Models and prices
Every model runs on Blixild's own GPUs in Poland and answers through one API that works with any OpenAI SDK. A model's name leads to its page: what it is for, its limits and a first call to copy.
Nous Research's Arete 27B, a general-purpose chat model with reasoning and tool calling, fine-tuned from Qwen3.8 27B. Text only.
4-bit (W4A16, AutoRound) · up to 104,448 tokens per request
- Input
- $0.10 / M tokens
- Output
- $1.50 / M tokens
Details
Arete 27B is Nous Research's fine-tune of Qwen3.8 27B, a general-purpose model for chat, writing, analysis and code. It reasons before it answers: the reasoning comes back separately (reasoning_content) and its tokens count as output. It calls tools as OpenAI's API does: offer them in tools, and its calls come back in tool_calls, streamed too; leave room in max_tokens for the reasoning before a call. Text only.
- Precision
- 4-bit (W4A16, AutoRound)
- Max request
- 104,448 tokens, prompt and answer together
- Shared GPU
- up to 8 requests at once across all customers, 2 per key and 4 per customer; when it's busy, requests wait their turn
- Made by
- Nous Research, from Qwen3.8 27B
- Runs on
- Blixild's own RTX 3090, in Poland (EU)
- Speed
- about 45 tokens/s for one request
- Languages
- tested in English and Polish
- Tools
- OpenAI-style tool calling (tools, tool_calls), streamed too
- Good for
- drafting, summaries, questions over long documents, code, agents with tools
- Not for
- images or audio; large batch jobs
- License
- Apache 2.0 · Model card
- basal-2.0-4.5bDecisions
basal 2.0 (4.5B), an open decision model for Polish and English. It answers typed questions about a text (yes or no, a choice, a score) with a probability for each answer and writes no text.
up to 32,768 tokens per request
- Input
- $0.10 / M tokens
- Output
- Not charged
Details
basal 2.0 (4.5B) is an open decision model fine-tuned from Bielik v3. Send it a text (state) and named questions on /v1/systemone, or OpenAI's Decisions format on /v1/decisions: yes or no (noul), one of 2 to 255 options (choice), or a level on a scale (score). Each answer comes back with a calibrated probability for every option, in one pass, so it can only be one of the options you gave. Several questions over one text cost little more than one. You pay for the input tokens only: it generates nothing. It works best when every option has a short description, the rules are in the text, and dates and amounts are worked out in your code. Not a chat model: /v1/chat/completions refuses it.
- Max request
- 32,768 tokens, the text and its questions
- Made by
- the basal project (github.com/rkinas/basal), fine-tuned from Bielik v3
- Runs on
- Blixild's own H200, in Poland (EU)
- Speed
- 10 ms for one question; 232 decisions/s with 32 clients, measured
- Languages
- Polish and English
- Questions
- yes or no, choice (2 to 255 options), score; several per request over one text
- Good for
- routing messages and tickets, rule checks with the rules in the text, triage with a confidence threshold, an agent's next step
- Not for
- writing text; date and amount arithmetic (work it out in your code); texts over 32,768 tokens (the middle is cut)
- License
- Apache 2.0 · Model card
- basal-2.0-miniDecisions
basal 2.0 mini (1.5B), the small and fast basal decision model for Polish and English. It answers typed questions about a text with a probability for each answer and writes no text.
up to 8,192 tokens per request
- Input
- $0.05 / M tokens
- Output
- Not charged
Details
basal 2.0 mini (1.5B) is the small basal decision model, fine-tuned from Bielik v3: the same requests as basal 2.0 (4.5B) on /v1/systemone or /v1/decisions, about twice as fast, for high volumes of simple decisions. It is as good as the 4.5B at routing and at picking an agent's next tool when every option is described; for questions where several labels interact, use the 4.5B. You pay for the input tokens only. Texts up to 8,192 tokens. Not a chat model: /v1/chat/completions refuses it.
- Max request
- 8,192 tokens, the text and its questions
- Made by
- the basal project (github.com/rkinas/basal), fine-tuned from Bielik v3
- Runs on
- Blixild's own H200, in Poland (EU)
- Speed
- 4.5 ms for one question; 446 decisions/s with 32 clients, measured
- Languages
- Polish and English
- Questions
- yes or no, choice (2 to 255 options), score; several per request over one text
- Good for
- high volumes of simple routing, an agent's next tool
- Not for
- writing text; questions where several labels interact (use basal-2.0-4.5b); texts over 8,192 tokens
- License
- Apache 2.0 · Model card
SpeakLeash and ACK Cyfronet AGH's Bielik 11B v3.1, an open model made for Polish, with tool calling and reasoning on request. Text only.
8-bit (FP8, W8A8) · up to 32,768 tokens per request
- Input
- $0.20 / M tokens
- Output
- $0.50 / M tokens
Details
Bielik 11B v3.1 is an open model from SpeakLeash and ACK Cyfronet AGH, trained with a focus on Polish and other European languages. It answers directly by default; send reasoning_effort (low, medium, high or max) and it reasons first: the reasoning comes back separately and its tokens count as output. It calls tools as OpenAI's API does: offer them in tools, and its calls come back in tool_calls, streamed too. Its model guidelines (about 1,000 tokens: who it is, the date, safety rules) come first in every request and count as input. For work over several tool calls it needs a complete system prompt with the procedure written out step by step. Text only.
- Precision
- 8-bit (FP8, W8A8)
- Max request
- 32,768 tokens, prompt and answer together
- Shared GPU
- up to 8 requests at once across all customers, 2 per key and 4 per customer; when it's busy, requests wait their turn
- Made by
- SpeakLeash and ACK Cyfronet AGH
- Runs on
- Blixild's own RTX 5090, in Poland (EU)
- Speed
- 328 tokens/s in Polish and 186 in English for one request, measured
- Languages
- Polish first; tested in Polish and English
- Tools
- OpenAI-style tool calling (tools, tool_calls), streamed too
- Reasoning
- on request (reasoning_effort low, medium, high, max); off by default
- Good for
- Polish text and chat, Polish documents without personal data, agents with a written procedure
- Not for
- personal data, which SpeakLeash's terms of use (bielik.ai/en/terms) forbid processing with it; images or audio; prompts over 32,768 tokens; date and time arithmetic without a tool
- License
- Apache 2.0 · Model card
Qwen3.8 27B from Alibaba's Qwen team, a general-purpose model with reasoning and tool calling and a 262,144-token context. Text only.
8-bit (FP8) · up to 262,144 tokens per request
- Input
- $0.20 / M tokens
- Output
- $1.50 / M tokens
Details
Qwen3.8 27B is the Qwen team's model for chat, writing, analysis, code and agents, served in Qwen's own 8-bit build. It reasons before it answers: the reasoning comes back separately (reasoning_content) and its tokens count as output. For shorter reasoning send reasoning_effort low or medium; to have it answer at once, send chat_template_kwargs {"enable_thinking": false}. It calls tools as OpenAI's API does: offer them in tools, and its calls come back in tool_calls, streamed too. A request, prompt and answer together, may hold up to 262,144 tokens. A long prompt takes a while before the first token (about 17 seconds at 130,000 tokens, about a minute at 250,000), so stream the answer or allow for the wait; further questions about the same document start within seconds. Text only.
- Precision
- 8-bit (FP8)
- Max request
- 262,144 tokens, prompt and answer together
- Shared GPU
- up to 8 requests at once across all customers, 2 per key and 4 per customer; when it's busy, requests wait their turn
- Made by
- Qwen (Alibaba Cloud)
- Runs on
- Blixild's own NVIDIA H200, in Poland (EU)
- Speed
- 196 tokens/s in English and 171 in Polish for one request, measured
- Long prompts
- facts found anywhere in English and Polish documents up to 254,000 tokens, measured
- Languages
- tested in English and Polish
- Tools
- OpenAI-style tool calling (tools, tool_calls), streamed too
- Good for
- questions over long documents or many files at once, analysis, code, agents with tools
- Not for
- images or audio
- License
- Apache 2.0 · Model card
Prices are in US dollars per million tokens, including VAT where it applies, taken from your credit as you go. Add credit from $5 by card, Apple Pay or Google Pay. When your credit reaches $0, requests stop: no bill arrives later.