Included catalogue
Every tier can access the same catalogue. Tiers differ by capacity, SLA, support, context and priority, not by hiding models behind a premium plan.
| Model | Endpoint | Primary use | License |
|---|---|---|---|
| Qwen/Qwen3.6-35B-A3B | /v1/chat/completions | Fast, high-volume chat: assistants, RAG, classification, extraction and summaries | Apache 2.0 |
| Qwen/Qwen3.8-27B | /v1/chat/completions | The most capable: agents with tools, code, multi-step reasoning and images | Apache 2.0 |
| Qwen3-Embedding-8B | /v1/embeddings | Semantic search and retrieval | Apache 2.0 |
| Qwen3-Reranker-4B | /v1/rerank | Second-stage reranking for RAG | Apache 2.0 |
| nvidia/canary-1b-v2 | /v1/audio/transcriptions | Transcription with diarization; the most accurate on Spanish | CC-BY-4.0 |
| nvidia/parakeet-tdt-0.6b-v3 | /v1/audio/transcriptions | Transcription with diarization; the better of the two on English and the fastest | CC-BY-4.0 |
| whisper-large-v3 (+ turbo) | /v1/audio/transcriptions | Transcription for compatibility: Catalan, 99 languages and translation into English | MIT |
| Kokoro 82M TTS | /v1/audio/speech | Voice for IVR and agents | Apache 2.0 |
Chat models and aliases
There are two chat models, and the difference is not that one does text and the other images: one is the fast one and the other the most capable. Qwen3.8-27B is the smartest: a dense model that uses all of its 27B parameters on every token, it nearly doubles Qwen3.6 on the Artificial Analysis Intelligence Index (34 versus 18), and it is the one for agents that chain tools, code, multi-step reasoning and complex instructions; it also reads images. Qwen3.6-35B-A3B is a mixture-of-experts that activates only 3B parameters per token: it answers sooner and handles more volume, and it is the best option for chat, RAG, classification, extraction and summaries. Each one ships with two aliases: the base name answers in direct mode and the -thinking suffix enables extended reasoning.
| Literal | Mode | Input |
|---|---|---|
| Qwen/Qwen3.6-35B-A3B | Direct (default) | Text |
| Qwen/Qwen3.6-35B-A3B-thinking | Thinking | Text |
| Qwen/Qwen3.8-27B | Direct | Text + image |
| Qwen/Qwen3.8-27B-thinking | Thinking | Text + image |
Qwen3.6 versus Qwen3.8
Qwen/Qwen3.6-35B-A3B | Qwen/Qwen3.8-27B | |
|---|---|---|
| Profile | The fast one: lower latency, more volume | The most capable: smarter on hard tasks |
| Architecture | Mixture-of-experts: 35B total, 3B active per token | Dense: all 27B active on every token |
| Artificial Analysis Intelligence Index | 18 | 34 |
| GPQA Diamond (scientific reasoning) | 86.0 | 89.2 |
| LiveCodeBench v6 (code) | 80.4 | 90.3 |
| Speed | Faster per request and with more requests at once | Slower per request |
| Input | Text | Text and image |
| Best for | Chatbots and assistants, RAG, classification, extraction, summaries and large batches | Agents with tools, code, multi-step tasks, complex instructions, invoices, scanned documents, screenshots and photos |
Audio models
Transcription has four literals. nvidia/canary-1b-v2 and nvidia/parakeet-tdt-0.6b-v3 are the recommended ones: they run on GPU, transcribe long real conversations better than Whisper, accept files of up to 4 hours and separate speakers in multi-speaker conversations with diarize=true. Both cover 25 European languages. whisper-large-v3 and whisper-large-v3-turbo stay available for compatibility. Speech synthesis is kokoro-82m, also on GPU.
| Literal | What for | Detail |
|---|---|---|
nvidia/canary-1b-v2 | Spanish and other known European languages | 6.6% difference on real Spanish calls; language required |
nvidia/parakeet-tdt-0.6b-v3 | English, unknown language, high volume | 7.6% on real English interviews; ×300–400 real time |
whisper-large-v3, whisper-large-v3-turbo | Existing integrations, Catalan, translation | No diarization; up to 25 MB per file |
kokoro-82m | Speech synthesis | About 0.2 s per short sentence; voices in Spanish, English and 6 more languages |
Practical compatibility
- Use the official OpenAI SDK and change
base_urltohttps://api.tesseraai.cloud/v1. - Keep your
messages,tools,response_formatand streaming structures where the endpoint supports them. - Commercial model aliases can map internally to preserve contractual stability.
Model freeze and upgrades
Contracts can freeze a model family for 12 months. Upgrades are offered as opt-in to avoid silent changes in quality, token cost or latency.