ruby_llm-providers-infomaniak

Open-weight models hosted in Switzerland, through the RubyLLM API you already know.

This gem adds Infomaniak AI Tools as a provider to RubyLLM. Kimi, Qwen, Gemma, Mistral and Apertus run in Infomaniak's Swiss data centres; your Ruby code keeps using RubyLLM.chat, tools, schemas, streaming and embeddings as with any other provider.

chat = RubyLLM.chat(model: 'moonshotai/Kimi-K2.6', provider: :infomaniak)
chat.ask('Explain Ruby blocks in one sentence.').content
# => "A Ruby block is a chunk of code you pass to a method ..."

• Chat, streaming, tool calling, structured output, image input, embeddings
• Reranking, image generation (Flux) and transcription (Whisper)
• Thinking on/off per request, with the model's reasoning on response.thinking
• Only an API token to configure: the AI Tools product is looked up for you
• A model catalog with context sizes and capabilities, refreshable at runtime
• Irons out Infomaniak's differences from OpenAI's API, such as Kimi's broken JSON while thinking

• Installation
• Configuration
• Models
• Usage
• How Infomaniak differs from OpenAI
• Development
• Releasing

# Gemfile
gem 'ruby_llm-providers-infomaniak', require: 'ruby_llm/providers/infomaniak'

Requires Ruby 3.2+ and RubyLLM 2.x.

You need an Infomaniak API token with the AI Tools scope (Infomaniak Manager > API tokens).

require 'ruby_llm/providers/infomaniak'

RubyLLM.configure do |config|
  config.infomaniak_api_key = ENV['INFOMANIAK_API_KEY']
end

That is all. On first use the provider asks GET /1/ai which AI Tools product the token belongs to, and caches the answer per token.

The chat and embedding models Infomaniak served on 2026-09-30, and how each behaved when probed. Your product lists its own with bundle exec rake models, or live with RubyLLM::Providers::Infomaniak.refresh_models!.

Photomaker, which needs reference photos on its own route, is not covered.

Any model id your product serves works as given, catalog or not. The gem ships a catalog, models.json, so that ruby_llm also knows each model's context size and capabilities. That enables:

• with_thinking and with_thinking(false), which need to know how a model switches thinking
• RubyLLM.chat(model: 'moonshotai/Kimi-K2.6') without provider:
• RubyLLM.models.by_provider(:infomaniak) without a network call

The catalog is a snapshot from the gem release. To know about models Infomaniak added since, load the live list at boot (two API calls):

RubyLLM::Providers::Infomaniak.refresh_models!

RubyLLM.models.refresh! does not do this: it skips providers that ship their own catalog.

chat = RubyLLM.chat(model: 'moonshotai/Kimi-K2.6', provider: :infomaniak)

chat.with_instructions('Answer like a Swiss train conductor.')
chat.ask('When does the next train to Bern leave?').content

chat.ask('Tell me a story about a marmot.') { |chunk| print chunk.content }

class Weather < RubyLLM::Tool
  description 'Current weather for a city'
  parameter :city, description: 'City name'

  def execute(city:) = WeatherService.current(city)
end

chat.with_tools(Weather).ask('Do I need an umbrella in Lausanne today?').content

schema = {
  type: 'object',
  properties: { city: { type: 'string' }, population: { type: 'integer' } },
  required: %w[city population],
  additionalProperties: false
}

chat.with_schema(schema).ask('Largest city in Switzerland?').parsed
# => {"city" => "Zurich", "population" => 443000}

Kimi and Qwen think before they answer, unless told not to. The reasoning comes back separately from the answer:

response = chat.ask('Is 1001 prime?')
response.thinking&.text   # the model's reasoning
response.content          # the answer

chat.with_thinking(effort: :none).ask('Quick: 17 * 23?')   # thinking off: faster, cheaper
RubyLLM.chat(model: 'google/gemma-4-31B-it', provider: :infomaniak)
       .with_thinking(effort: :high).ask('Plan a 3-day Ticino trip.')   # thinking on

With the model in the catalog, with_thinking and with_thinking(false) work too.

chat.ask('What is on this receipt?', with: 'receipt.jpg').content

RubyLLM.embed('Grüezi mitenand', model: 'Qwen/Qwen3-Embedding-8B', provider: :infomaniak).vectors

Sort retrieved passages by relevance before handing them to a chat model:

rerank = RubyLLM.rerank('Which metal is the densest?', passages,
                        model: 'BAAI/bge-reranker-v2-m3', provider: :infomaniak, top_n: 3)
rerank.results.map { |result| [result.score, result.document] }

image = RubyLLM.paint('A red Swiss train crossing a stone viaduct, watercolor',
                      model: 'flux', provider: :infomaniak, size: '1024x1024') # or 1024x1792, 1792x1024
image.save('train.jpg')

Prompts work best in English, and are limited to 77 tokens. Infomaniak's extra options pass through provider_options:, e.g. provider_options: { style: 'photographic' }.

RubyLLM.transcribe('meeting.m4a', model: 'whisper', provider: :infomaniak, language: 'de').text

Infomaniak runs transcriptions as background jobs: the gem uploads the file, then polls until the transcript is ready, so the call blocks for about as long as the job takes. Streaming (passing a block) is not possible. mp3, mp4, m4a, wav, flac, ogg, opus, aac, wma and webm are accepted.

The catalog carries Infomaniak's list prices, so ruby_llm computes costs from the token counts:

response = chat.ask('Summarise this contract.', with: 'contract.txt')
response.cost.total   # => 0.0021 (CHF)
chat.cost.total       # the whole conversation

Costs from this provider are in CHF, exactly as Infomaniak bills them, although ruby_llm documents its prices as USD. Don't add them to costs from other providers without converting. Chat, embedding and rerank costs are computed; Flux and Whisper bill per minute, which ruby_llm cannot compute, so their prices are only in model.metadata[:price_per_minute]. The prices date from 2026-09-30 and live in lib/ruby_llm/providers/infomaniak/pricing.rb.

The gem handles these, so the RubyLLM API behaves as usual:

bundle install
cp test/.env.example test/.env    # put your API token in test/.env (gitignored)

test/.env also takes INFOMANIAK_MODEL (the console and live-test model, default Kimi-K2.6), INFOMANIAK_EMBEDDING_MODEL (adds embeddings to rake live) and RUBYLLM_DEBUG=1 (logs every request).

Tags publish to RubyGems through trusted publishing, without a stored API key. See RELEASING.md.

MIT. Not affiliated with Infomaniak.