Skip to content
English

Switch theme

// Ant Ling / Ring / Ming · Text-to-speech / audio

Ming-3.0-audio-flash: VRAM & local deployment

126B total · 6.5B active (MoE) parameters. About 84.5 GB at 4-bit and 278 GB at BF16.

released 2026-09-03licence: Commercial OKmemory figures are estimates

01Specs

PublisherAnt Group inclusionAI · inclusionAI
Family / seriesAnt Ling / Ring / Ming · Ming-3.0-audio-flash
TypeText-to-speech / audio (text-to-speech)
Parameters126B total · 6.5B active (MoE) Source: official model card“126B total parameters, 6.5B activated per inference step”
Released weightsBF16 · 256 GBdownload
LicenceMIT · Commercial OKPermissive licence (Apache-2.0 / MIT / CC-BY / OpenMDW): commercial use allowed.
Released2026-09-03 (Hugging Face repo created)

02How much VRAM it needs (estimate)

Audio models are small; the weight size is close to the total memory needed.

PrecisionWeightsWeights + overheadSmallest GPUMac
BF16 / FP16 (original)252 GB 278 GB 4 × RTX PRO 6000 (384GB)Mac 512GB
FP8 / INT8 / Q8_0134 GB 148 GB 4 × RTX PRO 6000 (384GB)Mac 256GB
4-bit (Q4_K_M / AWQ / GPTQ-Int4)76.4 GB 84.5 GB RTX PRO 6000 96GBMac 128GB

Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →

03How to run it

Only runtimes the publisher or the runtime’s official library names for this model.

Which inference engine to choose →

04Official repositories & quantised builds

1 official repo(s) from inclusionAI.

05Other Ant Ling / Ring / Ming models

All 29 →
ModelParams4-bit totalReleased
Ling-3.0-tiny-UBTree183M1.0 GB2026-10-21
Ling-3.0-flash-UBTree482M1.5 GB2026-10-14
Ming-3.0-asr-tiny8.87B6.4 GB2026-10-14
Ling-3.1-flash560B / 25B375 GB2026-10-13
Ming-3.0-tts-0.5B1.48B1.5 GB2026-09-22
Ming-3.0-tts-tiny8.98B / 1.3B6.5 GB2026-09-22
Ming-Image-0.1-Design6.16B4.6 GB2026-09-17
Ling-3.0-flash-VL124B / 5.5B83.6 GB2026-09-04

06Sources

Data checked 2026-10-22. Parameters, context and licence are copied from the official repository; memory figures are our estimates.

Data updated: Self-host open models(UTC+8)