Changelog.

Muse Glimmer 30B

Added muse-glimmer:latest to the OpenAI-compatible API and model picker. It is a separate model ID with a 131,072-token context window. Qwen remains separately addressable.

{"model":"muse-glimmer:latest","messages":[{"role":"user","content":"Reply with: connected"}]}

The right model for the work

The new Router page explains how to use Auto Router today and separates the planned policy-first, agent-oriented routing architecture from currently available API capabilities.

The factory behind the API

Our new vision page connects research, evaluation, routing, and serving, with a clear distinction between what is available today and what we want to build next.

Larger output budgets for Bielik

• Explicit requests to bielik-11b-v3 and its bielik-11b alias now accept max_tokens above 4096, including 8192 and 16384 for short prompts, in chat and batch requests.
• Bielik has a shared 32,768-token context window: the prompt, chat formatting and requested output must fit together. The output parameter allows up to 32,768, but the available output budget is smaller once the input is included.
• The default remains max_tokens: 2048. An explicit budget is a maximum, not a guarantee that an answer uses every token. Other models' limits are unchanged.

{"model":"bielik-11b-v3","messages":[{"role":"user","content":"Rozwiąż zadanie krok po kroku i sprawdź wynik: ile wynosi 17 × 23?"}],"max_tokens":8192}

Public uptime and proxy telemetry

• Added /status, a public API status dashboard with calculated 1h, 24h, 7d and 30d uptime, p50/p95 latency, HTTP code distribution and grouped incidents.
• The dashboard now combines API request outcomes with Caddy gateway errors, including 502, 503 and 504 responses that never reach the application.
• Added a live busy signal for in-flight, queued and running work. Prompts, responses and customer identifiers are not exposed.

Public streaming chat

• Added /chat, a no-login web chat for trying Bielik, Qwen, and Auto Router through the production streaming demo.
• Responses arrive as live streamed deltas in the Fabryka interface. Conversation state stays in the browser and is not retained after delivery.
• Added model selection, clear conversation, starter prompts, mobile layout, and links to API documentation.

Logprobs take the front page

• A prominent homepage banner now links directly to the changelog, above the fold on desktop and mobile.
• The homepage now leads with the logprobs API: full-vocabulary probabilities on Bielik, top alternatives on Qwen, and a direct link to the guide, exporter, and distillation examples.
• Live checks confirmed all 32,128 Bielik vocabulary tokens. The guide now identifies the current Qwen limit of 20 alternatives and uses Bielik for its full-distribution request.

Longer answers and full Bielik context

• Chat and batch requests now default to max_tokens: 2048 instead of 256, including Bielik and Qwen. The live demo also uses 2048 output tokens.
• Bielik 11B v3 now has 32,768 tokens of context per request, shared by the prompt, chat formatting and generated answer. The API no longer restricts explicit Bielik requests to 4096 tokens of total context.
• Explicit max_tokens values still take precedence. This parameter counts tokens, not characters; the default is an output budget, not a guarantee that every answer uses it all.

{"model":"bielik-11b-v3","messages":[{"role":"user","content":"Napisz szczegółową odpowiedź po polsku."}],"max_tokens":2048}

Auto Router and readable model controls

• Added model: "auto" to the existing OpenAI-compatible POST /v1/chat/completions endpoint.
• Tool and structured requests route to Qwen, ordinary Polish requests to Bielik, and long Polish context to PLLuM. The hybrid pipeline remains an explicit choice.
• Responses and streams report the actual selected model; response headers also expose the requested model, route reason, and streaming mode.
• Replaced native model menus with accessible task-oriented choices and raised compact labels to readable desktop and mobile sizes.

curl -N https://fabryka.ai/v1/chat/completions \
  -H "Authorization: Bearer $FABRYKA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"Explain this in Polish."}],"stream":true}'

One Fabryka product surface

• Consolidated the live streamed demo, API-key onboarding, runnable code snippets, public usage evidence, API documentation, and Hermes setup under fabryka.ai.
• Added an English, aggregate-only /usage report covering requests, tokens, response time, throughput, and daily production trends.
• Added a copyable coding-agent configuration prompt and an English Hermes guide using the canonical Fabryka endpoint.
• Rebranded OpenAPI and Swagger documentation, including a proper Bearer authorization workflow.
• New API keys use the fab_live_ prefix; existing keys remain valid.

Native streaming and Qwen tool calling

• Bielik 11B v3 and Qwen3.8 27B now forward native upstream deltas through POST /v1/chat/completions and the public demo instead of waiting for a completed response.
• Direct Qwen calls disable the private thinking channel so the output budget produces a visible streamed answer; the hybrid pipeline still uses reasoning internally.
• Qwen3.8 27B accepts tools, tool_choice, assistant tool_calls, and role: tool results.
• PLLuM and the hybrid pipeline continue to use streamed client delivery with explicit pipeline progress.

curl -N https://fabryka.ai/v1/chat/completions \
  -H "Authorization: Bearer $FABRYKA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3.8-27b","messages":[{"role":"user","content":"Hello"}],"stream":true}'

Hybrid demo and public usage

• Introduced the multi-stage Qwen + Bielik pipeline with SSE progress events.
• Published aggregate service usage without prompts or customer data.
• Added anonymous test keys with renewable daily credit.

Public batch API

• Opened the asynchronous /v1/batches interface.
• Added status retrieval, result delivery, balance accounting, and actual-usage charging.