DeepSeek V4.1 Flash, Mistral Large 4 and three more models

Five text models join Tchavi: DeepSeek V4.1 Flash, Mistral Large 4, Command A+, Gemini 3.5 Flash-Lite and MiMo-V2.6-Flash. For high volume and quick answers, look at MiMo and Gemini Flash-Lite; for code, DeepSeek; for writing in French, Mistral Large 4; for answering from your own documents, Command A+.

In short

• Tchavi, an OpenAI-compatible AI API gateway you pay for with mobile money, adds five text models: deepseek-v4.1-flash, mistral-large-4, command-a-plus, gemini-3.5-flash-lite and mimo-v2.6-flash.
• All of them read images, call tools and return JSON.
• Prices run from 308 to 9,175 credits per million tokens (about KES 23 to KES 695), depending on the model and the direction (input or output).

What each one is for

DeepSeek V4.1 Flash. The 17 October 2026 version of DeepSeek's Flash model, the one DeepSeek serves itself. According to DeepSeek, it improves on code and gives shorter answers than the previous one. It thinks before answering (about 600 tokens and 15 seconds on a small coding task in my tests), and the thinking is billed as output. 1M-token context, output up to 393K.

Mistral Large 4. Mistral AI's flagship model, released on 6 October 2026 as a public preview. It writes carefully in French and English: announcements, emails, customer replies. On Tchavi it answers directly; with reasoning_effort: "high" it reasons first, which helps with calculations. In that case, message.content arrives as a list of blocks (thinking, then text), not a plain string.

Command A+. The latest in Cohere's Command A family, published in May 2026 under the Apache 2.0 licence. Its strength: answering from a text you supply while citing the passage it used (terms of sale, procedures, FAQs). One language per request: a test mixing English, Portuguese and Arabic mixed the languages three times out of three. tool_choice only accepts auto. 128K-token context.

Gemini 3.5 Flash-Lite. Google's model for high-volume tasks: classifying messages, extracting data from an invoice, translating in bulk. It answers without a thinking phase by default, in 2 to 3 seconds in my tests. 1M-token context, images and PDFs.

MiMo-V2.6-Flash. Xiaomi's fast model: under 2.5 seconds in my tests, no thinking. It suits extracting amounts and references from payment SMS messages (an M-Pesa confirmation, for example), sorting tickets, summarising. It reads images, not PDFs.

Prices

Credits first, then KES at October 2026 rates (1 credit is about KES 0.076):

Input and output are billed separately, on exact tokens. For models that think, the thinking counts as output. These are the prices on each model's page.

One call to try it

curl https://tchavi.com/api/v1/chat/completions \
  -H "Authorization: Bearer $TCHAVI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash",
    "messages": [
      {"role": "user", "content": "Write a JavaScript function that extracts the amount and the reference from a mobile money payment SMS."}
    ],
    "max_tokens": 1500
  }'

To switch models, change model. Leave enough max_tokens for the models that think.

How to try them

• Model pages: DeepSeek V4.1 Flash, Mistral Large 4, Command A+, Gemini 3.5 Flash-Lite, MiMo-V2.6-Flash. Each shows a real demo.
• Read the chat completions page and the list of models.

FAQ

Does Mistral Large 4 reason by default?

No, not on Tchavi: it answers directly, and reasoning_effort: "high" turns thinking on. Only the values "none" and "high" are accepted.

Why does my code break with Mistral Large 4 in reasoning mode?

Because message.content becomes a list of blocks. Read the text block instead of treating content as a string.

Which model for sorting a lot of messages?

Test MiMo-V2.6-Flash and Gemini 3.5 Flash-Lite on a sample of your real messages, then compare cost and accuracy.

create your account and top up KES 250

This article is part of the series "Building an AI product from Africa". Tchavi is an OpenAI-compatible AI API gateway: one key gives access to over 100 text, image, audio and video models, and you pay for it with mobile money, no bank card (countries and operators). More guides on the blog. Follow us: LinkedIn and X.