EVREN model selection for text, images, and video

Short answer

There are seven EVREN rows on LLMTR, and the first difference between them is the input type. Two rows take text only (DeepSeek V4 Flash TR and GLM-5.3 FP8), three take images as well as text (Gemma 4 31B, MiMo V2.6 Pro and DeepSeek V4.1 Flash TR), and two accept video on top of images (Qwen3.8 Flash Next and Qwen3 VL 30B). Within one input type, the choice is decided by the role EVREN gives each model.

We recommend using EVREN through LLMTR: you register your own EVREN key once in Settings, and the same code then reaches both EVREN models and the rest of the catalog. The key stays yours, and EVREN's own logging and retention policy still applies unchanged.

• LLMTR documentation: EVREN

The seven rows at a glance

The context windows, input types and roles in the table come from EVREN's published model information. All of them return text; there is no price column because LLMTR charges nothing for these rows.

• LLMTR documentation: EVREN
• Source: SAYZEK EVREN announcement (checked 25 September 2026)

Text work

The row EVREN positions as its most capable is GLM-5.3 FP8: a mixture-of-experts model with 753 billion total parameters and roughly 40 billion active per token. It is recommended for deep reasoning, code generation, multi-step agent loops and long document analysis. Its context is 512,000 tokens; the architecture allows 1 million, but EVREN serves this row at 512,000.

DeepSeek V4 Flash TR is lighter, with 284 billion total and roughly 13 billion active parameters, and it carries a 1,000,000-token context. EVREN presents it as its code and agent specialist. Because the number of parameters active per token shows how much computation a model does for each token, rows with fewer active parameters usually answer faster. DeepSeek V4 Flash is described as having an adjustable thinking budget, but that setting is not offered through LLMTR.

For a task that needs only text, Gemma 4 31B is also an option: a 30.7-billion-parameter dense model and the everyday row EVREN recommends for fast chat and document reading.

MiMo V2.6 Pro is the mixture-of-experts model EVREN presents as the largest in its fleet, with about 1.02 trillion total parameters and roughly 42 billion active per token, positioned for long-running agent work. Its 1,048,576-token window is shared by input and output. It offers tool calling, JSON mode and a way to turn thinking off; the model card and the EVREN documentation have the details.

DeepSeek V4.1 Flash TR is the mixture-of-experts model EVREN presents as its next-generation code and agent model, with about 763 billion total parameters, of which about 8 billion are active per input token and 16 billion per output token; it is the successor EVREN recommends for the older DeepSeek V4 Flash row. Its 1,048,576-token window is shared by input and output, and it takes images as well. Like MiMo V2.6 Pro, it offers tool calling, JSON mode and a way to turn thinking off.

Images and video

Each of the five rows that accept images has a different emphasis. For Gemma 4 31B, EVREN lists image understanding, object detection with bounding boxes, screen and interface interpretation, chart reading and handwriting recognition. Qwen3 VL 30B is offered for heavy visual reasoning and frame-level event extraction. Qwen3.8 Flash Next also takes images and video, but EVREN describes vision on this row as secondary; its main strength is high concurrency. MiMo V2.6 Pro and DeepSeek V4.1 Flash TR also take images; on these two rows an image can be combined with tool calling and a long context in the same request.

For video, the row EVREN puts forward is Qwen3 VL 30B. The provider states 2 frames per second and up to 520 frames, which is about 260 seconds of video. A longer recording exceeds that capacity; splitting it into segments is the way to stay within the limit. Images are added to the message content as an image_url part, in the OpenAI-compatible chat format, as a base64 data URL; EVREN does not download an image from an http(s) address.

Sending an invoice image to Gemma 4 31B

import base64, os
from openai import OpenAI

client = OpenAI(base_url=os.environ["LLMTR_BASE_URL"] + "/v1", api_key=os.environ["LLMTR_API_KEY"])

with open("invoice.png", "rb") as f:
    image = "data:image/png;base64," + base64.b64encode(f.read()).decode()

res = client.chat.completions.create(
    model="evren/gemma-4-31b",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Give the seller name and the total amount on this invoice."},
            {"type": "image_url", "image_url": {"url": image}},
        ],
    }],
    max_tokens=500,
)
print(res.choices[0].message.content)

What these rows do not offer

What these rows lack affects the choice as much as the input type. Tool calling, JSON mode and thinking control are not offered on the EVREN rows other than MiMo V2.6 Pro and DeepSeek V4.1 Flash TR. An agent framework that relies on tool calling will not work as expected with those rows; that kind of task needs one of those two rows or another model in the catalog that supports tool calling.

Streaming on these rows is not real streaming: the answer is completed and then delivered in chunks. No separate output ceiling is published either; the limit on the model card is the context window. Because the same code also reaches the other models in the catalog, changing the model identifier is enough to see whether a task does better on an EVREN row or on another model.

• Source: EVREN OpenAPI schema (checked 25 September 2026)

Frequently asked questions

Should I use EVREN directly or through LLMTR?

We recommend using EVREN through LLMTR. Your requests still go to EVREN on your own EVREN key and LLMTR charges nothing for them; the difference is that the same code also reaches the other models in the catalog. EVREN's logging and retention policy is not affected by the choice.

Which EVREN models accept images?

Five rows: Gemma 4 31B, Qwen3.8 Flash Next, Qwen3 VL 30B, MiMo V2.6 Pro and DeepSeek V4.1 Flash TR. DeepSeek V4 Flash TR and GLM-5.3 FP8 take text only.

Which row suits video best?

Qwen3 VL 30B, which EVREN describes as its video row. The provider states 2 frames per second and up to 520 frames, about 260 seconds of video. Qwen3.8 Flash Next also takes video, but its vision is secondary.

Do EVREN rows support tool calling?

Only on the MiMo V2.6 Pro (evren/mimo-v2.6-pro-tr) and DeepSeek V4.1 Flash TR (evren/deepseek-v4.1-flash-tr) rows. The other five EVREN rows do not offer tool calling, JSON mode or thinking control; work that depends on those capabilities uses one of those two rows or other models in the catalog.

Related posts

• Building a team application without sharing the EVREN key
• The difference between an EVREN request record and a prompt record

• EVREN guide
• Model catalog