Ultra is the full-size model, trained further. Fewer errors than the original on every test we ran.
18% fewer across 25 languages.
28% fewer on full-length talks.
18% fewer across 25 languages.
28% fewer on full-length talks.
1
0
15
1,093
Today we're releasing Moondream Parakeet Redux and Moondream Parakeet Ultra. Two speech-to-text models based on NVIDIA's Parakeet. 25 languages. They run locally, wickedly fast.
Redux is for CPUs. Ultra is for GPUs.
Redux is for CPUs. Ultra is for GPUs.
14
53
609
29,880
Shrinking a model usually costs accuracy, but redux holds up. Same or better results across 25 languages. Noisy audio is where it slips. For that, use Ultra.
1
0
13
1,083
Redux shrinks the model from 1.2 GB to 178 MB. Ternary quantized, so every encoder weight is -1, 0, or +1.
On a CPU, speed is limited by moving weights through memory. Less to move means faster: 1 hour of audio in 32 sec on 8 cores, 2.5x faster than parakeet.cpp. On a MacBook Air, 1 min 35 sec.
On a CPU, speed is limited by moving weights through memory. Less to move means faster: 1 hour of audio in 32 sec on 8 cores, 2.5x faster than parakeet.cpp. On a MacBook Air, 1 min 35 sec.
2
0
35
2,246
Photon can hear now.
Photon 2.1 adds streaming speech recognition for Whisper, Qwen3-ASR, and Parakeet. It won all 16 matched H100/B200 tests against vLLM, Qwen-ASR, and NeMo, with up to 3.1× the throughput.
Free:
pip install -U moondream
moondream.ai/blog/photon-2-…
Photon 2.1 adds streaming speech recognition for Whisper, Qwen3-ASR, and Parakeet. It won all 16 matched H100/B200 tests against vLLM, Qwen-ASR, and NeMo, with up to 3.1× the throughput.
Free:
pip install -U moondream
moondream.ai/blog/photon-2-…
4
19
286
27,896
AI struggles with live video. Here's why.
General-purpose engines are built for throughput. Cameras need fresh context, bounded memory, predictable latency, and hours-long stability.
We found five failure modes. We're building Photon to solve them.
moondream.ai/blog/live-vide…
General-purpose engines are built for throughput. Cameras need fresh context, bounded memory, predictable latency, and hours-long stability.
We found five failure modes. We're building Photon to solve them.
moondream.ai/blog/live-vide…
2
1
33
3,037
4. Long-run stability
Reports document a vLLM deployment reaching OOM after days of VRAM growth, an SGLang VLM leak, and a DeepStream pipeline crashing after ~7 hours as latency drifted.
Robots need 24/7 uptime.
Reports document a vLLM deployment reaching OOM after days of VRAM growth, an SGLang VLM leak, and a DeepStream pipeline crashing after ~7 hours as latency drifted.
Robots need 24/7 uptime.
1
0
4
247
5. Cold starts
Two published vLLM measurements found cold starts of roughly 3–8 minutes.
On robots, a power cycle or restart after a fault can mean minutes of downtime.
We're building Photon Live Video and taking three design partners. Running this workload? DM us.
Two published vLLM measurements found cold starts of roughly 3–8 minutes.
On robots, a power cycle or restart after a fault can mean minutes of downtime.
We're building Photon Live Video and taking three design partners. Running this workload? DM us.
0
0
5
220
Photon 2.0 is out.
An inference engine built for Physical AI. Supports Moondream, Qwen, and Gemma (more soon). Beats vLLM and SGLang throughput, up to 2.3x.
Here's why Physical AI needs its own engine 🧵
moondream.ai/blog/photon-2-…
An inference engine built for Physical AI. Supports Moondream, Qwen, and Gemma (more soon). Beats vLLM and SGLang throughput, up to 2.3x.
Here's why Physical AI needs its own engine 🧵
moondream.ai/blog/photon-2-…
21
33
477
145,779
Available now: `pip install moondream`.
Engine: Apache 2.0
Megakernels: free to run
Compiler: private (ours)
Engine: Apache 2.0
Megakernels: free to run
Compiler: private (ours)
1
1
38
3,588
Inference is entering its compiler era.
Building robotics / live video / industrial perception / computer-use agents? DMs open. Bring your weirdest workload.
Moondream: Every Machine Will See.
Building robotics / live video / industrial perception / computer-use agents? DMs open. Bring your weirdest workload.
Moondream: Every Machine Will See.
0
1
43
2,017
Photon 1.3 is out!
It's now free and runs Moondream up to 70% faster. More images/sec, lower latency, on Mac, Windows & NVIDIA.
Good, fast, cheap: we picked all three!
moondream.ai/blog/photon-1-…
It's now free and runs Moondream up to 70% faster. More images/sec, lower latency, on Mac, Windows & NVIDIA.
Good, fast, cheap: we picked all three!
moondream.ai/blog/photon-1-…
4
3
80
20,440