From the factory · 07 Sep 2026Changelog Longer answers. Full Bielik context.Follow the API as it evolves. See what’s new
Bielik + Qwen · OpenAI-compatible API

An API with
full-vocabulary
logprobs.

Token probabilities for distillation, model comparisons, and evaluation. Full vocabulary on Bielik. Top alternatives on Qwen.

32,128 Bielik tokens Top 20 Qwen alternatives JSON + SSE with token IDs
Explore the logprobs API → Try a live model

Bielik: full distributions for 1–8 generated positions per request. Guide, Python exporter, and distillation examples included.

§ 01 · Live production test

Try the actual API.

No account and no mock response. Choose a route, send a prompt, and watch the answer arrive through the same streamed interface available to developers.

Route response with

Not sure? Auto Router chooses a live model before streaming starts.

Ready · prompts and responses are not retained
§ 02 · Production evidence · last 30 daysOpen live usage report →
4Models publishedBielik · Qwen · PLLuM · hybrid
—Completed requestsAggregated production activity
—Tokens processedInput and output combined
LIVEPublic usage dataUpdated every minute
§ 02 · Product map

One factory, three connected systems.

Inference generates revenue and operational data. Research improves the serving system. CodeSOTA supplies external measurement and model evidence.

01 · Production

Inference API

Use open models through one endpoint with streaming, authentication, usage tracking, and an OpenAI-compatible request shape.

Open the API →02 · Laboratory

Learn from every token

Get full-vocabulary log probabilities from Qwen and Bielik. Build teacher datasets, distill a student, and measure model behavior.

Explore logprobs & distillation →03 · Measurement

CodeSOTA

Independent benchmark evidence answers which model is strongest for a specific capability before it reaches production.

Browse the registry ↗
§ 03 · Production lines

Models available now.

Choose the smallest system that reliably completes the workload, or let Auto Router choose. Every row maps to an identifier exposed by /v1/models.

# Model Role Language / mode Interface Status Action
00 Auto Routerauto Task-aware model selection Automatic STR ROUTE Live
01 Bielik 11B v3bielik-11b-v3 Fast general inference Polish STR Live
02 Qwen3.8 27Bqwen3.8-27b Agents and tool calling Direct answer STR TOOL Live
03 PLLuM 12Bpllum-12b-instruct-2512 Longer Polish context Polish · 16k STR Live
04 Qwen + Bielikqwen-bielik-hybrid Reason, review, edit Polish synthesis STR LAB Live
Table 1 · Current public model and route registry. Availability is an operational signal, not a claim of task-level superiority.
§ 04 · Factory method

Research that returns to production.

The factory loop is deliberately closed: compute produces inference and observations; observations become experiments; verified improvements return to serving.

  1. 01
    Serve

    Run real workloads and measure latency, throughput, failures, and cost.

  2. 02
    Evaluate

    Compare model behavior on task-relevant evidence rather than generic claims.

  3. 03
    Optimize

    Test quantization, batching, routing, prompts, and hardware utilization.

  4. 04
    Ship

    Return proven changes to the API and document the operational delta.

§ 06 · Agent-ready endpoint

Give this to your AI agent.

Paste the instruction into Hermes, Claude Code, Codex, or another coding agent. It tells the agent exactly how to configure Fabryka without exposing your key in source control.

Read the complete Hermes setup guide →
CONFIGURATION PROMPT 
Configure this project to use Fabryka AI as an OpenAI-compatible provider.

Base URL: https://fabryka.ai/v1
Default model: auto
API key environment variable: FABRYKA_API_KEY

Requirements:
1. Read the key only from FABRYKA_API_KEY. Never hardcode, print or commit it.
2. Use POST /chat/completions and render streamed SSE deltas incrementally.
3. Preserve OpenAI-format tools, tool_choice, assistant.tool_calls and role: tool messages.
4. Add a connection check using GET /models.
5. Stop and report 401 or 402. Use capped exponential backoff for 429, 502, 503 and 504.
6. Do not send secrets, local files or tool output without explicit user approval.
7. Run one minimal request, then show me the observed result and files changed.
§ 07 · Brand promise

More intelligence
per GPU.

Better servingLower token costUseful experimentsMeasured output
§ 08 · Start producing

One key. Four open systems.

Generate a test key, inspect the API contract, or connect a background Hermes agent.

Generate API key ↑Read API docsBuild a Hermes agent →
API base URLhttps://fabryka.ai/v1Customer dashboard ↗