Qwen-Image and FLUX ran on ModelScope API-Inference without being in the model list

I’ve been using ModelScope’s API-Inference (OpenAI-compatible for LLMs) for proofreading and agent experiments since testing function calling on Qwen3.7 Plus.
In September, my account stopped showing a monthly call quota and showed a Magicube balance instead.
As a test I sent one request to the Qwen3.8 Max I normally use.
The balance dropped by 0.5.

No idea what the other models cost, so I measured what each call takes off the balance.
While at it, I hit everything the endpoint serves beyond the chat models in the model list.

Test setup

Item	Details
Client	Apple M1 Max / 64GB / macOS 26.5
Endpoint	ModelScope API-Inference (OpenAI-compatible for LLMs), called with curl
Auth	ModelScope access token
Cost check	Magicube balance read by eye in the browser
Date	September 4, 2026

All inference runs on the API side, so the local machine is irrelevant, and I checked the balance in the browser after each call and wrote down the difference.
Response headers did not carry the balance.

How Magicubes work

The pricing and grant rules are on the official API-Inference Usage Limits and Magicube Rewards Program pages.

Item	Details
Cost per call	Lightweight models about 0.5, standard about 1, flagship about 2
Daily grant	200 for logging in, plus 50 if an Alibaba Cloud account is linked
Short-term Magicubes	Expire 24 hours after they are granted
Long-term Magicubes	Expire after 90 days. One-time grants for email verification, profile setup and so on
Failures	Tasks that fail on a system error are refunded in full
Scope	LLMs, multimodal models, and text-to-image
Prerequisites	A profile with a verified email. API use also needs a linked Alibaba Cloud account with real-name verification

The daily grants are good for that day only and do not roll over.

The model list and what actually runs

/v1/models returned 49 entries.
DeepSeek-V4-Pro, GLM-5.2, MiniMax-M3, the Qwen3.5 and Qwen3.8 sizes, the three Qwen3-Embedding sizes, and so on.
The only image model listed is Qwen/Qwen-Image-Edit.

Of the listed ones, DeepSeek-V4-Pro, GLM-5.2, MiniMax-M1-80k, Qwen3.5-397B-A17B and the three Qwen3-Embedding sizes ran as listed.
But Qwen/Qwen-Image, which is not listed, ran fine on the image endpoint.
Meanwhile MiniMax/MiniMax-M3, which is listed, was rejected on chat with “no provider supported”.

According to the Usage Limits page, supported models show a usage entry point and code examples on the right side of their model page.
Opening the Qwen-Image-2512 page, it is there at the top right.

The Wan2.1-T2V-1.3B, Qwen3-ASR-1.7B and Qwen3-TTS pages I opened have no such panel.

Image generation

The documented flow is POST /v1/images/generations with a model name and prompt, an async-mode header to get a task ID back, then polling GET /v1/tasks/<id> at intervals until the result is ready.

curl -H "Authorization: Bearer $MODELSCOPE_TOKEN" \
  -H "Content-Type: application/json" \
  -H "X-ModelScope-Async-Mode: true" \
  "$BASE_URL/images/generations" \
  -d '{"model":"Qwen/Qwen-Image","prompt":"a red apple on a white table"}'
# => {"task_status":"SUCCEED","task_id":"..."}

curl -H "Authorization: Bearer $MODELSCOPE_TOKEN" \
  -H "X-ModelScope-Task-Type: image_generation" \
  "$BASE_URL/tasks/<task_id>"
# => {"task_status":"SUCCEED","output_images":["https://...png"]}

The four Qwen-Image generation and editing models

The ID Qwen/Qwen-Image maps to the original repository from August 2025 on ModelScope.
Newer versions have their own IDs, so I sent 2512, plus the original and 2511 editing models, to the same endpoint.
For editing, the input was the red apple generated by the original Qwen-Image, with the instruction to turn it green.

Model ID	Instruction	Result
Qwen/Qwen-Image	Red apple on a white table	Clean photographic output
Qwen/Qwen-Image-2512	Same	Worked. Closer framing on the apple
Qwen/Qwen-Image-Edit	Make the apple green	Composition kept, apple stayed red, image slightly smeared
Qwen/Qwen-Image-Edit-2511	Make the apple green	Apple turned green, composition kept

The images follow the table order: original Qwen-Image, Qwen-Image-2512, original Edit, Edit-2511.

The original Edit used the input image but ignored the instruction.
One attempt only, but the gap to 2511 showed in a single image.

Sending the Edit model without an image_url did not error and generated an image from the prompt alone.

Image models other than Qwen

Models I’ve covered in earlier posts are also on ModelScope, so I ran the same prompt through them. The prompt was a standing anime-style girl with a red side ponytail and a white shirt.

Model ID	Result
Tongyi-MAI/Z-Image-Turbo	Worked. Matched the prompt
Tongyi-MAI/Z-Image	Worked. Painterly, side view
krea-community/Krea-2-Turbo	Worked. Matched the prompt
MusePublic/FLUX.1-Kontext-Dev	Worked. Ponytail sits toward the back
MAILAND/majicflus_v1 (the example in the official docs)	Worked. Photorealistic
ModelE/yumemono-illustrious-v-1-0 (a LoRA for Illustrious-XL)	“submit failed with code: 40212”. No charge. No image

The images follow the table order: Z-Image-Turbo, Z-Image, Krea-2-Turbo, FLUX.1-Kontext-Dev, majicflus_v1.

The rejected one turned out, per its model metadata, to be a LoRA for Illustrious-XL, and Illustrious-XL is built on the SDXL architecture.
The docs’ resolution table has an “SD series” entry, but this one did not go through.
Krea-2-Turbo does not exist as krea/Krea-2-Turbo. The correct ID is krea-community/Krea-2-Turbo.

A LoRA repo ID passed straight in as the model

Searching for “Anima” turned up four community repositories I checked, and all four are tagged LoRA in their metadata. Their bases were Krea-2-Turbo, Qwen-Image-Edit-2511 or Z-Image. Anima itself did not appear in the search results.

Putting the repo ID from the search results straight into model went through and returned an image.
The docs describe a separate loras parameter that takes up to six ModelScope LoRAs with weights summing to 1.0.
Same prompt as the previous section, no seed set.

Repo ID	Base	Result
Martis/Anima	Z-Image	Same painterly look as Z-Image alone. Seated pose, face broken
Devilworld/anima-krea2	Krea-2-Turbo	Almost the same image as Krea-2-Turbo alone
Creolim/WenM-Anima	Krea-2-Turbo	Also almost the same image as Krea-2-Turbo alone
Blueee/Anima.f	Qwen-Image-Edit-2511	Anime style, side view, white background

The images follow the table order.

The two Krea-2-Turbo LoRAs produced almost the same image as the base alone.

All of the images so far came back as 760x1280 PNGs hosted on Aliyun OSS, and each task cost 1 coin.

Embeddings

/v1/embeddings exists.
The first call failed with encoding_format must be 'float' or 'base64', got ''. Adding encoding_format as the error asked made it go through.

curl -H "Authorization: Bearer $MODELSCOPE_TOKEN" \
  -H "Content-Type: application/json" \
  "$BASE_URL/embeddings" \
  -d '{"model":"Qwen/Qwen3-Embedding-0.6B","input":"apple","encoding_format":"float"}'

Model ID	Result
Qwen/Qwen3-Embedding-0.6B	Worked. 1024 dimensions
Qwen/Qwen3-Embedding-4B	Worked. 2560 dimensions
Qwen/Qwen3-Embedding-8B	Worked. 4096 dimensions
BAAI/bge-m3, bge-large	”Invalid model id”
jinaai/jina-embeddings-v3, v4	”Invalid model id”
intfloat/multilingual-e5-large	”Invalid model id”
Alibaba-NLP/gte, iic/gte	”Invalid model id”
nomic-embed-text-v1.5, embeddinggemma-300m, all-MiniLM-L6-v2	”Invalid model id”
Qwen/Qwen3-VL-Embedding-2B	”Invalid model id”

Of the IDs I tried, the three in the list went through.
For rerankers (models that re-order search results), /v1/rerank, /v1/reranking and /v1/rerankers were all 404.
Each call cost 1 coin. The 400s for invalid IDs cost nothing.

Video

/v1/videos/generations exists as a path.
It returned an application-level JSON error rather than a 404.

Model ID	Response
Wan-AI/Wan2.2-T2V-A14B	”Invalid model provider”
Wan-AI/Wan2.2-TI2V-5B, Wan2.5, Wan2.6	”Invalid model provider”
Wan-AI/Wan2.1-T2V-1.3B	”X-ModelScope-DataInspection is empty”

Only the 1.3B model got a different response, demanding an X-ModelScope-DataInspection header.
The first try was true. The error that came back said a bool could not be unmarshalled into a struct, so the second try was an empty object, {}.

Header value	Response
true	”cannot unmarshal bool into Go value of type parameter.MuseCsiConfig”
{}	”failed to call provider API”

Neither call produced a video, and the docs scope API-Inference to “LLMs, multimodal, and text-to-image” with no mention of video.
Video generation is on the Civision side. The Wan2.1-T2V-1.3B model page had a “Generate Now” button instead of an API panel.

Speech

I tried /v1/audio/transcriptions, /v1/audio/speech, /v1/audio/generations, /v1/asr/generations, /v1/tts/generations, /v1/speech/recognition and /v1/audio/synthesis. All 404.
These are plain nginx 404s, so the routes don’t exist.

I also tried the OpenAI-compatible way of attaching audio to chat completions (input_audio).
A Japanese WAV made with macOS speech synthesis went to Qwen3-ASR-1.7B, SenseVoiceSmall and Qwen3-Omni. All three came back “has no provider supported”.
Requesting audio output from Qwen3-TTS and CosyVoice2 via modalities failed the same way.
The ASR and TTS model pages have no API-Inference panel either.

Sending the same audio to the discounted Qwen3.8 Max I normally use returned a 200, but choices was null with no body, and it still cost 0.5.

Large LLMs

I also hit the other vendors’ flagships in the list once each.
With “reply OK only” and max_tokens=8, some models spent it all on thinking and returned an empty body, but the call itself went through.

Model ID	Result	Cost
deepseek-ai/DeepSeek-V4-Pro	Worked	2
zai-org/GLM-5.2	Worked	2
Qwen/Qwen3.5-397B-A17B	Worked	2
MiniMax/MiniMax-M1-80k	Worked	1
MiniMax/MiniMax-M3	Listed, but “no provider supported”	0
moonshotai/Kimi-K3, K2.6, K2.7-Code	”no provider supported”	0

The discounted models under the Qwen Ambassador namespace that I normally use cost 0.5 for Qwen3.8 Max and 0.2 for Qwen3.7 Plus on the same call.

Kimi-K3 has a repository on ModelScope and was even updated on September 3. In the file list I fetched, an .eval_results folder was the newest entry, with a commit message adding an evaluation result.
In the top 20 entries the weight shards still dated from July 28.

Measured costs

Call	Cost
Public flagship LLMs (DeepSeek-V4-Pro, GLM-5.2, Qwen3.5-397B)	2
MiniMax-M1-80k	1
Discounted models under the Qwen Ambassador namespace (Qwen3.8 Max / Qwen3.7 Plus)	0.5 / 0.2
Image generation or editing (per task)	1
Embeddings (per call)	1
400s, 404s, and the video “failed to call provider API”	0
Empty response (a 200 with null choices after the audio was ignored)	0.5 (the discounted model’s normal rate)

The model metadata for Qwen/Qwen-Image, incidentally, shows a creation date of August 2, 2025 on modelscope.cn and September 5, 2025 on modelscope.ai, a month apart for the same ID.