Ming-flash-omni 2.0 Audio DiT — Native Component

Extracted and validated on 2026-10-27.

This repository exports the trained local audio DiT from Ming-flash-omni 2.0. The scope is exactly cfm.model.* within talker/model.safetensors. It produces audio-latent flow predictions; complete speech, sound or music synthesis still requires the matching Talker and AudioVAE.

Included scope

• All 8 attention-transformer layers, timestep, input and condition embeddings, rotary state, normalization and output head.
• 103,117,920 parameters, 124 tensors and 206,235,840 payload bytes in original BF16.
• Unmodified native DiT and CFM sources. The cfm.model. prefix is removed without changing tensor values or dtypes; per-tensor SHA-256 and key mappings are included.

Native interface and dependencies

Inputs are [B,current_patch,64] noisy audio latents, [B,history_patch,64] history and [B,1,896] Talker hidden states. Native defaults are current patch 4, history 32 and 10 CFM steps. Preserve the nonuniform EPSS times, sway transformation, conditional + strength*(conditional-null) guidance semantics and explicit SDE noise inputs.

Talker language weights, aggregator, speaker conditioning, text normalization and AudioVAE weights are excluded. Their native dependency configurations are included for interface documentation. This package does not directly accept text or emit waveform audio.

import sys, torch
from huggingface_hub import snapshot_download
path = snapshot_download("gatilin/ming_flash_omni2_audio_dit")
sys.path.insert(0, path)
from component import load_component
dit = load_component(path, dtype=torch.float32, device="cpu")
# Interface smoke fixture only; use real Talker/AudioVAE inputs for synthesis.
noise = torch.randn(1, 4, 64)
history = torch.zeros(1, 32, 64)
condition = torch.randn(1, 1, 896)
prediction = dit.predict(noise, 0.4, condition, history)
sampled_latents = dit.sample(condition, history, noise)

Source and reproducibility

• Official source: inclusionAI/Ming-flash-omni-2.0.
• Weight revision: c1ed5565e64e85a76826255c000b66a20d4634c3.
• Source file: talker/model.safetensors; official full-file SHA-256: 1403c23a0a915b31a03dd05be7854ddb22ca71f776a3642ac310cd9886284f39.
• Export SHA-256: 53c27afb4fe815ac6ca08942ed24ca4416c055f19dee336e52cfcdb71dae9bef.

See extraction.json, source_code_manifest.json and requirements.txt. Loading requires explicitly importing the provided component code. No complete language model is bundled.

Validation and limits

All 18 recorded checks passed: strict native loading, source-weight integrity, full-depth prediction, independent attention comparison, condition/time responses, batch isolation and native sampling trajectories.

Tests use real checkpoint parameters, CPU float32 execution and small synthetic conditions/latents. They do not measure real-prompt generation, waveform/image decoding, perceptual quality or complete upstream task performance. Storage dtypes remain unchanged. See validation.json; rerun with python validation_reference.py ming_audio.

Attribution and licenses

The original inclusionAI/Ming MIT license is included. Native source-file copyright and attribution notices, including the DiT references, are retained. Dependency models and third-party packages keep their original terms.