Changelog
Changelog
Releases and notable platform changes, newest first.
Last updated: 2026-11-03
What changed, when, and why. Newest first.
Between 24 May and 22 September 2026 several changes were not logged here, among them audio, embeddings and rerank, images as input and the Responses API. The pages they affect describe how they work today.
2026-10-25 — New platform
The site, the dashboard and the API moved to a new platform on our machines. Each of them now runs on several machines at once, and the database is kept on three: a machine that stops no longer stops the service. Addresses, keys, settings and documents stay as they were.
The move interrupted the service for a few minutes on 25 October, at different times for each site, between 11:44 and 14:41 (Swiss time). During an interruption the API answered 503; those requests were not charged. No data was lost.
2026-10-24 — Models by hardware, and larger documents
Models and hardware. The new Models page of the dashboard lists each model on each type of hardware that serves it: the name to use in the API, the speed we measured with one request at a time, the price per million tokens there and the country. The model's name alone lets us choose among the machines that serve it, and the Usage page says, request by request, which hardware answered. The model's name followed by @ and a hardware's name, for example ministral-3-14b@spark, runs only on that hardware; if none of it is free, the request is refused with 503 and not charged.
Keys by country. Where we have machines in more than one country, the API keys page lets a key be limited to the machines of one of them. A model that does not run there answers with an error that says where it runs.
Qwen3-Coder-Next. The catalogue has qwen3-coder-next, Qwen's newest model for programming. It has 80 billion parameters, of which about 3 billion work on each token, and it is served in FP8. Its context is 262,144 tokens, and it calls tools and follows a JSON schema. It answers about as fast as qwen3-coder-30b and does better on changes across a whole project.
Large documents in knowledge bases. Uploading a large document could fail while it was being indexed: the passages went to the embedding model 32 at a time, and under load a batch took longer than the 60-second limit. They now go 8 at a time, and the total time is the same. A document left half-indexed by a restart is marked as failed after 40 minutes, so it can be uploaded again.
2026-10-24 — Refused requests, explained
Refused requests in the dashboard. The dashboard's home and the Usage page now list the requests refused in the last 7 days, grouped by key, model and reason, with the most recent first. Each group says what happened and what to do about it. Where the message has numbers, the group uses them, for example the prompt's tokens and the model's context. It also shows how many times the refusal happened, when it last happened and the ID of the latest request, which you can quote to support. The home shows the five latest groups and links to the full list on Usage. Refused requests are not charged.
- What is counted: every
4xxanswered to a valid key, such as a wrong model name, a parameter that does not validate, the rate limit, credit running out, the key's spending limit or a refusal from the model's engine. A429or a402is counted at most once a minute per key, so a client that keeps retrying past a limit fills one line, not thousands. - A request that failed on our side and was not answered by another machine is listed too, and says that the error was ours and the request was not charged. A request that failed on one machine and was answered by the next one is not a refusal and is not listed.
- When a model's engine refuses a request for a reason we recognise, the error body of the API carries
error.hint. Today that covers a prompt longer than the model's context and a JSON schema the engine cannot use. The text of the hint is the same advice the dashboard gives. See Errors.
2026-10-14 — Reasoning control, a spending limit per key, a page for each model
The changes of 7 to 11 October, released on 14 October.
Reasoning. Chat completions accept reasoning_effort and OpenRouter's reasoning object. Until now they were refused as unknown fields. On gpt-oss-120b they set its level, and none gives low, because this model cannot stop reasoning. On the Qwen3 models none switches reasoning off, and any other level leaves it on, as without the field. On models that do not reason, the field changes nothing. chat_template_kwargs is still refused, and the error now points to reasoning_effort. Answers of qwen3.8-27b now start with their first character: two blank lines that the model sends after its hidden reasoning were passed on as the start of the answer. See Every parameter.
A spending limit per key. The API keys page of the dashboard shows what each key has been charged this month. Each key can be given a monthly limit and an output allowance per request. Before a chat completion reaches a model, its highest possible cost is reserved: the prompt's tokens at the input price plus the allowance at the output price, counting the key's requests that are still running. If that does not fit in what is left of the limit, the answer is 402 with type and code spending_limit_reached, and no model runs. On a key with a limit, a request that sets no max_tokens gets the allowance as max_tokens. The default allowance is 8,192 tokens. The month starts on the 1st, in UTC. See Errors.
Models. Models catalogue is now a table for comparing models. It shows countries as flags, context, tool calls, images, reasoning and price, and filters by country and by what a model can do. It shows countries, never zone codes. Each model has its own page, /live/models/<id>, with what we serve: publisher, architecture, licence, precision, context, longest answer, what it reads, tool calls, JSON schema, reasoning, countries, status, price, a request to copy and its limits. Where we have not tried something yet, the page says so. Service status shows each chat model by country.
GET /v1/models is built on every call from what the brand serves. It lists the chat models of the catalogue, bge-m3, the reranker, picture and video models while a machine makes them, whisper-1, whisper-1-hd and tts-1. The voices are no longer listed as models: they are the voice of tts-1, at GET /v1/audio/voices. tts-1-hd is still accepted but not listed, and owned_by is the brand's name. Each chat model also carries context_length, max_completion_tokens, its input modalities, supported_parameters, pricing, countries, license and precision, with OpenRouter's field names. The OpenAI SDKs ignore fields they do not know.
Integrations. A new section, Integrations, explains what any OpenAI-compatible tool needs here and what the API does not do. It has a page for each of these tools: Cline, Continue, Roo Code, Kilo Code, OpenCode, Aider, Zed, Open WebUI, n8n, LangChain, LlamaIndex, LiteLLM, Vercel AI SDK and the OpenAI Agents SDK. Six of them were run against the live API on 11 October, and their pages show the version and what came back. The pages for the others say that their settings come from the tool's own documentation.
Playground and chat. The playground in the dashboard has been rewritten. It is translated, and it lists models with the flags of their countries. It offers reasoning levels where the model honours them. After each call it shows the model that answered and its country, the tokens, the time and the cost. It shows the same request in curl, Python and JavaScript on the brand's own API address: before, the example went to siati.ai's address on every brand. The tier menu, which the playground did not use, is gone. In the web chat the menu has one option per model and country, with the country's flag, and the request stays on that country's machines. Code blocks are readable again: their lines showed as white bars. Amounts in the dashboard follow the page's language.
Fixes in the API.
- An engine that refuses a request, for example because the prompt is over the context or a schema cannot be compiled, gets a
400 invalid_request_errorwith the engine's reason. No other machine is tried, because it would refuse too. Before, this came back as502, which looked like an outage. toolssent to a model served without tool calling (apertus-70b-instruct) get a400withparam: toolsthat lists the models that call tools. Before, this was also a502.- A
422from the validation of chat completions now also carries OpenAI'serrorobject, withmessage,type,paramandcode. The SDKs showed it as an unknown error.messageanderrorsare still there. - With
stream_options: {"include_usage": true}, the last chunk before[DONE]has an emptychoicesand theusageof the request, which are the counts that are billed. The option used to be accepted with no effect. See Streaming. - On
/v1/responses,function_callandfunction_call_outputitems sent back ininputare read as the model's tool call and its result. Before, they became empty user messages, and the model asked for the same tool again. n8n and the OpenAI Agents SDK use this API by default. - After a machine fails to answer in time, the next requests try it last for five minutes. After a
5xxor a refused connection, they try it last for one minute. On 6 October an engine hung while its health check still answered, and every request waited the full timeout before moving to the next machine. - Zone codes now give only the country and a number, such as
CH-1orIT-2, andX-Zonasends them since 8 October. Usage records made before that keep the code they had. /v1/rerankis no longer charged. Where a brand makes video, a request withoutseedgets a random one: before, the same description made the same clip every time. A failed video job answers with a sentence, without the engine's traceback.POST /api/v1/me/deletereturns acancel_urlthat opens: it pointed to the API's host, which answered404.
2026-10-06 — Transcription in the zones, and the documentation checked against the code
Transcription in the zones. A whisper-1 transcription now runs on a transcription machine in one of the zones of the key, where the brand has one, and the response carries X-Zona with that zone. The rest runs on the server in Switzerland, outside the zones, as before, and its responses have no X-Zona: whisper-1-hd, word timings, requests with a prompt, audio over 10 minutes, files over 25 MB, 3GPP and AMR, and every request when the machine in the zones is down or fails. Speech synthesis runs in Switzerland. The answers of the machine in the zones are rewritten in the shape of that server's, and phrases it invents on silence become empty text: see What we change in responses and Audio.
Fixes in the API, found while checking the documentation against the code:
- The limits per endpoint count per key, as the limit of the tier always did. They were meant to from 6 October, but on every
/v1route the limit ran before the key was read, so it still counted per address: a class behind one school address shared 60 chat requests a minute. A request with a wrong key is refused before it is counted, and the app API counts per signed-in user. See Rate limits. - Errors of the developer API are in English on every brand. Some were still in Italian: a path that does not exist, a model on request, an image the model cannot take, a refused or misspelt parameter, no machine in the zones. The model on request also linked the same contact page on every brand: it now names the brand's own support address. Where the brand's language is Italian, a
422saidvalidation.requiredinstead of a sentence. See Errors. - Errors of the app API (
/api/v1) are in English on every brand too, with the same codes. Sign-in, tokens, sign-up, audio, billing and purchases answered in siati.ai's Italian on every brand, the school's English one included, and a disabled account was told to contact support without an address: it now gets the brand's. A conversation, knowledge base, document or invoice that does not exist got404 unknown_endpointon the API's own host, as if the path did not exist: it now gets the same404as on the dashboard's host, with a message that names what is missing. The fixed answers of a knowledge base, when nothing is ready or nothing matches, are in English on both APIs: three of them were in Italian on every brand. See Authentication and Knowledge bases API. /v1/completionswithoutstreamand/v1/responsesnow sendX-Zona, as chat completions do: they rewrite its answer into their own shape, and the header was lost on the way.nabove 1 withstream: trueis refused with400,code: invalid_value,param: n. A stream carries one answer: the pieces of several arrived glued together, and all of them were billed. Withoutstream,nworks as before.- The
userfield of a streamed request is stored with its usage record, as it always was without a stream. - Models catalogue shows the price of embeddings and rerank as it is charged, per million input tokens, the same as Pricing. It showed rerank per 1,000 searches, a unit the charge does not use.
- Models catalogue says what a model started on request answers until it is started:
503withmodel_activation_required, at once, and who to write to. It said the model was loaded on the first call, with a wait. Where the pages are in English, the precision of each model is in English too. - On the models served by Ollama, function calls arrive in OpenAI's shape, with and without a stream, and a call sent back on the next turn reaches Ollama in the shape it reads;
response_formatapplies in a stream too,{"type": "text"}gives free text instead of JSON, andn: 1is accepted. See What we change in responses. - The arguments of a function call keep their empty objects: a call without arguments had
"arguments": "[]"instead of"{}", which the OpenAI SDKs cannot unpack.
Corrections. An outside review of the documentation found pages that contradicted each other, or promised more than the service does. Each one was checked in the code and with requests to the API, and corrected:
- Knowledge bases: an API key lists them and asks them questions; creating them and uploading documents need a signed-in user. Every page now says so, with a table in Knowledge bases API.
- The OpenAI migration guide no longer says that image generation does not exist: it lists pictures and video, for the brands that have the machines.
- Models no longer copies the catalogue, which had moved on: it says how to read it, and the models with the images badge are not only Gemma.
- Every parameter said that nothing is accepted and ignored. Four parameters are, in a stream, and two are not applied on the models served by Ollama: the page lists them.
/v1/responsesignores the fields it does not know (Responses API). - Structured output promised valid JSON every time. A schema constrains the shape of a complete answer, not its values, and an answer cut at the token cap is not an object: the page now lists the checks to make, in order.
- The streaming examples now tell a complete answer from one cut at the token cap, an error and a broken stream (Streaming in TypeScript).
- Pricing listed the chat models only. It now lists every service — embeddings, rerank, knowledge bases, transcription, speech, and pictures and video where a brand makes them — with each price read where the charge reads it, and says what is not charged. Audio prices are back: since 3 October the charge and the price list read the same value.
- Rate limits says what is still counted per address, why that is twenty times wider on a school's own networks, and what a student sees when a limit is reached.
- Three examples, with measured numbers says in its introduction and in its conclusions that the document-search example was measured on the pipeline, not through the public API.
New pages.
- What works where: every endpoint on one page, with its modes, models, inputs, the parameters it honours, refuses or accepts without effect, its credential, its zones and what it bills. The other pages link to it.
- AI-generated media: which mark goes into pictures, video and voices, how to read it, what removes it, and why a visible note is still needed where you publish.
2026-10-03 — Zones, fallback between machines, brands, documentation in English
Zones. A zone is a group of machines in one country, with a declared owner. Every API key has the zones it may use: by default, all the zones its brand offers.
- A text generation request goes only to machines in the zones of its key. If none is available there, the answer is
503and names the zone: the request is never moved to another zone. - Chat completion responses say where they were processed, in the
X-Zonaheader, and the usage records keep the zone and the site of every request. - Only text generation goes through zones for now. Embeddings, rerank and audio do not yet.
- A new machine first enters a test class, where only internal keys reach it, and serves customers after that.
Fallback between machines. If the chosen machine fails before the first token, the request moves to the next one, always inside the zones of the key. Until now the answer in that case was a 502.
Control plane and data planes. The platform is now split in two: a control plane in Switzerland, with the inventory of the machines, the catalogue and the zones; and one data plane per brand, in the brand's country, with accounts, keys, usage records and files. A brand downloads its configuration every minute and, once an hour, sends back only aggregated numbers: calls, errors, tokens and average time per model, machine, endpoint and zone, with no users, keys, texts or IP addresses. siati.ai is the first brand running this way. See Sovereignty.
Brands. The same platform serves several brands, each with its own name, domains, users and payments: by card, or with credits loaded by an administrator. A brand can open sign-up to everyone, or by request: the person asks for access with an address of an allowed domain, confirms it through a signed link, and an administrator approves. The request page does not reveal who already has access.
Documentation. The documentation is moving to English, one page at a time, and fills in the name, the addresses and the data country of each brand. A brand can replace a page with its own. The live pages — catalogue, service status, pricing — are in English: the catalogue shows the zones and the precision of every model, and the pricing page shows the cost formula the code applies, for brands that take payment by card and for those that load credits by hand.
Corrections. Pages that promised more than the service does were corrected:
- A tier does not give priority in a machine's queue. It chooses which machines may serve you, the rate limit of your key and the price multiplier. See Tiers.
- The documentation no longer names models that are not served, such as Mistral Large 2, Qwen 2.5 72B and DeepSeek-R1 Distill, added on 24 May.
- The web chat no longer shows the name and port of a machine in its error messages.
Contents of requests are no longer kept. The request log now keeps only numbers — endpoint, model, tokens, time taken, outcome — for 30 days. The text of requests and answers is kept only for the keys where you turn on "keep messages" in the dashboard, for debugging, encrypted, for 30 days. What the log held before was deleted on 3 October.
Credits. Every path that calls a model now records its usage: the web chat, the playground, knowledge bases (questions and indexing), the chat sessions of the app, and streams the client closes early (with tokens estimated from the text). On brands with prepaid credits each request takes its cost from the balance, and at zero paid calls stop with 402 insufficient_credits.
Fixed.
/v1/completionswith a JSON body answered500. It now answers.- Transcription: a recording that started with a few seconds of silence came back as empty text, because the silence check read only its beginning. It now reads the whole file, and declares silence only after reading all of it. A silent file with
response_format=textanswered500; it now answers with empty text. - After the move to the new machines, earlier on 3 October, uploads over 2 MB were refused. The limit is back to 50 MB per file.
- On brands where access is by request, sign-up through the app API is now refused: access is asked for on the brand's site.
- Every request now has a billing reference generated by us. The
X-Request-Idyou send stays in the logs and in the response, for correlation only.
2026-09-22 — The parameter list, and two false limits removed from the developer page
Reported by someone evaluating the service: the developer page of the site still listed as limits two things fixed on 18 August. They were right, and the page promised more than had been kept: "the list is updated at every release" and "the fix will be followed by its date in the changelog". This entry is that date, a month late.
No longer limits, and removed from the list:
response_format—json_objectandjson_schemawork, also withstream: true. Verified today through the public gateway: a schema withnumeroandtotalereturns{"numero": "2026-114", "totale": 1250.00}, keys and types respected. (Correction, 3 October 2026: on theqwen2.5models,response_formatis applied only without streaming.)top_p— forwarded to the engine. Verified today: attop_p: 0.01with temperature 1.5 the answer is the same three times out of three; at1.0it changes.
What had happened. Until 18 August the validation of the request kept only the fields listed in its rules and silently dropped everything else — top_p included, which was even in the table of this documentation. Fixed at the root: the list of parameters lives in one place, and an automated test fails if a field is validated without reaching either the engine or the list of refused ones. A parameter we do not recognise now gives 400 and is named.
Full list: Every parameter: honoured or refused.
Real limits that remain: max_tokens at 8192; no Assistants API, Realtime, Batch or fine-tuning; no SOC 2 Type II. And schemas constrain the shape of an answer, not its truth.
2026-05-24 — Catalogue expansion and chat improvements
Catalogue — three open-weight models added:
- Mistral Large 2 (123B) — reasoning, code, and many languages, Italian included.
- Qwen 2.5 72B Instruct — the best dense open model of its size when it was added.
- DeepSeek-R1 Distill Llama-70B — a reasoning model that shows its chain of thought in
<think>…</think>blocks.
Llama 3.1 405B removed: in its 4-bit quantised form it lost too much quality on instructed tasks.
Web chat:
- Answers stream token by token, instead of appearing all at once at the end.
- Markdown and math formulas are rendered while the answer arrives.
- The page scrolls with the answer, and stops following it while you scroll up.
- A Stop button interrupts the generation and keeps the partial answer.
- Choosing a model selects the fastest tier available for it.
- Every message shows the time to the first token, the total time and the tokens per second.
- Conversations get a title, generated in the background from the first message.
Chat sessions API: POST /api/v1/chat/sessions/{id}/messages streams the same chunks as the web chat. See Chat sessions.
Outgoing mail moved to a dedicated relay, with authenticated sending and spoofing blocked on the relay.
2026-05-20 — Better retrieval for RAG: reranker, Docling, hybrid search
Three pieces, each for a different weakness of plain vector search:
- Reranker —
bge-reranker-v2-m3. The pipeline takes 30 candidates from the vector store, a cross-encoder scores each of them against the question, and the best 5 are kept. - Docling replaces
pdftotextas the main PDF reader: it keeps tables, the reading order of pages in columns and the captions of figures, and reads scanned pages with OCR.pdftotextstays as the fallback. - Hybrid search — every collection holds dense vectors (BGE-M3) and sparse ones (BM25), combined with Reciprocal Rank Fusion. It finds exact terms — codes, names, acronyms — that dense vectors tend to bury.
Architecture and settings: RAG.
The knowledge bases created before this change were deleted: the old index could not hold both kinds of vectors.
2026-05-19 — RAG goes live, more machines, documentation rebuilt
RAG
- The RAG service is live in the dashboard and in the API.
- Vector store: Qdrant. Embeddings: BGE-M3, multilingual, 1024 dimensions, on a dedicated machine.
- Answers: Apertus 70B by default; any model of the catalogue can be chosen per request.
/dashboard/rag, for those who do not write code: upload PDF, DOCX, Markdown or text files, ask questions, get answers with citations.- API:
POST /api/v1/rag/kb,POST /api/v1/rag/kb/{slug}/docs,POST /api/v1/rag/kb/{slug}/chat.
Machines and routing
- More machines: for embeddings, for Apertus 70B and for the small models.
- Requests are routed by tier and model across several machines: by the weight of each machine for the tier, then by the shortest queue.
- Two inference engines, interchangeable behind the same API.
- A health check every 30 seconds.
- Machines are added and removed from the administration, without a release.
Site and documentation
- The copy of the site rewritten.
- Pricing page: plans billed monthly or yearly, and top-ups in five amounts from 5 to 500 CHF, with an estimate of the tokens they buy.
- This documentation rebuilt from scratch: several pages, search (⌘K), live pages read from the database, copy buttons on code blocks.
Pricing
- Four tiers:
slow,medium,fast,ludicrous. The separate tier for embeddings is gone. - The per-tier surcharge setting was removed: a latent bug in it doubled costs.
- Prices have a single source in the database.
2026-05-17 — Production platform live
- siati.ai is live: site, dashboard (sign-up, billing, playground, API keys), documentation (a single page at first), chat and API, one application on several domains.
- API for the mobile app: sign-in, chat sessions, in-app purchases, billing.
- Payments with Stripe, for subscriptions and top-ups; verification of App Store and Google Play purchases.
2026-05-15 — Backend rebuild started
- The backend is rebuilt from scratch.
- Existing accounts and API keys are carried over from the previous backend: passwords and keys keep working.
- API keys are stored as a salted HMAC-SHA256 hash, never in clear.
- Sign-in with tokens for the mobile app.