @TheStageAI + SOTA Nemotron-3-Diarization from @nvidia
We're shipping a new ios sdk soon with this in it, and there'll be pretty generous free limits so you can actually play with it.
Kirill Solodskikh retweeted
I made my first animated film.
12 minutes.
1 month.
12 AI tools.
1 person.
Countless failed generations.
In 2700, the oceans have taken almost all the land. Two brothers discover a way to grow it back.
This is EVERIS.
#SpecialForAAIFF
Huge thanks to the tools that became part of my one-person production pipeline along the way: @higgsfield , @Kling_ai , @ElevenLabs , @suno_ai_ , @midjourney , Seedance, @GoogleDeepMind Nano Banana
Kirill Solodskikh retweeted
Build real-time on-device multilingual TTS pipeline with TheStage AI Apple SDK.
Export your voice references and ship it with application bundle. In the result you will get dynamically switching language TTS.
Inference support @Alibaba_Qwen tts 0.6b and @neuphonicspeech multilingual nano model.
Read tutorial how to integrate it to your app: docs.thestage.ai/tutorials/s…
Kirill Solodskikh retweeted
W4A8 Metal kernels for @Apple M5. Run diffusion pipelines up to 3x faster locally on MacBook.
Github: github.com/TheStageAI/qlip-m…
Blogpost: app.thestage.ai/blog/qlip-ml…
Experimental release to utilize M5 GPU neural accelerator with int8 vector instructions.
Kirill Solodskikh retweeted
Up to 5.6x faster generation for diffusion pipelines with @ComfyUI support! New self-hosting engines!
@MiniMax_AI H3 4x faster, @krea_ai 2 turbo 5.6x faster, @ideogram_ai 1.6x faster, @AlibabaGroup Z-image-turbo 5.1x faster!
We are introducing new nodes for acceleration of your pipelines:
- Auto Sparse Attention
- Progressive generation (upsample feature maps during generation)
- Cache Nodes.
Read our tech report: app.thestage.ai/blog/ComfyUI…
Use ComfyUI nodes from our Github repo: github.com/TheStageAI/ComfyU…
Video is generated by Minimax H3.
Simulate @elonmusk.
@liquidai LFM + @TheStageAI thewhisper + @Alibaba_Qwen TTS. Fully on-device inference, no cloud costs. The best power consumption.
On-device SDK: github.com/TheStageAI/AppleS…
Read how we are building the best production grade on-device SDK using NPU inference.
Dive deeper to compression, power consumption reduction, thermal control and more!
Tech report on our SDK: app.thestage.ai/blog/Apple-S…
Github: github.com/TheStageAI/AppleS…
DM me for early SDK access and let's create great applications together!
Kirill Solodskikh retweeted
Want to learn about power consumption, thermal throttling, automated model compression, or why NPU is better than CPU and GPU?
Check out @TheStage AI Apple on-device SDK!
Read our tech report and request access to the SDK for early testing! Up to 6x power consumption reduction comparing to GPU inference.
LLMs, TTS, ASR with real-time streaming support and voice agent orchestration.
Tech report: app.thestage.ai/blog/Apple-S…
Github: github.com/TheStageAI/AppleS…
