---
title: Dia2-2B-bf16
canonical_url: "https://www.modelscope.cn/models/mlx-community/Dia2-2B-bf16"
md_url: "https://www.modelscope.cn/models/mlx-community/Dia2-2B-bf16.md"
repository: mlx-community/Dia2-2B-bf16
last_updated: 2026-10-07
license: apache-2.0
pipeline_tag: text-to-speech
tasks:
  - text-to-speech
base_model:
  - nari-labs/Dia2-2B
base_model_relation: quantized
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
language:
  - en
downloads: 47
stars: 0
tags:
  - mlx
  - mlx-swift
  - dia2
  - mimi
  - dialogue
  - text-to-speech
---

# Dia2-2B-bf16

> Dia2-2B-bf16 - mlx-community 在 ModelScope 开源的模型。nari-labs/Dia2-2B is Nari Labs' two-speaker dialogue TTS: a script with [S1] / [S2] turns renders as one take in which both speakers trade lines. This repo is that checkpoint converted for MLX-Swift. It is the…

mlx-community/Dia2-2B-bf16 是 ModelScope 魔搭社区上的text-to-speech模型，采用 apache-2.0 许可，基于 nari-labs/Dia2-2B 构建。

- **Repository**: mlx-community/Dia2-2B-bf16
- **License**: apache-2.0
- **Tasks**: text-to-speech
- **Base model**: nari-labs/Dia2-2B
- **Tags**: mlx, mlx-swift, dia2, mimi, dialogue, text-to-speech
- **Downloads**: 47
- **Stars**: 0
- **Last updated**: 2026-10-07

Source: https://www.modelscope.cn/models/mlx-community/Dia2-2B-bf16

---

# Dia2-2B — MLX, bf16

[`nari-labs/Dia2-2B`](https://huggingface.co/nari-labs/Dia2-2B) is Nari Labs' two-speaker dialogue TTS: a
script with `[S1]` / `[S2]` turns renders as one take in which both speakers trade lines. This repo is that
checkpoint converted for [MLX-Swift](https://github.com/ml-explore/mlx-swift). It is the weights repo of
[`mlx-dia2-tts-swift`](https://github.com/xocialize/mlx-dia2-tts-swift), a Swift port of upstream's runtime and
an [MLXEngine](https://github.com/xocialize/mlx-engine-swift) `tts` package.

| file | what | dtype |
|---|---|---|
| `model.safetensors` | the Dia2 decoder (28 × 2048, GQA 16/8) + depformer (4 × 1024, 5 weight sets) — upstream's keys, unchanged | bf16 (RMSNorm weights float32 at load) |
| `mimi.safetensors` | Kyutai's Mimi codec (32 codebooks, 12.5 Hz, 24 kHz), re-keyed onto the module layout of [`kyutai-labs/moshi-swift`](https://github.com/kyutai-labs/moshi-swift) | float32 |
| `config.json`, tokenizer files | upstream's config (+ `"mimi": {"num_codebooks": 32}`) and its GPT-2 BPE tokenizer | — |

**Layout note.** `model.safetensors` keeps upstream's keys. `mimi.safetensors` is in moshi-swift's layout: fused
attention `in_proj` with q/k back in Kyutai's interleaved RoPE order, and conv weights as MLX `(O, K, I)`. Loaders
expecting transformers' `MimiModel` keys will not read it.

## Conversion

Sources:
- `nari-labs/Dia2-2B` @ `7abae125471a73b0fc6b9d413cb15f4ae1e771d8`
- `kyutai/mimi` @ `89091b3e466eb6a9d11e537bf26b144f194978f7`

```bash
python Tools/oracle-capture/convert.py Dia2-2B-bf16 --size 2B --dtype bfloat16   # in mlx-dia2-tts-swift
```

## Parity and measurements

The Swift port is gated against upstream's PyTorch runtime (`nari-labs/dia2` @ `8687268`) in fp32. From the
package's `PORTING-SPEC.md` / `MEASUREMENTS.md`:
- **Gates:** the tokenizer is id-exact; the decoder is within 1.05e-5. Every one of 4 257 sampled distributions
  matches upstream within total variation 1.1e-5, except one documented near-tie at the CFG filter's boundary.
  Codes are token-exact under replay; the waveform is within 114–118 dB.
- **This tier:**
  - On upstream's own evaluation jobs it scores within sampling noise of upstream's fp32 run (scenes, single lines,
    voice prefixes).
  - On an M5 Max the package declares 4.5 GB resident + 3.0 GB activation (measured phys_footprint, including
    MLXEngine's 2 GiB buffer pool). The model's own working set is ≤ 1.2 GB, flat over takes of 3–101 s. It runs
    at ≈ 1.8× realtime.

## Use (Swift)

```swift
import MLXServeCore
import MLXDia2TTS

let engine = MLXServeEngine()
try await engine.register(Dia2TTSPackage.registration, configuration: Dia2TTSConfiguration())
try await engine.prepare(.tts)
let wav = try await engine.run(TTSRequest(
    text: "[S1] Did you hear that? [S2] Hear what? It's the wind. [S1] No, it was a voice.",
    metaData: ["seed": .int(7)]))
```

Voice prefixes: `voice: .referenceAudio(clip)` + `referenceTranscript` set speaker 1. Speaker 2 is set with
`metaData` `speaker2Audio` (a base64 `.wav`) + `speaker2Transcript`. Prefix both speakers for a consistent pair; a
single prefix conditions only weakly.

## Licence

- **Dia2:** `model.safetensors`, `config.json` and the tokenizer files are `nari-labs/Dia2-2B`, Apache License 2.0
  (`LICENSE`), © Nari Labs. Changes: cast to bfloat16, plus the `"mimi"` entry in `config.json`.
- **Mimi:** `mimi.safetensors` is **Mimi by Kyutai**, from [`kyutai/mimi`](https://huggingface.co/kyutai/mimi),
  licensed [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). Changes: re-keyed to the moshi-swift layout,
  attention q/k un-permuted and re-fused, conv weights transposed, stored float32.