Russian-language LLMs for tool invocation, translations, and financial analytics

Previous article with a selection of models for Russian

All models from the list were personally tested by me and behave exactly as described. My computer is equipped with 64GB of RAM. According to calculations, to run on GPU you need exactly two RTX 3090 graphics cards with 24GB of VRAM each on one machine: llama.cpp in cortex.so should automatically detect them and use their memory jointly (link).

1. Model for Supervisor agent

Gemma 3 is best suited for orchestrating a swarm of agents. You need to use the 27b version, as it works without hallucinations: version 12b is suitable for generating technical documentation, but calls non-existent tools. Version 4b ignores part of the system prompt. Version 4b spams with guesses and leading questions

Model link:https://huggingface.co/tripolskypetr/gemma-3-27b-it

2. Model for language translations

YandexGPT 5 Lite Instruct works very well for translations. Unlike Saiga, this version will not call tools, but the dataset is made by Yandex itself, which improves its linguistic abilities when translating from English to Russian. You can use it in a pipeline to make responses from subsequent models look nice

Model link: https://huggingface.co/tripolskypetr/yandexgpt_5_lite_8b_instruct_gguf

3. Model for financial analytics

Allows to generate reports of the following kind, in Russian

Report on the medium-term outlook for 19/05/2025
Report on the forecasted bitcoin trend (1-3 days)
Analysis date: May 19, 2025
Analysis of MACD indicator signals
Latest MACD signals:
May 19, 2025, 13:30 (UTC)
Signal type: BUY
Price: 103032.46
Reason: MACD crossed the signal line upward
Signal strength: Strong
Strategy: Crossing the signal line
May 19, 2025, 13:30 (UTC)
Signal type: BUY
Price: 103032.46
...
Reason: MACD histogram crossed the zero line downward
Signal strength: Medium
Strategy: Crossing the zero line
Trend forecast:
Based on the latest MACD signals, the following trends can be identified:
Positive signals (BUY): Starting from 13:15 (UTC), there is a series of buy signals, including strong signals such as MACD crossing the signal line upward and the histogram crossing the zero line upward. This indicates a possible recovery of the upward trend.
Negative signals (SELL): At 12:30 (UTC), strong sell signals were recorded, indicating a short-term price decline.
Conclusion:
In the medium-term perspective (1-3 days), an upward trend for bitcoin is likely, considering the predominance of strong buy signals in recent hours. However, possible short-term corrections should be taken into account. It is recommended to monitor further MACD signals for trend confirmation.

Model link: https://huggingface.co/tripolskypetr/Plutus-Meta-Llama-3.1-8B-Instruct-bnb-4bit

4. Alternative to Gemma 3 if you don’t like its communication style

Partially ignores system prompt, but can call tools, including enum. For those who don't like the glossiness of the model's gemma language

Model link: https://huggingface.co/tripolskypetr/Codestral-22B-v0.1-GGUF

5. Model for debugging System Prompt

If the system prompt was written by several developers, the takes inside may contradict each other. For detection, it is convenient to use the reasoning model. Notably, the Qwen3 32B model may not perform reasoning if the system prompt contains the inclusion of the text /no_think

Model link: https://huggingface.co/tripolskypetr/Qwen3-32B-GGUF