Skip to content

Changelog

This page lists all notable changes to the CompactifAI API. Deprecations and breaking changes are announced by email at least 15 days ahead: see Deprecations.

  • Guardrails on POST /v1/chat/completions – a deployment can enable guardrails that check every chat completion before it reaches the model. A blocked request returns 403 (content_policy_violation). On a deployment that requires the check, a request it cannot run on is not sent and returns 503 (policy_check_unavailable, retryable). Deployments without guardrails are unchanged. See Error handling.
  • quasar-438b – Quasar 438B is decommissioned and removed from the model catalog. Migrate to quasar-1-2-358b. See Quasar 1.2 358B.
  • Prepaid wallet balance never goes below zero – a request that costs more than the remaining balance is charged only up to the balance. The balance stops at $0.00 instead of going negative, and the part of that request’s cost that the balance did not cover is not charged.
  • qwen-3-8-27b output price – the output price is now $0.44 per 1M tokens (was $0.04). Input ($0.09) and cached input ($0.0171) are unchanged, with reasoning on or off. See Pricing.
  • New API prices – input, cached input and output prices are lowered for glm-5-3, glm-5-2, qwen-3-8-27b, quasar-438b, quasar-1-2-358b and carina-60b. Prices per 1M tokens, input / cached input / output: glm-5-3 $0.20 / $0.018 / $0.64, glm-5-2 $0.17 / $0.0153 / $0.54, qwen-3-8-27b $0.09 / $0.0171 / $0.04 (the same price with reasoning on or off), quasar-438b and quasar-1-2-358b $0.11 / $0.0099 / $0.33, carina-60b $0.022 / - / $0.085. Cached input is billed at 9% of the input price for GLM 5.3, GLM 5.2, Quasar 438B and Quasar 1.2 358B, and at 19% for Qwen 3.8 27B. See Pricing.
  • A system message in the middle of a conversation is honoured on POST /v1/messages – a message with role system that comes after a user or assistant turn is now treated as a developer-level instruction: it is binding from where it appears and overrides an earlier user request that conflicts with it, for example “calling any tool is forbidden, reply only with: Unable to query” after a user turn that asks for a tool call. Before, the model could read such a message as untrusted content and ignore it, so the tool was still called, as in Tencent’s onboarding report. This matches how the original model vendor behaves and how POST /v1/chat/completions and POST /v1/responses already treat the caller’s system messages. A system message at the start of the conversation, and requests without one, are unchanged.
  • POST /v1/responses no longer returns empty per-turn usage arrays – usage.input_tokens_details.input_tokens_per_turn, usage.input_tokens_details.cached_tokens_per_turn, usage.output_tokens_details.output_tokens_per_turn and usage.output_tokens_details.tool_output_tokens_per_turn are removed from the response body, from the response objects inside the stream events, and from the stored copy that GET /v1/responses/{id} returns. They were always [], which is not a valid token count, and the original model vendor does not return them. The token totals (input_tokens, output_tokens, total_tokens) and the other detail fields (cached_tokens, reasoning_tokens) are unchanged.
  • POST /v1/responses names output items in one format, streamed or not – reasoning items are rs_…, function calls fc_… and messages msg_…, and the call_id of a function call starts with call_. A streamed response used ids without these prefixes, and a response that was not streamed used chatcmpl-tool-… as call_id. The same applies to the item_id of the stream events, to response.completed and to the stored response.
  • POST /v1/responses echoes previous_response_id – a response to a request that carries previous_response_id returns it in previous_response_id (and in the response of the response.created, response.in_progress and response.completed events); it used to be null.
  • A streamed tool call can be answered with previous_response_id on POST /v1/responses – in a streamed response, response.completed could name an output item (id, and call_id for a function_call) differently from the response.output_item.done event of the same item, and the stored response keeps the names from the item events. A follow-up turn that sent only a function_call_output with the call_id from response.completed returned 400 function_call_not_found. response.completed now carries the same items as the item events, so the stream and the stored response name each item the same way.
  • POST /v1/detect checks whether a text was watermarked – send a model and a text and get back watermarked and a coarse confidence (none, low, medium, high). The raw score is not returned. A model that is not available for detection returns 400 (model_not_found). See Watermark Detection.
  • Non-streamed POST /v1/responses returns 400 for an input it cannot serve – an input of only ignored items (compaction_trigger, item_reference, computer_call, …), or one with custom tool items and no user message, used to return 200 with a failed response. Without streaming it now returns 400 (invalid_request, e.g. Input cannot be empty[<request id>]), as the origin provider does. Streamed requests are unchanged.
  • More input item types are ignored on POST /v1/responses – file_search_call, code_interpreter_call, image_generation_call, local_shell_call, shell_call, apply_patch_call (and their _output items), tool_search_call / tool_search_output, compaction, program / program_output and configuration_update used to return 400 (Prefill failed). They are now dropped, as the origin provider does. An unknown item type is dropped too, instead of returning the engine’s 400 validation dump.
  • Custom tool items without a user message fail like the origin on POST /v1/responses – an input that holds a custom_tool_call or custom_tool_call_output but no user message now fails with error.code invalid_request (400 without streaming, a response.failed and an error event when streaming), as the origin provider does. With a user message anywhere in the input, before or after them, a custom tool call and its output are sent to the model as a function_call / function_call_output pair and answered.
  • include as a string no longer fails on POST /v1/responses – a non-array include (for example "include": "web_search_call.results") used to return a 400 without streaming and a 502 upstream_unavailable with stream: true. It is now ignored, as the original model vendor does.
  • The response of POST /v1/responses no longer carries backend-only fields – input_messages, output_messages, kv_transfer_params and ec_transfer_params are removed from the response body and from the response objects inside the stream events (response.created, response.in_progress, response.completed, response.incomplete, response.failed), and from the stored copy that GET /v1/responses/{id} returns.
  • The model decides when to use web_search on /v1/responses – The web_search tool is now offered to the model as a tool. The model decides whether to search; when it calls the tool, the API runs the search, gives the results back to it and the model answers, up to WEB_SEARCH_MAX_USES searches per request (default 5). Before, /v1/responses searched on every request that carried the tool and include asked for its output, using the last user message as the query, even for prompts like “Reply with OK.” A request where the model does not search has no web_search_call item and no extra input tokens, and include no longer gates the search: it only chooses whether action.sources and results are returned. The output keeps the model’s reasoning and message items between the searches, a failed web_search_call carries an error string, and the model can read a page (web_search_call with action.type: "open_page"). usage adds up every model call, reasoning tokens included. A streamed request relays the model’s events live and pauses while a search runs. See Web search. POST /v1/chat/completions keeps searching on every request with an enabled web_search tool; its stream now carries the web_search sources on the closing chunk and no request_id. POST /v1/messages is unchanged: it accepts the web_search_* server tool and does not run it.
  • container_upload content blocks are accepted on POST /v1/messages – they used to fail with a 400 (Unsupported content block type 'container_upload'); they are now dropped before the request reaches the model, as the original model vendor does. There is no container or Files API behind the block, so the file is not read. A message that holds nothing but container_upload blocks is still rejected with a 400. See File uploads to a container.
  • Images sent to a text-only model on POST /v1/responses no longer fail – For a model without image support (capabilities.supports_image is false), each input_image part is now replaced by a short text notice and the request is answered normally, so the model replies that it cannot see the image. Before, the request failed with 400 and the provider’s raw <model> is not a multimodal model error. This matches how the original model vendor handles the same request. POST /v1/chat/completions and POST /v1/messages are unchanged and still reject images for such models. See Images sent to a model without image support.

  • thinking.display: "omitted" no longer hides the thinking text on POST /v1/messages – With thinking set to enabled or adaptive and display set to "omitted", the response used to carry a thinking block with an empty thinking string (and no thinking_delta events when streaming). It now returns the thinking text, as the original model vendor does, so display is accepted but has no effect. See Thinking.

  • container is accepted on POST /v1/messages – the field used to fail with a 400 (`container` is not supported by this API). It is now accepted in any shape (a string, an object with id and skills, …) and ignored, streaming and non-streaming, as the original model vendor does: there is no code-execution sandbox, so no container is created, reused or returned. See Request shape.

  • x-api-key accepts a Bearer prefix – On POST /v1/messages and POST /v1/messages/count_tokens, the x-api-key value may be the raw API key or Bearer <key>. Both resolve to the same credential, with the same billing and rate limits. Before, the prefixed form was rejected with 401.

  • input_file parts on /v1/responses are converted for the model – The models behind POST /v1/responses accept no file content part and rejected input_file with unsupported_content. The gateway now reads the file itself: PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx, first 1,000 rows per sheet) and text files are sent as text, and images (png, jpeg, gif, webp) as input_image. file_data, a data: file_url and file_id are handled, with OpenAI’s limits of 50 MiB per file and per request. A remote http(s) file_url is not supported and returns a 400 unsupported_file_url, and PDF page images are not sent, so scanned PDFs without a text layer return a 400 file_has_no_text. Other types and files over the limits return a 400 that names the problem. Safeguards are configurable (RESPONSES_FILE_MAX_PDF_PAGES, RESPONSES_FILE_MAX_CHARS, RESPONSES_FILE_MAX_PARTS, RESPONSES_FILE_MAX_CONCURRENT_EXTRACTIONS); when every extraction slot stays busy a request gets a 503 file_processing_busy. See OpenAI Compatibility.

  • top_p: 0 is accepted on chat completions and responses – POST /v1/chat/completions and POST /v1/responses now accept top_p values in [0, 1] and treat 0 as a request for greedy sampling: the gateway forwards temperature=0 and top_p=1 so the output is deterministic, taking precedence over any temperature sent in the same request. Before, chat rejected 0 with a validation error and responses forwarded it to the provider, which refused it. POST /v1/messages keeps its existing behavior of translating zero into top_k=1.

  • prompt is accepted on POST /v1/responses – any prompt (a template reference with id, version and variables, or a plain string) used to fail with a 400 from the model provider (prompt template is not supported). It is now dropped before the request reaches the model, streaming and non-streaming, and the request is answered from input and instructions. Templates are not resolved. See Responses.

  • background: true is accepted on POST /v1/responses – it used to fail with a 400 from the model provider; the response is now generated synchronously and returned in the same call.

  • Newer fields in /v1/responses responses – POST /v1/responses no longer returns 502 with error.code upstream_unavailable when the model’s response contains values newer than the API’s schema, such as reasoning.effort max or tool types computer, namespace and tool_search. Those responses now return 200 unchanged, and GET /v1/responses/{response_id} returns stored responses that contain them. A real provider outage still returns 502.

  • compaction_trigger input items are accepted on POST /v1/responses – they used to fail with a 400 from the model provider (Unsupported input item type); they are now dropped before the request reaches the model, so the rest of the input is answered normally. The conversation is not compacted. A request whose input holds nothing but ignored items returns 200 with a response whose status is failed and whose error.code is invalid_request (Input cannot be empty[<request id>], where <request id> is the response id without its resp_ prefix); with stream: true it arrives as a response.failed event followed by an error event, as the origin provider does.

  • item_reference input items are accepted on POST /v1/responses – they used to fail with a 400 from the model provider (Unsupported input item type); they are now dropped before the request reaches the model. The referenced item is not looked up: send its content inline or chain turns with previous_response_id.

  • Custom tool items are accepted on POST /v1/responses – custom_tool_call and custom_tool_call_output input items used to fail with a 400 from the model provider (Unsupported input item type); they are now sent to the model as a function_call / function_call_output pair, with the free-form input as {"input": "..."}. A custom_tool_call without a name returns 400 with error.code custom_tool_call_missing_name.

  • Replayed web_search_call and computer-use items are accepted on POST /v1/responses – they used to fail with a 400 from the model provider (Unsupported input item type). A web_search_call now reaches the model as an assistant message naming its queries and sources; computer_call and computer_call_output items are dropped.

  • Web search on /v1/responses – POST /v1/responses now runs a web search when the request has a web_search tool and include has web_search_call.action.sources and/or web_search_call.results. The response starts with a web_search_call output item carrying the search query and, depending on include, the result URLs (action.sources) and the results (results: title, snippet, url); the model answers from those results. Streaming adds the response.web_search_call.in_progress, .searching and .completed events. A web_search tool without those include values does not run a search.
  • Unsupported include values on /v1/responses – POST /v1/responses no longer returns 400 when include contains a value that is not supported. Unsupported values are ignored and have no effect on the response, with stream on or off. Supported values (code_interpreter_call.outputs, computer_call_output.output.image_url, file_search_call.results, message.input_image.image_url, message.output_text.logprobs, reasoning.encrypted_content) are unchanged.
  • x-api-key authentication on the Anthropic Messages endpoints – POST /v1/messages and POST /v1/messages/count_tokens accept the API key in the x-api-key header, as Anthropic SDKs send it, in addition to the existing Authorization: Bearer scheme. Same key, same billing and rate-limit identity. Send exactly one of the two headers: a request carrying both is rejected with 400, even when both are valid.

  • quasar-438b is back – Quasar 438B is served again and re-listed in the models catalog alongside Quasar 1.2 358B. Pricing stays $0.60 per 1M input tokens and $1.80 per 1M output tokens; prompt caching is not enabled for this model.

  • quasar-358b renamed to quasar-1-2-358b – The Quasar 1.2 358B model id is now quasar-1-2-358b (previously quasar-358b, and quasar-2-358b before that). Prices and capabilities are unchanged. Update the model field in your requests.
  • Disabling reasoning on /v1/responses – POST /v1/responses now refuses a request that disables reasoning on a model whose reasoning is always on, for example glm-5-3, with 400 and error.code reasoning_not_disableable, as POST /v1/chat/completions and POST /v1/messages already do. Before, the request reached the provider and failed with 502.
  • watermarking request field – POST /v1/chat/completions, POST /v1/completions and POST /v1/responses now accept watermarking (boolean). Chat and completions requests previously dropped it before reaching the model, so watermarking: false had no effect. Watermarking is on by default (true) when the model is configured for it. false opts the request out, and is only accepted in US deployments (us-*); in EU deployments it is refused with 400 and code: watermarking_opt_out_not_allowed. See the API reference.

  • Total token rate limits – Accounts can now be limited on input and output tokens combined, per minute, per hour, per day and over a rolling 30 days. Each request counts its prompt tokens plus its completion tokens against the total. Input and output limits are unchanged and still apply separately. See What counts against your limits.

  • thinking: {"type": "disabled"} on /v1/messages – POST /v1/messages no longer refuses thinking: {"type": "disabled"} on every model. It now turns reasoning off on models that can, as POST /v1/chat/completions does, and the response carries no thinking blocks. Models whose reasoning is always on, for example glm-5-3, still refuse it with 400 and error.code reasoning_not_disableable. Only thinking controls reasoning on this endpoint; other reasoning fields are ignored, not rejected.
  • Empty tools array accepted – tools: [] on POST /v1/chat/completions now returns 200, the same as omitting tools and as OpenAI. It previously returned 400.
  • thinking.clear_thinking defaults to true on /v1/chat/completions – Reasoning that a client sends back on earlier assistant turns is now dropped from the prompt unless thinking.clear_thinking is false, including when the request has no thinking object. Before, such a request kept the earlier reasoning, which added prompt tokens. Set thinking: {"clear_thinking": false} to keep it.

  • Image size limit – POST /v1/chat/completions (image_url) and POST /v1/responses (input_image) now reject requests whose inline images total more than 256 MB decoded with 400 and code: image_too_large, instead of failing with a 502/503 connection error. See Image size limit.

  • thinking.clear_thinking defaults to true on /v1/chat/completions – Reasoning that a client sends back on earlier assistant turns is now dropped from the prompt unless thinking.clear_thinking is false, including when the request has no thinking object. Before, such a request kept the earlier reasoning, which added prompt tokens. Set thinking: {"clear_thinking": false} to keep it.

  • quasar-1-2-358b – Quasar 1.2 358B chat model with tool calling, structured output, the Responses API, and the Anthropic Messages API. See Quasar 1.2 358B and the models catalog.
  • Video input for qwen-3-8-27b – capabilities.supports_video is now true: send video as a standard file content part in chat completions. See Video understanding.
  • Prompt caching for glm-5-2 – Input tokens served from the cache are billed at 20% of the input price ($0.22 per 1M tokens, versus $1.10 per 1M for non-cached input). See Pricing.
  • Anthropic Messages API support (supports_messages) — carina-60b, glm-5-2, glm-5-3, quasar-1-2-358b, and qwen-3-8-27b can be called via POST /v1/messages. See Anthropic Messages API.
  • quasar-438b – Removed from the model catalog. Migrate to quasar-1-2-358b.
  • nemotron-3-nano-omni and mistral-small-3-1 – Deprecated models removed from the catalog.
  • minimal reasoning effort – reasoning.effort and reasoning_effort now accept "minimal" alongside the existing levels. Each model still decides which levels it implements.
  • Prompt caching for glm-5-3 – Prompt caching is now enabled for GLM 5.3. Input tokens served from the cache are billed at 20% of the input price ($0.22 per 1M tokens, versus $1.10 per 1M for non-cached input). See Pricing.
  • No more giving up on a model too early – When a model isn’t immediately available, the API now tries an alternative before returning an error. Previously, if availability hadn’t been checked recently, it could skip that alternative and fail the request.
  • Stalled requests now report 503 instead of 500 – A request where the model never starts responding now returns the retryable 503 first_token_timeout instead of a generic 500 internal_error. Retry on first_token_timeout.
  • No more 404 Model not found for a model that exists – A request for an existing model no longer returns 404 Model not found. When the model cannot start responding, the client now receives a retryable 503 with error.code first_token_timeout or error.code request_deadline_exceeded instead of 500 or 404. Retry on either code.
  • chat_template_kwargs – POST /v1/chat/completions now accepts a top-level chat_template_kwargs object that is forwarded as-is to the model’s chat template renderer (vLLM/SGLang convention). The accepted keys depend on the model’s chat template — for example, Qwen models use {"enable_thinking": true} to toggle reasoning. This is model-agnostic: no catalog entry is required per model.
  • cai-mistral-small-3-1-slim – Removed from the model catalog.
  • gpt-oss-120b – Removed from the model catalog.
  • reasoning_enabled – The top-level reasoning_enabled request field is deprecated in favor of chat_template_kwargs. When both are sent, the user-supplied chat_template_kwargs takes precedence. The reasoning_enabled field will be removed in a future release.

This release contains breaking changes to the error format. If your integration inspects error response bodies, read the migration notes below before upgrading.

  • Standardised error format (breaking) – All errors from POST /v1/chat/completions, POST /v1/responses, POST /v1/audio/transcriptions, GET /v1/models and GET /v1/models/{model} now return a single top-level error object. Responses that previously used {"detail": ...} — including authentication failures, rate limits, request-validation errors and server errors — now use {"error": {"message", "type", "param", "code"}}. See Error Handling.
  • Request-validation errors (breaking) – Validation failures no longer return an array of field errors. They now return one error object whose param identifies the first offending field in dotted/indexed notation (for example messages[0].role), with a count of any remaining errors appended to message.
  • Streaming errors (breaking) – The data: line of an event: error SSE event is now a JSON error envelope rather than a plain-text message, so it can be parsed with the same JSON decoder as every other event. Clients that string-matched the old text will no longer match.
  • Rate limit responses (breaking) – The rate_limit_exceeded identifier has moved from the response body’s detail field to error.code. The Retry-After header is unchanged.
  • param and code are now populated – Both fields were previously always null. param now names the offending request field and code carries a stable machine-readable identifier suitable for programmatic branching. See the error code reference.
  • Removed the router_no_llm_provider error type (breaking) – 503 responses now use the standard server_error type. The router-specific identifier remains available as error.code (no_healthy_llm_provider).
  • Audio transcription status codes (breaking) – A non-boolean stream form value now returns 400 instead of 422. Model provider failures that are not client errors now return 502 instead of forwarding the provider’s own status code.
  • File inputs on /v1/responses – A file resolved from file_id is now attached to the request’s input (as an input_text part for text-decodable files, otherwise an input_file part) instead of being sent as a separate top-level file field. Text-only models can now read uploaded text files. See OpenAI Compatibility.
  • Error messages rewritten – Many messages are clearer, and messages for server-side failures are now deliberately generic so that deployment details are not disclosed. message has always been free-form; do not match on it — branch on error.code or error.type instead.
  1. Read errors from error, not detail. Replace response.json()["detail"] with response.json()["error"]["message"]:

    // Before
    {"detail": "Unexpected server error"}
    // After
    {"error": {"message": "Unexpected server error! Please retry shortly.", "type": "server_error", "param": null, "code": "internal_error"}}
  2. Branch on error.code, not on message text or HTTP status alone. The full list is in the error code reference.

  3. Parse streaming error events as JSON. The data: payload of an event: error is now an object; parse it with the same decoder you use for other events, excluding the data: [DONE] sentinel.

  4. Update rate-limit handling to read error.code == "rate_limit_exceeded" instead of checking detail, and continue honouring the Retry-After header.

  5. If you use POST /v1/completions, no changes are required — it retains the older {"detail": ...} error format. Prefer POST /v1/chat/completions for new integrations.

  • glm-5-2 – GLM 5.2 chat model with tool calling and structured output via POST /v1/chat/completions. Currently available for functional validation only; performance KPIs are not guaranteed during this beta period. See GLM 5.2 and the models catalog.
  • GitHub Copilot – Integration guide for using CompactifAI models directly inside GitHub Copilot via an OpenAI-compatible custom provider. See GitHub Copilot.
  • llama-4-scout, cai-llama-3-3-70b-slim, llama-3-3-70b, llama-3-1-8b, blackstar-10b, gpt-oss-20b, whisper-large-v3, glm-5-2 – These models have been removed. Migrate to an alternative model from the models catalog.
  • nemotron-3-nano-omni – Nemotron 3 Nano Omni multimodal model (text and image inputs via chat completions). See the models catalog and Multi Modality.
  • Cursor and OpenCode – Integration guides for using CompactifAI chat completions, custom model IDs, and the OpenAI-compatible base URL: Cursor and OpenCode.
  • hypernova-60b – Upgraded to version 2605.
  • DeepSeek R1 0528 Slim has been removed.
  • POST /v1/usage/completions (usage statistics API) – This endpoint is deprecated and removed from the API. Usage and billing visibility is now provided through the CompactifAI Dashboard; the route is no longer served.
  • View usage and billing in the CompactifAI Dashboard instead of calling the API. Use the dashboard to monitor consumption, manage API tokens, and review account settings.
  • Whisper Streaming Support – The Whisper audio transcription endpoint now supports streaming, enabling real-time transcription and lower-latency audio processing.
  • GLM-5 (Private Preview) – GLM-5 is now available as a private model in the US region.
  • Agentic Tool Calling Enhancements (Beta) – Improved support for agentic workflows and tool-calling capabilities across the following models:
    • gpt-oss-20b
    • gpt-oss-120b
    • hypernova-60b
  • Deterministic Tool Usage in Agent Loops – When tool_choice=required or tool_choice=<tool_name> is specified, the model consistently returns a tool call, improving reliability in agent-based workflows.
  • Improved Tool Call Reliability – Reduced tool-call hallucinations and improved adherence to defined tool schemas.
  • Improved inference throughput across several models, delivering ~30% higher throughput and better overall serving efficiency for:
    • cai-llama3-1-8b-slim
    • hypernova-60b
    • gpt-oss-120b
    • gpt-oss-20b
    • blackstar-10b
  • Fixed an issue where streaming with tool calling was not supported for some models. The following models now fully support streaming responses with tool calls:
    • gpt-oss-20b
    • gpt-oss-120b
    • hypernova-60b
  • Added tool calling support for hypernova-60b model
  • Fixed a bug where Audio Transcriptions endpoint was not working for all the file mime types specified in our API Reference.
  • Significantly improved the performance of the Audio Transcriptions endpoint using the whisper-large-v3 model, reducing latency and increasing the speed factor from 15x to 100x on a 10 minutes long audio file (The speed of your network connection might affect the speed factor).
  • Fixed a bug where the audio transcription endpoint was not working for audio files with a size greater than 1MB. Now, the endpoint can process audio files up to 25MB in size.
  • Added tool calling support for gpt-oss-20b and gpt-oss-120b
  • Added hypernova-60b
  • Added blackstar-10b model
  • Speech-to-text transcription endpoint /v1/audio/transcriptions with Whisper Large V3 support for multilingual transcription workflows.
  • Feature and API documentation detailing request parameters, Python examples, and guidance for the new speech-to-text capability.
  • Removed the deepseek-r1-0528 model from the API.
  • Multi-modality support for chat completions, enabling image-plus-text inputs across the API.
  • Added mistral-small-3-1 model with full multi-modal understanding and refreshed usage examples.
  • Function tool compatibility has been activated in all models except mistral.
  • Added deepseek-ai/DeepSeek-R1-0528 model, accessible via the deepseek-r1-0528 model ID.
  • Deprecated deepseek-r1.
  • Initial release of the CompactifAI inference API with the following features:
    • Models API endpoint for listing and retrieving available compressed models
    • Chat Completions API endpoint for conversational interactions
    • Completions API endpoint for text generation
  • OpenAI-compatible API design for easy migration and integration
  • Added the following models:
    • cai-llama-3-3-70b-slim
    • cai-mistral-small-3-1-slim
  • HTTPS encryption for all API requests
  • Secure authentication using Bearer token scheme