GPT-5 mini, Devstral, Qwen3-VL and more: nine models added
On September 20, 2026 we added nine models that developers already use widely but could not yet pick on FastMetal. We added them in the order people had been searching for them, and Devstral 2512 made it the first Mistral model in the catalog.
One of the nine, Llama 4 Maverick, is no longer offered: we took it out of service on November 9, 2026. The other eight are still served. See the model catalog for what is available today.
The models
| Model | Model id | Good for |
|---|---|---|
| GPT-5 mini | gpt-5-mini | The small model of the current GPT generation, 400K context |
| GPT-4.1 mini | gpt-4.1-mini | A mid-size model with a 1M-token context |
| GPT-5.4 nano | gpt-5.4-nano | The smallest current GPT, for high-volume routine jobs |
| Llama 3.3 70B Instruct | llama-3.3-70b-instruct | The standard 70B open-weight model |
| Qwen3-VL 235B-A22B | qwen3-vl-235b-a22b-instruct | A large vision model with a 262K context |
| Gemma 4 26B-A4B | gemma-4-26b-a4b-it | A light multimodal model from Google |
| gpt-oss-20B | gpt-oss-20b | The smaller sibling of gpt-oss-120B, at the low end of the price range |
| Devstral 2512 | devstral-2512 | Coding, and our first Mistral model |
| Llama 4 Maverick | (no longer served) | An open MoE model with a 1M-token context and image input |
What we tested before listing them
Every model got real requests before it went into the catalog.
- Tool calling: all nine returned a function call when given a tool.
- Image input: the six vision models (GPT-5 mini, GPT-4.1 mini, GPT-5.4 nano, Qwen3-VL, Gemma 4 26B and Llama 4 Maverick) each read a Japanese invoice image. All six pulled out the invoice number, the total and the due date correctly.
- Data handling: at launch we confirmed that each model answered with providers that train on or retain request data excluded. Several of these models run on different hosts now, so check the data policy on each model page for the current state.
Calling them
Change the model value. The endpoint, the API key and the balance are the same for every model:
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-5-mini",
"messages": [{"role": "user", "content": "Explain the time complexity of this function."}]
}'
FAQ
How much do these models cost?
Each is billed per million input tokens and per million output tokens. The current rates are in the model catalog and on the pricing page. Billing is prepaid from a balance you load by card, from $5, and you can never be charged more than that balance.
Can I use Mistral models on FastMetal?
Yes. Devstral 2512 (devstral-2512) was the first. It is built for coding and has a 262K-token context.
Can I compare these models with others?
GPT-5 mini, GPT-4.1 mini, GPT-5.4 nano, gpt-oss-20B and Llama 3.3 70B are in the model comparison, where you can read their real answers side by side with other models.