Models & limits
What you can call, how much, and which parameters work.
Models
| Fact | Qwen3.8 27B | Nemotron 3.5 Super VL | EuroLLM 22B | GLiDE 9B |
|---|---|---|---|---|
| Model id | qwen3.8-27b | nemotron-3.5-super-vl-120b | eurollm-22b | glide-9b |
| Vendor | Qwen | NVIDIA | EuroLLM consortium | Fastino |
| Parameters | 27B | 120B (12B active) | 22B | 9B |
| Quantization | FP8 | NVFP4 | BF16 | BF16 |
| Deployment | Live | Live | Live | Live |
| Context window | 1,000,000 tokens (prompt + reasoning + answer) | 1,000,000 tokens (prompt + reasoning + answer) | 32,768 tokens (prompt + reasoning + answer) | 16,384 tokens (state + one question) |
| Default output | 8,192 tokens | 8,192 tokens | 4,096 tokens | Not applicable |
| Max output, reasoning off | 32,768 tokens | 32,768 tokens | 16,384 tokens | Not applicable |
| Max output, reasoning on | 32,768 tokens (reasoning + answer) | 32,768 tokens (reasoning + answer) | Not supported | Not applicable |
| Reasoning | Off / Low / Medium / High · Medium by default | Off / Medium / High · Medium by default | Not supported | Not supported |
| Input | Text, Image | Text, Image | Text | Text |
| Capabilities | Chat, Streaming, Tool calling, JSON mode, Optional reasoning, Image analysis | Chat, Streaming, Tool calling, JSON mode, Optional reasoning, Image analysis | Chat, Streaming, JSON mode | Decisions |
| License | Apache-2.0 | OpenMDW-1.1 | Apache-2.0 | Fastino Community License |
| Price | Included in plan (€29/mo or €24/mo yearly, excl. VAT) | Included in plan (€29/mo or €24/mo yearly, excl. VAT) | Included in plan (€29/mo or €24/mo yearly, excl. VAT) | Included in plan (€29/mo or €24/mo yearly, excl. VAT) |
| Hosting | EU | EU | EU | EU |
GLiDE 9B is a decision model with its own endpoint: see Decisions below.
Inference uses shared capacity. Deployment status is listed above; current availability is shown in the console and on Status. A pending model is not available for general use.
GET /v1/models returns the chat models in this list as JSON, including deployment: "live" or "pending". Deployment status describes rollout; check model availability in the console before sending a request.
Limits
Inference is €29/month or €288/year (€24/month) per person, excl. VAT — or €19/month for the first 12 months / €192 for the first year (€16/month) when subscribed up to 1 December 2026 CET — covering all deployed models in Chat and the API. Paid plans include unlimited tokens and requests across all API keys and Chat. Existing pilot grants keep their pilot terms until expiry or subscription. Per-model concurrency limits apply to every account.
| Limit | Value | Details |
|---|---|---|
| Tokens | Unlimited | No daily or monthly token cap |
| Qwen3.8 27B concurrency | 2 admitted per account across all models | Preparation, queue and execution share this limit across all keys and Chat. At most 1 may reserve more than 262,144 input-plus-output tokens. |
| Nemotron 3.5 Super VL concurrency | 2 admitted per account across all models | Preparation, queue and execution share this limit across all keys and Chat. At most 1 may reserve more than 262,144 input-plus-output tokens. |
| EuroLLM 22B concurrency | 2 admitted per account across all models | Preparation, queue and execution share this limit across all keys and Chat. At most 1 may reserve more than 262,144 input-plus-output tokens. |
| Model capacity | Shared | 503 engine_busy when the queue is full or the wait exceeds 30 seconds. |
Requests receive an x-request-id header without per-minute token quota headers. Capacity is shared: requests may queue for up to 30 seconds, then receive a retryable 503. Each model has its own queue.
Supported
- Chat:
POST /v1/chat/completionswithsystem,user,assistantandtoolmessages.developeris treated assystem. - Streaming:
stream: true. Addstream_options.include_usageto get token counts in the last chunk. - Tools:
tools,tool_choice,parallel_tool_calls, with Qwen and Nemotron. EuroLLM 22B does not call tools. - Web search: not built into the API. Pass your own search function in
toolsand run it in your app. Chat has its own Search. - JSON output:
response_formatwith{"type": "json_object"}. - Sampling:
temperature,top_p,stop,seed,presence_penalty,frequency_penalty. Values you leave unset follow each model card: Qwen uses temperature 1.0, top_p 0.95 and top_k 20 with reasoning, and temperature 0.7, top_p 0.8, top_k 20 and presence_penalty 1.5 without. Nemotron uses temperature 1.0 and top_p 0.95, or temperature 0.6 when the conversation has images. EuroLLM's card gives no values, so Vora adds none. - Length:
max_tokensormax_completion_tokens, up to the selected model's output limit in the table above. The default is listed separately asdefault_output_tokens. Input, reasoning and the final answer must fit in the model's context window. - Reasoning: Use the model-specific efforts in the table. Qwen accepts
none,low,medium,high, plusminimalas Low andxhighormaxas High. Nemotron acceptsnone,mediumandhigh. Their default is Medium, and each request can choose direct answers or thinking. EuroLLM 22B has no thinking mode: omitreasoning_effortor sendnone. - Also passed through:
top_k,min_p,repeat_penalty. No model supportslogprobsortop_logprobs.
Thinking modes
Choose an effort with Reasoning in Chat, or set reasoning_effort in each API request. Omitted or null effort uses the model's default (marked in the table below). Use "none" for a direct answer: this is useful when you want a short response without a separate thinking phase.
Qwen and Nemotron use native effort controls, which do not promise a fixed number of thinking tokens.
| Level | reasoning_effort | Qwen3.8 27B: reasoning effort | Nemotron 3.5 Super VL: reasoning effort |
|---|---|---|---|
| Off | none | Off | Off |
| Low | low | Low | Not supported |
| Medium | medium | Medium (default) | Medium (default) |
| High | high | High | High |
Higher effort does not guarantee a better answer.
Qwen and Nemotron ignore client template controls; choose the effort through reasoning_effort. Supplied legacy settings must still have valid types. Unsupported efforts and invalid types return 400.
Reasoning can help with complex questions, but takes longer and uses more tokens. It does not guarantee a correct answer.
Direct answer:
curl https://api.voracompute.com/v1/chat/completions \
-H "Authorization: Bearer $VORA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.8-27b",
"messages": [
{
"role": "user",
"content": "Explain an API key in one sentence."
}
],
"max_tokens": 256,
"reasoning_effort": "none"
}'Thinking enabled:
curl https://api.voracompute.com/v1/chat/completions -N \
-H "Authorization: Bearer $VORA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.8-27b",
"messages": [
{
"role": "user",
"content": "Compare two approaches to retrying a failed HTTP request and explain when each is appropriate."
}
],
"max_tokens": 2048,
"reasoning_effort": "medium",
"stream": true,
"stream_options": {
"include_usage": true
}
}'The Python and Node SDKs accept reasoning_effort directly. The examples print reasoning to standard error and the final answer to standard output. A client's effort control must send this field; see Cherry Studio setup.
Response fields
- With
stream: true, reasoning arrives inchoices[0].delta.reasoning_content, and the final answer inchoices[0].delta.content. - With
stream: false, the complete response containschoices[0].message.reasoning_contentandchoices[0].message.contentseparately. - A model may answer directly even when thinking is enabled. Treat
reasoning_contentas optional: it is omitted when none was generated. Tool calls remain intool_callsand usefinish_reason: "tool_calls". - For Qwen and Nemotron, preserve returned
reasoning_contentseparately in assistant history for the next turn. This supports continuity and reuse of the conversation prefix. Never copy reasoning intocontent. - EuroLLM 22B returns no
reasoning_content, and drops any it finds in the history, such as turns written by another model.
Tokens, timing and cancellation
max_tokens limits the combined reasoning, closing tokens and final answer. Qwen and Nemotron have no separate numeric thinking budget. The total output limit applies to every model: a short limit can end during reasoning with finish_reason: "length" and no final answer. Increase the output limit within the model's ceiling or turn reasoning off and retry.
Very small limits can still produce no final answer.
usage.completion_tokens includes reasoning and the final answer. Neither paid plans nor pilot grants impose a daily token cap. TTFT measures the first generated token, including reasoning; it is not the wait until the final answer begins. If a request is interrupted before final usage arrives, recorded token counts are an approximate lower bound.
The deadlines are described below. Streaming is recommended for longer replies. Stop in Chat, or abort the HTTP request in your client, to cancel generation. Chat preserves any text already received; without streaming, a cancelled request cannot return partial output.
The interface stops immediately, but cancellation can take time to reach the server. Tokens may still be generated and the request may occupy a concurrency slot until the server receives it. Release checks observed a delay of about 30 seconds; this is an observation, not a guaranteed cancellation time.
Changing Reasoning in Chat applies to the next message. The reasoning panel starts collapsed and can be opened while the model is thinking.
GET /v1/models keeps max_output_tokens for direct answers and reports reasoning.supported, reasoning.default_enabled, reasoning.max_output_tokens, reasoning.supported_efforts and reasoning.default_effort.
Qwen long context
Qwen is available with the limits below.
The context window is 1,000,000 tokens, including the exact rendered input and reserved output. Output defaults to 8,192 tokens and may be raised to 32,768. The complete JSON request must also fit within 15 MiB. Oversized requests are rejected without truncation.
Each account can have 2 admitted requests in the Qwen H100 pool, all API keys and Chat, including preparation, waiting and generation. At most 1 can reserve more than 262,144 input-plus-output tokens. Open conversations that are not generating do not occupy a slot. Exceeding either account limit returns 429 with Retry-After; unavailable model capacity returns 503.
Waiting Qwen requests use one FIFO across its replicas. There are at most 4 waiting requests in the Qwen H100 pool. A large request at the head is not bypassed by smaller requests. A request leaves the queue with 503 if it cannot start within 30 seconds. Once inference has been sent, Vora does not automatically repeat it on another engine.
Long inputs can require substantial preparation before a first token. Qwen allows up to 30 minutes before the first token and 60 minutes in total, measured from request reception. These are timeout bounds, not expected response times. Configure the client and any intervening proxy for at least that total duration; allowing one extra minute avoids ending the connection before Vora reports its deadline. A client idle timeout must also allow the prefill wait. The SDK examples generated for this model include the extended timeout.
Conversation prefixes may be reused in volatile GPU memory within the same account. Accounts never share prefix-cache entries, and cache contents are not written to disk. Vora does not persist prompts, images or completions.
Nemotron long context
Nemotron is available with the limits below.
The context window is 1,000,000 tokens, including the exact rendered input and reserved output. Output defaults to 8,192 tokens and may be raised to 32,768. The complete JSON request must also fit within 15 MiB. Oversized requests are rejected without truncation.
Each account can have 2 admitted requests in the Nemotron pool, all API keys and Chat, including preparation, waiting and generation. At most 1 can reserve more than 262,144 input-plus-output tokens. Open conversations that are not generating do not occupy a slot. Exceeding either account limit returns 429 with Retry-After; unavailable model capacity returns 503.
Waiting Nemotron requests use one FIFO across its replicas. There are at most 4 waiting requests in the Nemotron pool. A large request at the head is not bypassed by smaller requests. A request leaves the queue with 503 if it cannot start within 30 seconds. Once inference has been sent, Vora does not automatically repeat it on another engine.
Long inputs can require substantial preparation before a first token. Nemotron allows up to 30 minutes before the first token and 60 minutes in total, measured from request reception. These are timeout bounds, not expected response times. Configure the client and any intervening proxy for at least that total duration; allowing one extra minute avoids ending the connection before Vora reports its deadline. A client idle timeout must also allow the prefill wait. The SDK examples generated for this model include the extended timeout.
Conversation prefixes may be reused in volatile GPU memory within the same account. Accounts never share prefix-cache entries, and cache contents are not written to disk. Vora does not persist prompts, images or completions.
EuroLLM 22B
EuroLLM 22B was built by European universities and companies, funded by the European Union and trained on the MareNostrum 5 supercomputer. It writes and translates in 35 languages, including every official EU language, Catalan and Galician. It accepts text only, has no thinking mode and does not call tools.
EuroLLM is available with the limits below.
The context window is 32,768 tokens, including the exact rendered input and reserved output. Output defaults to 4,096 tokens and may be raised to 16,384. Oversized requests are rejected without truncation.
One replica serves EuroLLM. Up to 4 requests generate at once, and together they reserve at most 32,768 input-plus-output tokens, so a request with a large max_tokens can wait for others to finish. Up to 4 requests wait in its FIFO; a request leaves the queue with 503 if it cannot start within 30 seconds. Each account can have 2 admitted requests across all models.
Conversation prefixes may be reused in volatile GPU memory within the same account. Accounts never share prefix-cache entries, and cache contents are not written to disk. Vora does not persist prompts or completions.
Image input
With Qwen and Nemotron, a user message can contain text and image_url parts; EuroLLM 22B accepts text only. Each image must be a base64 data URL containing JPEG, PNG or WebP. Remote URLs are never downloaded. Limits apply across the complete request, including conversation history: 8 images and 10 MiB of decoded data. Qwen scales images larger than 2 megapixels down and uses up to about 2,048 tokens per image. Nemotron scales images larger than about 3.4 megapixels down and uses up to about 3,300 tokens per image.
{
"model": "qwen3.8-27b",
"messages": [{"role": "user", "content": [
{"type": "text", "text": "Describe this image."},
{"type": "image_url", "image_url": {"url": "data:image/png;base64,BASE64_IMAGE"}}
]}],
"reasoning_effort": "medium",
"max_tokens": 1024
}
Not supported
The Responses API (/v1/responses) and Anthropic Messages API (/v1/messages) are not available. Requests to those endpoints return 404. See agent compatibility before configuring a client.
These return 400 with code unsupported_parameter or unsupported_content:
ngreater than 1logit_biasaudio, andmodalitiesother than text- Video and audio
Other request parameters are ignored. Custom templates, reasoning budgets and reasoning-format overrides are not passed through; Vora sets these internally.
Use /api/status?model=MODEL_ID for each model's current availability. A deployed model can still be temporarily unavailable or at capacity.
Decisions
GLiDE 9B, by Fastino, answers questions about a text with a probability for each answer: yes or no, one of several options, or a level on a scale. It answers each question in one pass and never generates text, so answers usually take well under a second. It is not a chat model and has its own endpoint, the same as Jev's System One API: clients built for Jev work with Vora's base URL and an API key.
Request
POST https://api.voracompute.com/v1/systemone
curl https://api.voracompute.com/v1/systemone \
-H "Authorization: Bearer $VORA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "glide-9b",
"state": "Shoes arrived two weeks late and in the wrong size. Also I see two charges on my card.",
"questions": {
"billing": { "type": "noul", "instructions": "Is there a billing problem?" }
}
}'model: glide-9b (glide-latest and jev-latest are also accepted).
Response
{
"model": "glide-9b",
"answers": {
"billing": { "type": "noul", "noul": 0.979 }
},
"usage": { "input_tokens": 77, "output_tokens": 0, "reasoning_tokens": 0 },
"latency_ms": 310
}Test thresholds on your own data before relying on them.
Questions
noul: yes or no.criteriamay describetrueandfalse.choice: one option.criteriais an object of 1 to 255 options, each with a description ornull.score: a level.criteriais a list of 1 to 255 levels, in order.
score is the expected level, from 0 to the number of levels minus 1; probabilities gives each level.
Options
GLiDE has no reasoning options; an options object sent by Jeeves clients is accepted and ignored.
Limits
- Up to 4 questions per request.
- The state with each question fits in 16,384 tokens.
- 2 requests at a time per account.
- At most 255 options or levels per question.
- Each question reads the whole state, so time grows with both: a short text takes well under a second; a state of 16,000 tokens takes about 10 seconds per question.
- Evaluated by Fastino in English. Vora also checked it on Spanish test sets.
License
Built with Fastino GLiDE.
Fastino Community License: free for organizations with under US$10 million in annual revenue. Larger organizations need a commercial agreement with Fastino to use it in production.
Where
GLiDE 9B runs in the EU. When that replica is full or down, a backup in the United Kingdom answers. The API stores usage metrics only, never states, questions or answers. See Privacy.