中/EN
Open Source
Open Source

OpenMontage Pitfalls: Defaults to Foreign Models Unusable in China, with Domestic Alternatives (46.5k stars)

OpenMontage (46,548 stars, Python/AGPL-3.0) is an open-source agentic video production system whose defaults depend on foreign cloud models (OpenAI/Veo/Runway/ElevenLabs) blocked in China by VPN, payment, and phone-number barriers; but PROVIDERS.md natively supports Kling/Qwen/Jimeng/Doubao/DeepSeek and local Wan2.1 -- with a full domestic-alternative mapping table and minimal working .env config.

Published August 10, 20269 min read
<!-- openmontage-resource | open-source | OpenMontage Pitfalls: Defaults to Foreign Models Unusable in China, with Domestic Alternatives (46.5k stars) -->

OpenMontage, at 46,548 stars, calls itself "the world's first open-source agentic video production system"--12 production pipelines, 100+ tools, 700+ agent skills. Pair it with an AI coding assistant (Claude Code/Cursor/Copilot) and it runs the full chain from topic to final cut. But developers in China hit a wall on step one: its default provider list is entirely foreign cloud models--OpenAI, Google Veo, Runway, ElevenLabs, fal.ai--blocked by VPN, overseas payment, and overseas phone-number requirements. The good news is docs/PROVIDERS.md natively supports a full set of domestic and local models; swap a group of environment variables and it runs in China. This piece is that swap list. Repo calesthio/OpenMontage, Python, AGPL-3.0, created 2026-03-29, last push 2026-08-03.

Boundaries first: this piece is assembled from OpenMontage's GitHub README and docs/PROVIDERS.md, with facts as of 2026-08-10. Star counts move in real time--treat GitHub as the source of truth. This is a standalone deep dive on OpenMontage, complementary to but independent of this site's 5 open-source AI video production tools compared. The comparison covers "which of five routes to pick"; this one drills into OpenMontage's China usability. The domestic alternatives below are drawn from PROVIDERS.md's listed support, not independently stress-tested by this site--check the official docs for interface stability.

1. What OpenMontage is: an agent-driven video production system

First, understand the route OpenMontage bets on. It's not another text-to-video model, nor a preset-script automation pipeline--it hands "video production" to an AI agent to orchestrate. You configure an AI coding assistant (Claude Code, Cursor, Copilot, Windsurf, Codex all work); the agent reads your production brief, dispatches OpenMontage's Python tools on demand, and walks through scripting, assets, editing, subtitles, music, and assembly.

Three mechanisms hold up this paradigm. First, 12 production pipelines cover documentaries, explainer videos, creative shorts, and other genres--each pipeline is a template combining a set of tools and skills. Second, 100+ tools plus 700+ agent skills and production-knowledge files let the agent match tools on demand rather than via hardcoded logic. Third, and most critical: provider selection isn't hardcoded. The agent scores candidate providers across seven dimensions--quality, speed, cost, availability, quota, feature fit, historical performance--and picks the best fit for the current task. This means the same pipeline can swap its entire underlying model stack by changing provider config, which is exactly why domestic substitution is possible.

A quality gate backs it up: ffprobe validation, frame sampling, audio-level checks, subtitle checks, and delivery-promise verification. The agent doesn't hand off directly; it runs automated QA first, catching black frames, missing audio, and misaligned subtitles before delivery. This site compared OpenMontage alongside MoneyPrinterTurbo, video-use, Palmier, and Remotion in 5 open-source AI video production tools compared--no need to repeat the route differences here.

2. The core pitfall: defaults depend on foreign cloud models, unusable in China

This is the wall developers in China hit first. OpenMontage's default provider list is a full stack of foreign cloud models, and the out-of-box config basically won't run in China. Lay the list out.

Default foreign-model dependencies:

  • OpenAI: GPT Image 2 for images + OpenAI TTS for narration
  • Google: Gemini/Veo for video + Imagen for images + Lyria for music + Chirp3-HD for narration
  • Runway Gen-4: video generation, $12/mo+, officially the highest quality
  • ElevenLabs: TTS + music + SFX, 10K characters/mo free tier
  • HeyGen: avatar video, also a gateway to VEO/Sora/Runway/Kling
  • fal.ai: unified gateway, one key calls FLUX for images + Kling/Veo/MiniMax for video + Recraft for images
  • Recraft: image generation; Sora (via HeyGen); Grok: images + video

Why it's unusable in China--three barriers stack. First, network: these API endpoints are all overseas, unreliable or unreachable on direct connections from China; you need a proxy or dedicated line, and a network hiccup mid-pipeline breaks the whole chain. Second, payment: OpenAI, Runway, ElevenLabs, fal.ai all require overseas credit cards or PayPal, and Chinese cards are routinely rejected. Third, accounts: some services require an overseas phone number for signup verification, and the Google stack is outright unavailable in China. Any one of the three blocks the default config.

Many people give up at this point and conclude OpenMontage "can't be used in China." That's a misunderstanding.

3. The good news: PROVIDERS.md natively supports domestic + local models

The key is docs/PROVIDERS.md. This document lists every provider OpenMontage supports--the foreign set is just the default, not the whole picture. Domestic and local models are first-class citizens in the same support list, just not enabled by default. Here they are.

Domestic/China-usable models (equally supported in PROVIDERS.md):

  • Kling official direct (Kuaishou): video/image/TTS/avatar/lip-sync, endpoint api-singapore.klingai.com, no third-party gateway
  • MiniMax (Hailuo): video generation, via fal.ai or Atlas Cloud
  • Jimeng 3.0 Pro (ByteDance Volcano Engine): text-to-video/image-to-video, req_key jimeng_ti2v_v30_pro, IAM V4 signing
  • Doubao Speech TTS (Volcano Engine): narration synthesis, strong Mandarin
  • Alibaba DashScope: Qwen-Image for images + Qwen-TTS for narration + Qwen-ASR word-level timestamps, smoothest for Chinese production
  • Bloome: multi-agent collaboration layer, natively supports Claude/ChatGPT/DeepSeek--use DeepSeek for LLM orchestration
  • Atlas Cloud: 300+ models behind one unified API, including Seedream/Nano Banana/GPT Image for images + Kling/Seedance/Hailuo for video

Local models (zero cost, offline):

  • VIDEO_GEN_LOCAL_MODEL supports wan2.1-1.3b / wan2.1-14b (Alibaba), hunyuan-1.5 (Tencent), ltx2-local, cogvideo-5b (Zhipu)
  • Zero API key is possible: make setup configures a local/free setup

The point: these aren't community hacks. They're officially listed in OpenMontage's PROVIDERS.md. What domestic developers need to do isn't to modify the project--it's to swap a group of environment variables to replace the default foreign providers with domestic or local ones.

4. The domestic-alternative mapping table

This is the core value of this piece. The "foreign model -> domestic alternative" mapping, laid out at once.

CapabilityOpenMontage default (foreign)Domestic alternative (equally supported)Notes
LLM/orchestrationClaude/GPTDeepSeek/Qwen (via Bloome)Bloome natively supports DeepSeek
ImagesFLUX/Recraft/GPT ImageQwen-Image (DashScope)/JimengDashScope has stronger Chinese-prompt understanding
VideoVeo/Runway/SoraKling (Kuaishou)/MiniMax/Jimeng/local WanKling official direct, high quality
Narration TTSElevenLabs/OpenAI TTSDoubao Speech/Qwen-TTSDoubao has strong Mandarin
MusicSuno/ElevenLabsLocal/royalty-free PixabayOpenMontage auto-finds royalty-free assets
Fully local, zero cost—Wan2.1/Hunyuan/CogVideo + WhisperXGPU required

A few points. First, video is the biggest item: the default Veo/Runway/Sora are all unusable in China, but Kling's official direct connection delivers solid quality, and Jimeng 3.0 Pro and local Wan2.1 are mature alternatives covering both text-to-video and image-to-video. Second, don't fixate on ElevenLabs for narration: Doubao Speech's Mandarin naturalness is first-tier among Chinese TTS, and Qwen-TTS is equally capable--plus Qwen-ASR gives word-level timestamps, saving a big step on subtitle alignment. Third, for images, Qwen-Image's understanding of Chinese prompts is noticeably better than the FLUX family, so visuals land more on-topic for Chinese content. Fourth, music is the easiest: OpenMontage itself auto-finds royalty-free assets (Pixabay etc.), so you don't have to bind to Suno/ElevenLabs. Fifth, the fully-local zero-cost route genuinely exists: Wan2.1-14b + WhisperX can run video generation and subtitles offline, at the cost of needing a GPU.

5. Minimal China-usable config

Here's a .env example covering the three minimum-necessary capabilities--video, images, narration--with API keys redacted.

bash
# Video: Kling official direct (Kuaishou, no third-party gateway)
KLING_API_KEY=kl-xxxxxxxxxxxxxxxx
KLING_API_ENDPOINT=https://api-singapore.klingai.com

# Images + narration: Alibaba DashScope (Qwen, smoothest for Chinese)
DASHSCOPE_API_KEY=sk-xxxxxxxxxxxxxxxx

# Narration alternative: Doubao Speech (Volcano Engine, strong Mandarin)
DOUBAO_SPEECH_API_KEY=xxxxxxxxxxxxxxxx
DOUBAO_SPEECH_APP_ID=xxxxxxxxxxxxxxxx

# LLM orchestration: DeepSeek via Bloome (replaces Claude/GPT)
BLOOME_LLM_PROVIDER=deepseek
DEEPSEEK_API_KEY=sk-xxxxxxxxxxxxxxxx

# Fully local video generation (zero API key, GPU required; pick one of cloud or local)
# VIDEO_GEN_LOCAL_MODEL=wan2.1-14b

After configuring, run make setup and OpenMontage validates the config and pulls dependencies. When the agent runs a task it scores across seven dimensions and picks the best fit from your configured provider pool. The idea behind this config: the cloud path uses Kling for video, Qwen for images and narration, DeepSeek for orchestration--a fully Chinese, fully domestically-reachable chain; if you have a machine, the local Wan2.1 path is zero API cost. This site's zero-cost multimodal workflow SOP breaks down more local + free combinations.

6. Getting started and use cases

Three steps to start. First, clone the repo and run make setup; it walks you through choosing a provider mix, with local/free/cloud presets. Second, configure .env with at least video, image, and narration API keys (or enable local models). Third, configure an AI coding assistant to drive it--the README explicitly supports Claude Code, Cursor, Copilot, Windsurf, and Codex; whichever you use, the agent plugs into OpenMontage's toolchain via skill files.

Who it's for. First, producers working on documentaries, explainers, and creative shorts--genres needing complex orchestration where OpenMontage's agent paradigm beats preset pipelines. Second, developers already living in an AI-coding-assistant workflow who want to fold video production in. Third, solo teams that need to batch-produce Chinese video content without getting blocked by foreign models--with domestic providers configured, the whole chain is directly reachable in China. This site's AI agent tool-calling SOP and free multimodal API comparison help you stabilize the agent-dispatch and provider-selection steps.

7. Limitations and pitfalls

Four points to state clearly. First, AGPL-3.0 commercial risk. OpenMontage is AGPL-3.0, not MIT--AGPL's network clause means if you turn it into an online service offered to others, you must open-source your modifications. Read the full license before commercial use; don't treat it as MIT.

Second, high stars don't mean stable interfaces. 46,548 stars and 5,789 forks look impressive, but the project was created 2026-03-29--less than six months old--and provider interfaces, tool signatures, and pipeline structures are still moving fast. Check the changelog before upgrading; don't blindly git pull.

Third, local models need a GPU. Local video models like Wan2.1-14b and Hunyuan-1.5 are VRAM-hungry; consumer GPUs can't run them or are too slow to be practical--14b models typically need 24GB+. Without a card, go cloud: Kling/Qwen pay-per-use is more realistic.

Fourth, DashScope is not OpenAI-compatible. Alibaba DashScope's API isn't OpenAI-compatible; it requires the native SDK. The provider adapter in PROVIDERS.md handles this difference, but if you extend or modify providers yourself, don't assume every provider speaks OpenAI format.

8. FAQ

Q: Can OpenMontage really be used in China? Doesn't the default config require a VPN for everything? A: The default config does depend on foreign cloud models and won't run directly in China. But docs/PROVIDERS.md natively supports Kling, Alibaba DashScope, Jimeng, Doubao, DeepSeek, and other domestic models, plus local models like Wan2.1/Hunyuan/CogVideo. Swap a group of environment variables to replace the providers with domestic or local ones, and the whole chain is directly reachable in China without a VPN.

Q: Can AGPL-3.0 be used commercially? A: AGPL-3.0 is not MIT. Its network clause requires that if you turn OpenMontage into an online service offered to others, you must open-source your modifications under AGPL. Pure internal or personal use isn't subject to this. Read the full license before commercial use; for closed-source commercial products, seek legal advice.

Q: Can I run local models without a GPU? A: Local video models like Wan2.1-14b and Hunyuan-1.5 have high VRAM requirements (typically 24GB+); consumer GPUs can't run them or are too slow to be practical. Without a GPU, go cloud: Kling for video, Qwen-Image for images, Doubao/Qwen TTS for narration--pay per use, zero hardware cost.

Q: What's the difference between OpenMontage and MoneyPrinterTurbo? A: MoneyPrinterTurbo bets on "one-click batch"--a preset script runs the full "script -> assets -> music -> assembly" pipeline automatically, suited for social-media content matrices. OpenMontage bets on "agent-driven"--the agent dispatches tools on demand and picks providers across seven dimensions, suited for documentaries and explainers needing complex orchestration. This site's 5 open-source AI video production tools compared has the full breakdown.

Q: With this many stars, is the project mature? A: 46,548 stars is impressive, but the project was created 2026-03-29--less than six months old--and provider interfaces and pipeline structures are still moving fast. Run it through your own genre before production integration; don't bolt it on blindly.

References

This article is AI-assisted and human-edited. Last updated: 2026-08-10

FAQ

Can OpenMontage really be used in China? Doesn't the default config require VPN?
The default config does depend on foreign cloud models and won't run directly in China. But docs/PROVIDERS.md natively supports Chinese models (Kling, Alibaba DashScope, Jimeng, Doubao, DeepSeek) and local models (Wan2.1/Hunyuan/CogVideo). Swap providers via env vars and the whole chain works in China without VPN.
Can AGPL-3.0 be used commercially?
AGPL-3.0 isn't MIT -- its network clause requires that if you modify OpenMontage into an online service, you must open-source your modifications under AGPL. Pure internal use isn't affected. Read the full license before commercial use; seek legal advice for closed-source products.
Can I run local models without a GPU?
Local video models like Wan2.1-14b and Hunyuan-1.5 need high VRAM (typically 24GB+); consumer GPUs can't run them or are impractically slow. Without a GPU, use cloud: Kling official for video, Qwen-Image for images, Doubao/Qwen TTS for narration -- pay-as-you-go, zero hardware cost.
How does OpenMontage differ from MoneyPrinterTurbo?
MoneyPrinterTurbo bets on "one-click batch" -- preset scripts running a fully automated pipeline, good for self-media matrix content. OpenMontage bets on "agent-driven" -- the agent schedules tools on demand and picks providers across 7 dimensions, good for documentaries and explainer topics needing complex orchestration. Our comparison review has the full breakdown.
So many stars -- is the project mature?
46,548 stars is impressive, but the project was created 2026-03-29 -- less than six months old -- and provider interfaces and pipeline structure are still changing fast. Run your own topic through it before production; don't go all-in.

Related

Open Source

M3.1 Flash Weights Are Out, GitHub Comes Up Empty

On October 23, 2026, MiniMax open-sourced M3.1 Flash weights: a native multimodal MoE with 428B total and 23B active parameters, a 1M-token context, MSA model-wide sparse attention (arXiv:2606.13392), and DSpark speculative decoding replacing EAGLE-style MTP. The weights live in two places: Hugging Face at MiniMaxAI/MiniMax-M3.1-Flash (public, not gated) and ModelScope at MiniMax/MiniMax-M3.1-Flash (page and API both verified 200 on release day); GitHub has no repo named after the model (github.com/MiniMax-AI/MiniMax-M3.1-Flash returns 404; the org's 37 repos include only the kernel library MiniMax-AI/MSA). The license, per the LICENSE file itself: MiniMax Community License - free for non-commercial use; commercial use requires visible attribution Built with MiniMax M3.1-Flash; annual revenue above USD 20 million requires prior written approval ([email protected]), with commercial use defined broadly to cover API resale and fine-tuned deployment. Three throughput calibers coexist: over 100 trillion tokens in the first 72 hours (official announcement via media), 200T over the 13 anonymous days (official, Oct 19), and 56.7T on OpenRouter during anonymity (via Zhidx, single platform); officially, agentic usage cost fell 70 percent versus the predecessor. Ecosystem: ten chip vendors including Huawei Ascend, T-Head, and Kunlunxin plus vLLM, SGLang, and FlagOS; Kunlunxin completed Day 0 while most others are still pushing first-day integration. The only official M3-to-M3.1 Flash note: keep the core MoE, change three things - sparse attention in the first three layers, QAT, and DSpark. Division of labor: the Oct 21 unmasking piece covers the anonymous saga and pricing; this one covers weights landing, repos, and the license.

Oct 24, 20269 min read
Open Source

Mistral Large 4: Model Is Live, Weights and License Are Not

Mistral announced Large 4 (nicknamed Le Chonk) in public preview on October 6, 2026, with the API open the same day: 1.05T total parameters activating 52B per token (the HF repo name Mistral-Large-4.0-1T05-A52B uses this figure; 49B excludes the embedding and output layers), a 1.6B vision encoder, training across 160+ languages, and 3,800 Grace Blackwell GPUs in Mistral's own European datacenters; the API context is listed at 1M while OpenRouter and AA test at 512K. Pricing is $1.36/$0.14/$4.18 per million tokens list, with an undated half-price column of $0.68/$0.07/$2.09 on the same docs page. But as of October 23 the open weights have not landed: the HF repo is a countdown page with zero downloadable files, the license field reads coming soon, the docs render only the word Open, and Artificial Analysis records isOpenWeights as false, classifying it as proprietary. GTC Berlin on October 21 reconfirmed the end-of-October weight release via the Mistral-NVIDIA industrial alliance. Self-reported benchmarks: DeepSWE v1.1 61.7 percent, Cybench 93 percent, and an 82 percent vulnerability reproduce-and-patch score where third-party data shows Opus 5.5 refuses 98.5 percent and GPT-6 Astra 100 percent of the same tasks; AutomationBench 59.9 despite the model card's own chart giving GLM-5.3 a higher 62.2. The third-party picture differs: AA Intelligence Index 38 (Opus 5.5 at 58), Omniscience at -5.3 meaning bare recall fabricates more than it admits, and 33rd of 45 on the vals index, roughly 9-12 months behind the frontier. Community reaction splits between 'about Kimi K3 level' performance doubts, an OpenRouter price roughly 10x DeepSeek's, and genuine EU data-residency defense. Running it locally takes an estimated 400GB+ of fast memory - today only a 512GB Mac Studio M5 Ultra theoretically qualifies.

Oct 23, 20269 min read
Open Source

DeepInfer Open-Sourced: 155T Tokens/Day, Engine or Factory?

On October 20, 2026 DeepSeek announced full open sourcing of DeepInfer, its production-grade inference framework - the first time the official production engine is available instead of open-source parts plus guesswork. The repository deepseek-ai/DeepInfer ships under Apache-2.0 and measured 17,902 stars, 1,836 forks, and 188 open issues on launch day, while the companion token codec deepseek-recipe is MIT-licensed; the two repositories carry different licenses. Every performance figure below is vendor-reported and has no third-party reproduction yet: a single cluster serves 155 trillion tokens per day (about 38 dollars per card per hour at off-peak pricing), first-token latency of 400 milliseconds, 230 tokens per second generation, 96 percent prefix-cache hit rate, and Prefill throughput per card rising from 60,000 to 1.5 million tokens per second when caching counts. SWA Bounded Replay pins the recompute of cache misses to the last 128 tokens, which nearly halves total Prefill compute and makes cost independent of context length; DSpark speculative decoding drafts up to 5 tokens, adds another 6 percent to live throughput, and yields 14K tokens per card per second; MoE expert replicas reshuffle every 10,000 steps so the busiest card runs only 9 percent above average load. Hardware covers both NVIDIA SM100 and Huawei Ascend 950 NPUs, and setup means a Python 3.12 virtual environment, the DEEPINFER_ACCELERATOR variable, and a source build that pulls DeepSeek-V4.1-Flash-1017 weights by default. The boundaries matter: DeepSeek states that Prefill and Decode disaggregation, data-parallel load balancing, and the KVCache management layer are still not fully open because they couple too deeply to internal infrastructure - this is the engine, not the whole factory - and no official benchmark against vLLM, SGLang, or TensorRT-LLM exists. In production, 8-card servers handle Prefill while 64 to 128-card supernodes handle Decode, with KVCache kept beyond 72 hours to avoid recomputation.

Oct 22, 202610 min read