The Samsar agent now supports full suite of Alibaba cloud models via composable model adapters. Choose Happy-Horse 1.1 , WAN 2.7 and QWEN 3.7 inference engine to create T2V in 1-shot. We also present Quinn - A2A connector for agent calls with default Alibaba Cloud model adapters.
- Update- I had earlier said QWEN 3.8 inference is available, as of today, QWEN 3.8 inference is only available via Token plan which is not suitable for production agentic usage. I have reverted the QWEN adapter back to 3.7 for now until it is available on the standard API plan. I have shared some internal tests using QWEN 3.8 below.
Installation tutorial -
The Integration
We already had SoTA inference models, however the recent Alibaba hackathon challenge gave the opportunity to add something we've been looking forward to for a long time - SoTA inference model from the East, served via native adapters.
Get your Alibaba Cloud API key and hosted endpoint, the container now supports QWEN 3.8 for Token Plan API keys. To run with Alibaba simply start the docker container
Optionally deploy the entire stack into your Alibaba cloud account on a 32 GB VM
Three Ways to Connect-
The Samsar docker container is adapter configurable, and you can specify multiple adapters or same model or use a deployed samsar-js key as universal fallback.
- Connect via native Alibaba Cloud config.
During the container setup you will be asked to enter adapter config from the various supported model adapters.
- Connect via supported Provider adapters
You can also use the full-suite of Alibaba cloud models by adding adapters for supported models.
If you prefer using OpenRouter, add OpenRouter API key to enable QWEN 3.7 inference. (From our experiments, inference is a bit slower than native adapter integration)
You can also add Fal API key to enable Happy Horse 1.1, WAN 2.7 Pro with excellent inference speed and render quality.
- Connect via Samsar-js universal fallback
You can also create an account and an API key at app.samsar.one, (simply sign-up for an account, get enough credits for render and create an API key) then add the API key in the universal fallback section.
You can add multiple adapters for the same config. If multiple adapters are added for the same config, the system prioritizes native adapters for the model, then third-party adapters and samsar-js fallback in that order respectively.
To enable the full-set of features, user will need to add a fallback config, like Fal or Samsar-js.
Architecture-
The basic architecture is the same as the the generic model agnostic agent. For this hackathon I made improvements and enhancements such as providing native Alibaba Cloud adapters.
Nuances & Challenges
Unlike other flagship models, QWEN 3.7 Max is text only, vision is supported in QWEN 3.7 Plus.
WAN 2.7 while creating great images doesn't render texts accurately. For rendering accurate texts, the Qwen image family models work better.
HappyHorse 1.1 often adds a bluish tinge to the teeth when there is a lot of blue in the background which was interesting behavior.
Some learning curve in the Cloud model studio regarding the various types of API plans and usage limits of each.
We found that QWEN 3.7 Max inference is fast and reliable via Alibaba native cloud, we also support this inference model via Openrouter adapter however large requests such as narrative and theme generation seem to often time-out. To better handle this on the router side and since we intend to provide full-support to all adapters and models, we switched QWEN inference via OpenRouter to QWEN 3.7 Max to Plus and added additional error handling (malformed JSON response, server timeouts and 429 Rate limits) , while still routing to QWEN 3.7 Max for inference when routed via native Alibaba cloud credentials.
Another interesting I have observed in Happy Horse I2V (and indeed other SoTA I2V models) if the image has bluish tint, it will shade the characters face, and teeth blue during speaker scenes.
Some demo 1-shot prompt to videos
All of these demoes are from QWEN 3.7 model, we will be publishing a few demoes via QWEN 3.8 shortly here.
Some demoes we created while building this integration with QWEN 3.7 inference and various media model combinations such as Happy Horse 1.1 and Cosmos 3 Super for video render stage, as well as WAN 2.7, Seedream 3 for the image generation stage.
Dune theme renders-
Every benchmark needs a Dune theme demo, so we created a few of ours. All of these used QWEN 3.7 model for inference. For image and video various diffrent models were used.
1-shot render settings - QWEN 3.7 Inference, WAN 2.7 Image, Cosmos 3 Super Video
Some demoes from the upcoming QWEN 3.8 version from internal tests. (Not yet setup in public github, since it is as of today Jul 21, I will enable it in public Github once it is available in the Pay-as-you-go API plan.
Bel Riose-The Empire’s Last Great General.
QWEN 3.8 Preview Inference. Cosmos 3 + Seedream
QWEN 3.8 Inference. Happy Horse 1.1 + Seedream
Try it here-
Full stack generative video cloud in a single docker container.
samsarone
Check out the gallery for some 1-shot T2V renders here
Check out the code for the Quinn A2A connector, which allows agents to autonomously authenticate, purchase credits and submit T2V requests using Alibaba model defaults.
samsarone