KosComputeAPI
Updates

Changelog

User-visible API, model, capability, limit, pricing, and reliability changes.

View Markdown

2026-10-13

  • Added nvidia/parakeet-tdt-0.6b-v3 (aliases parakeet-tdt-0.6b-v3, parakeet-v3) for POST /v1/audio/transcriptions: multilingual speech-to-text for 25 European languages including Czech, with punctuation, segment and word timestamps, and json, text, verbose_json, srt and vtt output. Audio up to 60 minutes per request; per-minute pricing is published in the model catalog. The language hint is accepted but not required; the model detects the language itself.
  • Removed the non-turbo Whisper Large v3 model and its aliases from the public API. Requests for it now return 404. Use nvidia/parakeet-tdt-0.6b-v3 or openai/whisper-large-v3-turbo instead.
  • Raised the kosmik/mimo-v2.6-flash-uncensored context window from 131,072 to 262,144 tokens; max_output_length stays 32,768. Very long prompts take longer to reach the first token (roughly 3-4 minutes at 250K input tokens), and several long requests in flight are processed one after another.

2026-10-09

  • Added kosmik/mimo-v2.6-flash-uncensored (alias mimo-v2.6-flash-uncensored): uncensored MiMo-V2.6 Flash RL for text chat, reasoning, tool calling and structured output on rental spare capacity, with published token pricing. It has no reserved capacity and returns 503 while unavailable.

2026-10-01

  • Removed the MiMo model family from the public API.
  • Qwen3.8 27B now runs only on available rental spare capacity. No GPU is reserved for the API by default. Rental jobs take priority and can interrupt active Qwen requests. When no Qwen backend is ready, requests return 503 rather than falling back to another model.
  • Host GPU checks also take priority over spare-capacity Qwen workers. Qwen readiness can briefly change even when no new rental appears.

2026-09-29

  • Added dealignai/mimo-v2.6-flash-uncensored with the mimo-v2.6-flash-uncensored alias, text chat, reasoning, tools, structured output, 131,072-token context, and published token pricing.
  • Kept qwen/qwen3.8-27b available on opportunistic GPU capacity. Compatible text requests may use the configured MiMo fallback when no Qwen backend is ready; the requested model ID and Qwen pricing remain in effect.

2026-08-17

  • Rebuilt the public documentation as an English-only, Markdown-first developer portal.
  • Added generated OpenAPI 3.1 JSON/YAML, integrated API reference, raw Markdown pages, and /llms.txt.
  • Made the public model catalog live from GET /v1/models with a timestamped fallback snapshot.
  • Clarified total completion-token budgets, stateless Responses behavior, Anthropic-specific 529 mapping, combined vision upload limits, and Zero Data Retention metadata boundaries.
  • Added contract, example, link, rendering, drift, and public-artifact leak validation.

Future entries include only changes that affect API users. Internal deployment and infrastructure details remain in the private operations knowledge base.

Type to search guides, models, and API reference.