Rumble 2 Flash for the Astral voice pilot

Rumble transcribes the speaker on device. LFM2.5 1.2B Instruct generates a text reply, and the Apple speech synthesizer speaks it in the selected language. The download contains three GGUF files totaling 1,228,712,832 bytes (1.23 GB).

Download

Run this from the directory that will hold the pilotโ€™s Models folder.

hf download Kentucky-Open-Science/Rumble-2-Flash-Astral --revision v1.0 --include '*.gguf' --local-dir Models

For on-demand installation, read astral-manifest.json from this release. Download each file from its commit-pinned url, verify its byte count and SHA-256, and install it at its relative path under Documents/Models in the pilotโ€™s app container. The release also includes model-hashes.json for clients that consume a path-to-SHA-256 map. Download manifest ยท Checksums

Purpose File Pilot runtime
Transcription rumble/Rumble-2-Flash-AD-Q5_K_M.gguf llama.cpp with libmtmd
Audio encoder rumble/mmproj-Rumble-2-Flash-Q8_0.gguf Load as the Rumble audio projector
Reply generation chat-1.2b/LFM2.5-1.2B-Instruct-Q4_K_M.gguf llama.cpp

The pilot uses 16 kHz mono speech input and streams generated text to speech in phrases. Its GGUF runtime was built from llama.cpp revision 8af68d4.

Spoken output on iOS

English uses the default iOS English voice. Arabic, Mandarin Chinese, French, German, Japanese, Korean and Spanish use an installed Apple voice for the selected language. If a voice is missing, install it in iOS Accessibility settings. The iOS speech synthesizer manages the voice assets separately from the model download.

Sources and licenses

  • Reson8 created Rumble 2 Flash. AtomicChat produced the AD-Q5_K_M recognition weights and Q8_0 audio projector distributed here. Both files retain the upstream Apache 2.0 license. GGUF source ยท License
  • Liquid AI produced the LFM2.5 1.2B Instruct Q4_K_M chat weights. These retain the upstream LFM Open License, including its commercial-use terms. GGUF source ยท License

Kentucky Open Science assembled the three upstream files with their original contents and published checksums and language settings for the Astral pilot.

Downloads last month
-
GGUF
Model size
1B params
Architecture
lfm2
Hardware compatibility
Log In to add your hardware

4-bit

5-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Kentucky-Open-Science/Rumble-2-Flash-Astral

Quantized
(1)
this model