LFM 2.5 2.6B
Neural Network
LFM 2.5 2.6B is a compact reasoning model by Liquid AI for agentic scenarios, data extraction, and RAG.
Specifications
- Input
- text, images, documents
- Output
- text
- Context size
- 65 536 tokens
- Max answer length
- 8 192 tokens
- Reasoning effort
- low, medium, high
- Capabilities
- reasoning, function calling
How it works LFM 2.5 2.6B?
LFM 2.5 2.6B is a compact reasoning model from Liquid AI. Its small size makes it convenient for steady streams of similar requests: agentic scenarios, data extraction, RAG, and long context tasks. The developer notes it is not intended for agentic programming, so choose another for code pipelines. In BotHub, you can run it without a VPN or foreign card; payment is made with Russian cards in rubles, charged per token without a subscription. All models are in one window, making it easy to compare outputs and switch between them. A unified OpenAI-compatible API allows replacing the model in a multi-model system without rewriting integrations. Requests are encrypted, and teams have access to contracts, invoices, EDI, and an admin panel with limits. Use cases: an agent that parses and routes incoming requests; extracting fields from documents and emails into a structured format; an internal knowledge base assistant via RAG; brief summaries of long reports; a fast first layer before a larger model.Frequently asked questions about LFM 2.5 2.6B
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
The model works with text, accepts documents and images, can reason before answering, and call functions. The context window is 128,000 tokens, so you can upload long materials: reports, correspondence, code. Suitable for analysis, summarization, debugging, and integration into services via API.
In one response, the model outputs up to 8,192 tokens, which is roughly several pages of text. This is usually enough for an article, detailed analysis, or a large code snippet. If the material is longer, ask it to continue: previous parts will remain in the 128,000-token context.
Yes, reasoning before answering is confirmed by our data. The model breaks the task into steps before providing a result, which significantly helps with logic, math, and code analysis. For simple queries, the difference is small, but for multi-step conditions, the answers are more accurate.
Yes. Function calling and structured JSON output are supported, so the model can be connected to external services, databases, and custom tools. BotHub uses a unified OpenAI-compatible API: you define the function schema once and change the model in the request without rewriting the integration.
You can upload images and documents — the model accepts them as input along with text: ask it to describe an image, analyze a diagram, or extract data from a file. Audio and video are not listed as input formats in our data, so for voiceovers or recording analysis, choose a specialized model.
We do not have data on the quality of performance for specific languages, so we won't make any promises — test it with your own tasks: a couple of queries will show more than any description. Access to the model in BotHub is available without a VPN or foreign card; payment is made with Russian cards in rubles.