Azim K

Co-founder & chief AI at

[Azim K @quaz1m]

Open source!

We’ve fine-tuned Whisper models to handle arbitrary audio chunks and compressed them with ANNA.

This enables optimized inference on @NVIDIA and @Apple devices with 2x faster time-to-first-token and full streaming support.

Clone the repo and start building

[Azim K @quaz1m]

@GarchFather @TheStageAI @GoogleDeepMind Great to see it deployed to the iPhones!

[TheStage AI @TheStageAI]

The smallest checkpoints for Gemma 4 E2B and E4B for local inference. Results for E2B:

size: 9.3 GB → 1.4 GB
speed: 113 tok/s on Apple M3
quality: -3% on ifEval
runs with: MLX, llama.cpp (coming)

Pareto-optimal, open source! Links to the blog post and GitHub repo ⬇️

@GoogleDeepMind @lmstudio @ollama @huggingface @ggerganov

[Azim K @quaz1m]

@TheStageAI Let's go!

[TheStage AI @TheStageAI]

TheStage AI Platform is now open to everyone.

Automatically accelerate your models and download them to run in the cloud or on smartphones.

[TheStage AI @TheStageAI]

Beyoncé heard cursing. TheWhisper heard Arsenal.

The fastest Whisper in the world.
Open-source real-time ASR.
Top 5 on OpenASR benchmarks.
1800 RTFx.

Built for live captions, transcription, and voice apps.

See the repo

[TheStage AI @TheStageAI]

For AI engineers, latency is product.

Wan 2.2 in Elastic Models now generates 5s of video in 34s on H100. Elastic Models is a library of accelerated open-source models.

Also new: TheWhisper at 1800 RTFx on a single H100 and instant FLUX LoRA switching.

Try it

[TheStage AI @TheStageAI]

How do you make text-to-music run in real time in production?

The model has to keep audio generation ahead of playback.

Our new case study with @MireloAI shows how inference optimization delivered up to 2.4х higher throughput.

See the full case study ↓

[TheStage AI @TheStageAI]

Proud to team up with @brilliantlabsAR and @neuphonicspeech on Halo’s on-device privacy engine.

Coming to Brilliant Labs’ Halo smart glasses: real-time voice + vision, POV stays private.

ANNA + GPU/NPU SDK + memory manager for wake word, STT, TTS, diarization.

SDK demo 👇

[TheStage AI @TheStageAI]

Are you a big fan of jacket potato?

This is an open-source, real-time multilingual ASR for live speech.

It stays robust in heavy noise – even at SNR 0 dB.

That’s why it understands speech where people struggle to hear.

Use it for transcription, research, and multilingual apps

[Azim K @quaz1m]

We integrated @nvidia cuDNN Paged Attention into our Elastic Models library at @TheStageAI.

It’s already running on B200: INT8 Llama-8B reaches ~200 tok/s per sequence @ batch 16 (≈3.2k tok/s total) – and we’re still tuning. Elastic Models keeps HF APIs, but can deliver up to ~4x faster inference.

Details:
https://t.co/JIa80XczH9

[TheStage AI @TheStageAI]

We know what you mean @Adele

[Azim K @quaz1m]

TheWhisper is now SOTA for Arabic, Chinese, and Hindi.

A single open-source speech-to-text model handling multiple languages and noisy audio.

On Jetson Thor it’s faster than torch.compile: RTFx 147.6 vs 106.2, TTFT 0.039s vs 0.122s

Read the full tech report →

[Azim K @quaz1m]

Up to 8x lower Time to First Token and 4x higher Real-Time Factor than other Whisper libs.

That reduces latency and lets you run more streams per GPU, like real-time captions and voice calls.

You get these gains in TheWhisper, our open source self-hosted speech-to-text engine

[Azim K @quaz1m]

We posted a tutorial on real-time on-device transcription using TheWhisper, our optimized open-source model at @TheStageAI.

It runs short windows on Apple Silicon with sub-150 ms latency and about 2 W power.

Build fast speech apps on your Mac

[Kirill Solodskikh @GarchFather]

We believe that everyone will become a model builder! That's why we are creating an automated acceleration and deployment stack which undestands ai engineers needs

[Azim K @quaz1m]

@0xabhishekc @nvidia @Apple @0xabhishekc diarization is in development now, we will release it soon

[Azim K @quaz1m]

Open source!

We’ve fine-tuned Whisper models to handle arbitrary audio chunks and compressed them with ANNA.

This enables optimized inference on @NVIDIA and @Apple devices with 2x faster time-to-first-token and full streaming support.

Clone the repo and start building

[Azim K @quaz1m]

I build ANNA. ANNA builds optimized compression for faster inference.

Seeing models breathe on @Modal always feels special. Follow our simple guide for optimized, production-ready diffusion model deployment.

[TheStage AI @TheStageAI]

We’ve made it easy to run text-to-image models on @Modal with the speed you’d expect from top inference providers.

Follow our quick guide to deploy containers with an @OpenAI compatible API and get 2× faster performance.

Big thanks to @MireloAI for the soundtrack magic 🎶

[Azim K @quaz1m]

Great result in MLPerf Inference v5.1 (@MLCommons) .

We used ANNA, our automated NNs acceleration stack.

It accelerates inference while keeping output quality high, delivering fast, stable, and production-ready performance for real workloads.

Enjoy ↓

[Azim K @quaz1m]

Validation is a key step when compressing or accelerating models.

It shows if the network still performs well.

Our research team @TheStageAI shared evaluation methods for sharpness, tone, color, object placement, and more