Tele-AI · lzx1413 · Aug 5, 2026 · Aug 4, 2026 · Aug 4, 2026 · Aug 4, 2026
diff --git a/.gitignore b/.gitignore
@@ -162,4 +162,10 @@ telefuser/_version.py
 !examples/data/lingbot_world_fast/image.jpg
 !examples/data/lingbot_world_fast/poses.npy
 !examples/data/lingbot_world_fast/intrinsics.npy
-.models
+# MiniMax H3 runnable example requests and frozen official reference inputs
+!examples/data/minimax-h3/fl2va-reference.png
+!examples/data/minimax-h3/ref2va-reference.mp4
+!examples/data/minimax-h3/ref2va-voice.mp3
+!examples/data/minimax-h3/ref2va.json
+!examples/data/minimax-h3/provenance.json
+.models
diff --git a/README.md b/README.md
@@ -17,6 +17,11 @@ TeleFuser is a high-performance runtime for world model inference and multimodal
 
 ## News 📰
 
+- ✨ **2026-08-05**: Added [**MiniMax H3**](examples/minimax_h3/README.md) T2VA, FL2VA, and Ref2VA joint
+  audio-video generation with standard `telefuser serve` support. On the matched 768p, five-second, 50-step T2VA
+  benchmark after warmup, the resident **4 x H100 80 GB** profile matched pinned local SGLang SP2+TP2 performance
+  while using less GPU memory; see the
+  [reproducible validation and performance notes](examples/minimax_h3/README.md#measured-four-gpu-profile).
 - ✨ **2026-08-03**: Validated LingBot-World v2 target-side real-time generation on **4 x H100 80 GB** at
   832x480 and 16 FPS. The current 77-frame gate reached **17.14 steady compute FPS**; see the
   [reproducible benchmark](docs/en/benchmark_aiperf.md#current-four-h100-real-time-gate).
@@ -230,6 +235,7 @@ telefuser/
 | `WanVideo` (Wan2.1 / Wan2.2) | T2V, I2V, FL2V | Main video generation family, including async and service examples in [examples/wan_video/README.md](examples/wan_video/README.md) |
 | `HunyuanVideo` | T2V, I2V | Supported via [examples/hunyuan_video/README.md](examples/hunyuan_video/README.md) |
 | `LTX Video` | I2V + Audio | Unified audio-video generation via [examples/ltx_video/README.md](examples/ltx_video/README.md) |
+| `MiniMax H3` | T2VA, FL2VA, Ref2VA + Audio | Local 768p joint audio-video generation via [examples/minimax_h3/README.md](examples/minimax_h3/README.md) |
 | `LongCat-Video` | T2V, I2V, VC | Long-form generation and continuation via [examples/longcat_video/README.md](examples/longcat_video/README.md) |
 | **NEW** `LingBot-Video` | T2I, T2V, TI2V, MoE refiner | Dense/MoE generation with native CFG/SP and an in-memory base-to-refiner path; see [examples/lingbot_video/README.md](examples/lingbot_video/README.md) |
 

diff --git a/README_zh.md b/README_zh.md
@@ -17,6 +17,11 @@ TeleFuser 是一个面向世界模型推理与多模态生成的高性能运行
 
 ## News 📰
 
+- ✨ **2026-08-05**：新增 [**MiniMax H3**](examples/minimax_h3/README.md) T2VA、FL2VA 和 Ref2VA 联合
+  音视频生成，并支持标准 `telefuser serve` 服务模式。在相同的 768p、5 秒、50 步 T2VA 请求和一次预热
+  条件下，常驻 **4 张 H100 80 GB** 配置的性能与固定版本的本地 SGLang SP2+TP2 对齐，同时使用更少
+  GPU 显存。复现条件和性能分析见
+  [MiniMax H3 示例文档](examples/minimax_h3/README.md#measured-four-gpu-profile)。
 - ✨ **2026-08-03**：LingBot-World v2 已在 **4 张 H100 80 GB** 上通过 832x480、16 FPS 的目标侧实时生成
   验证。当前 77 帧门禁达到 **17.14 steady compute FPS**，复现方法见
   [基准文档](docs/zh/benchmark_aiperf.md#当前-77-帧实时计算门禁)。
@@ -220,6 +225,7 @@ telefuser/
 | `WanVideo` (Wan2.1 / Wan2.2) | T2V, I2V, FL2V | 主力视频生成家族，含异步和服务示例，见 [examples/wan_video/README.md](examples/wan_video/README.md) |
 | `HunyuanVideo` | T2V, I2V | 见 [examples/hunyuan_video/README.md](examples/hunyuan_video/README.md) |
 | `LTX Video` | I2V + Audio | 统一音视频生成，见 [examples/ltx_video/README.md](examples/ltx_video/README.md) |
+| `MiniMax H3` | T2VA, FL2VA, Ref2VA + Audio | 本地 768p 联合音视频生成，见 [examples/minimax_h3/README.md](examples/minimax_h3/README.md) |
 | `LongCat-Video` | T2V, I2V, VC | 长视频生成与续写，见 [examples/longcat_video/README.md](examples/longcat_video/README.md) |
 | **NEW** `LingBot-Video` | T2I, T2V, TI2V, MoE refiner | 支持原生 CFG/SP 的 Dense/MoE 生成与内存直传 base-to-refiner，见 [examples/lingbot_video/README.md](examples/lingbot_video/README.md) |
 

diff --git a/docs/en/parallel.md b/docs/en/parallel.md
@@ -461,9 +461,8 @@ pool.
 # Device count must equal product of parallel degrees
 world_size = dp * cfg * sp_ring * sp_ulysses * pp * tp
 
-# SP and TP cannot be enabled simultaneously
-if sp_degree > 1 and tp_degree > 1:
-    raise ValueError("SP and TP are mutually exclusive")
+# SP and TP may be combined as independent mesh dimensions.
+# The selected pipeline must implement and validate the requested combination.
 ```
 
 ## Usage Examples

diff --git a/docs/en/testing.md b/docs/en/testing.md
@@ -365,6 +365,10 @@ If not set, defaults to `"model_zoo"` (relative to working directory).
 | prompt | str\|null | null | Override generation prompt |
 | input_image_path | str\|null | null | Input image for I2V/edit pipelines |
 | input_video_path | str\|null | null | Input video for VSR/continue pipelines |
+| target_video_length | float\|null | null | Override generated duration when the example accepts it |
+| use_run_with_file | bool | false | Use the standard file entrypoint to preserve container-level output such as audio tracks |
+| require_audio | bool | false | Require an audio stream and compare its contract and decoded waveform |
+| audio_cosine_min | float | 0.95 | Video audio: minimum decoded-waveform cosine similarity |
 | ppl_config_overrides | dict | {} | Override PPL_CONFIG keys |
 | psnr_min | float | 25.0 | Video: minimum PSNR vs baseline |
 | ssim_min | float | 0.85 | Video: minimum SSIM vs baseline |
@@ -407,7 +411,8 @@ Example: `20260402_120000_wan_video__wan21_1_3b_text_to_video_h100_1gpu.log`
 
 The runner compares outputs against baselines using:
 
-- **Video**: PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity)
+- **Video**: PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity); when `require_audio` is enabled, audio
+  stream presence, sample rate, channels, duration, and decoded-waveform cosine similarity
 - **Image**: Mean pixel difference
 
 #### Metrics Thresholds
@@ -417,6 +422,8 @@ Configure in YAML or per-pipeline:
 ```yaml
 psnr_min: 25.0      # Higher = stricter
 ssim_min: 0.85      # Range [0, 1], higher = stricter
+require_audio: true      # Require and compare the output audio stream
+audio_cosine_min: 0.95  # Higher = stricter
 pixel_diff_max: 0.02 # Range [0, 1], lower = stricter
 ```
 

diff --git a/docs/zh/parallel.md b/docs/zh/parallel.md
@@ -456,9 +456,8 @@ stage 包含大量 CPU 计算时才需要提高该值；它不会修改父进程
 # 设备数必须等于各并行度乘积
 world_size = dp * cfg * sp_ring * sp_ulysses * pp * tp
 
-# SP 和 TP 不能同时启用
-if sp_degree > 1 and tp_degree > 1:
-    raise ValueError("SP and TP are mutually exclusive")
+# SP 和 TP 可以作为独立的 mesh 维度组合。
+# 所选 pipeline 必须实现并校验请求的组合。
 ```
 
 ## 使用示例

diff --git a/docs/zh/testing.md b/docs/zh/testing.md
@@ -365,6 +365,10 @@ python examples/run_examples.py --all
 | prompt | str\|null | null | 覆盖生成提示词 |
 | input_image_path | str\|null | null | I2V/编辑 pipeline 的输入图像 |
 | input_video_path | str\|null | null | VSR/续写 pipeline 的输入视频 |
+| target_video_length | float\|null | null | 示例支持时覆盖生成时长 |
+| use_run_with_file | bool | false | 使用标准文件入口，保留音轨等容器级输出 |
+| require_audio | bool | false | 要求输出包含音轨，并比较音频契约和解码波形 |
+| audio_cosine_min | float | 0.95 | 视频音频：解码波形余弦相似度下限 |
 | ppl_config_overrides | dict | {} | 覆盖 PPL_CONFIG 配置 |
 | psnr_min | float | 25.0 | 视频：最低 PSNR 阈值 |
 | ssim_min | float | 0.85 | 视频：最低 SSIM 阈值 |
@@ -407,7 +411,8 @@ work_dirs/example_outputs/
 
 Runner 使用以下指标对比 baseline：
 
-- **视频**：PSNR（峰值信噪比）和 SSIM（结构相似度）
+- **视频**：PSNR（峰值信噪比）和 SSIM（结构相似度）；启用 `require_audio` 时还会校验音轨存在性、采样率、声道数、
+  时长和解码波形余弦相似度
 - **图像**：平均像素差异
 
 #### 指标阈值
@@ -417,6 +422,8 @@ Runner 使用以下指标对比 baseline：
 ```yaml
 psnr_min: 25.0      # 越高越严格
 ssim_min: 0.85      # 范围 [0, 1]，越高越严格
+require_audio: true      # 要求并比较输出音轨
+audio_cosine_min: 0.95  # 越高越严格
 pixel_diff_max: 0.02 # 范围 [0, 1]，越低越严格
 ```
 

diff --git a/examples/data/minimax-h3/fl2va-reference.png b/examples/data/minimax-h3/fl2va-reference.png
diff --git a/examples/data/minimax-h3/provenance.json b/examples/data/minimax-h3/provenance.json
@@ -0,0 +1,24 @@
+{
+  "schema_version": 1,
+  "source": "MiniMax H3 official reproducible request inputs frozen for TeleFuser parity validation",
+  "files": [
+    {
+      "path": "fl2va-reference.png",
+      "bytes": 1263485,
+      "sha256": "352e959e08b886406d3766cee25ebe7bbf679e6d53eeffbd1c607b8dc93e2068",
+      "original_uri": "https://cdn.hailuoai.com/prod/hailuo_demo/testsets/H3_AA_I2VA/gallery/sr_v17_variants_seed42_43_20260724/inputs/4a3a90bf9100_KDmcbkhzYo5sjjxr9FqcVmWVnzb.png"
+    },
+    {
+      "path": "ref2va-reference.mp4",
+      "bytes": 4396484,
+      "sha256": "d6edc20ac82aa980c9cbd0b05b371f60d12dc8d018492e571b42de0086819c41",
+      "original_uri": "https://cdn.hailuoai.com/prod/hailuo_demo/testsets/h3_promo_eval_ref2va/gallery/sr_v2p26_trio_seed42_20260724/inputs/297573323635_00_%E8%A7%86%E9%A2%911_YnyRbxEwio_video_20260525_163755_1927e9d3.mp4"
+    },
+    {
+      "path": "ref2va-voice.mp3",
+      "bytes": 204432,
+      "sha256": "2a48d447b343b38a2c504d5281a0c97c688db0ba87931ed9282c598bd6752454",
+      "original_uri": "https://cdn.hailuoai.com/prod/hailuo_demo/testsets/h3_promo_eval_ref2va/gallery/sr_v2p26_trio_seed42_20260724/inputs/f463d523c5ce_01_%E9%9F%B3%E9%A2%911_RSLcbpzJPo_6%E6%9C%885%E6%97%A5(1).mp3"
+    }
+  ]
+}
diff --git a/examples/data/minimax-h3/ref2va-reference.mp4 b/examples/data/minimax-h3/ref2va-reference.mp4
diff --git a/examples/data/minimax-h3/ref2va-voice.mp3 b/examples/data/minimax-h3/ref2va-voice.mp3
diff --git a/examples/data/minimax-h3/ref2va.json b/examples/data/minimax-h3/ref2va.json
@@ -0,0 +1,25 @@
+{
+  "task": "ref2va",
+  "prompt": "Preserve the subject, composition, motion, and soundtrack from <Video 1>. Use the calm male voice timbre from <Audio 2> for the spoken line: Follow the wind, live free.",
+  "conditions": [
+    {
+      "type": "video",
+      "uri": "ref2va-reference.mp4",
+      "role": "reference"
+    },
+    {
+      "type": "audio",
+      "uri": "ref2va-voice.mp3",
+      "role": "reference"
+    }
+  ],
+  "target": {
+    "short_edge": 768,
+    "aspect_ratio": "auto",
+    "duration_seconds": 5
+  },
+  "seed": 0,
+  "flow_shift": 12.0,
+  "audio_flow_shift": 3.0,
+  "num_inference_steps": 50
+}
diff --git a/examples/example_config.yaml b/examples/example_config.yaml
@@ -101,6 +101,23 @@ pipelines:
     ppl_config_overrides:
       attn_impl: TORCH_SDPA
 
+  # ============================================================
+  # minimax_h3 — Joint Audio-Video Generation
+  # ============================================================
+
+  minimax_h3_t2va_4gpu:
+    script: minimax_h3/minimax_h3_fl2va_h100.py
+    gpu_count: 4
+    output_type: video
+    timeout_seconds: 1800
+    seed: 0
+    resolution: "768p"
+    target_video_length: 5
+    use_run_with_file: true
+    require_audio: true
+    ppl_config_overrides:
+      num_inference_steps: 50
+
   # ============================================================
   # lingbot_world_fast - Image to Video
   # ============================================================