Infercom Release Notes and Product Updates

October 1, 2026

Deprecation: MiniMax-M2.7 retired

• MiniMax-M2.7 was retired on September 30, 2026 at 6:00 PM CEST. Requests to MiniMax-M2.7 now return HTTP 410
• Migrate to gpt-oss-120b or gemma-4-31B-it, both EU-hosted
• Model overview, rate limits, reasoning, function calling, OpenAI compatibility and deprecations updated

Global Model Catalog: MiniMax-M3 added

• MiniMax-M3 is available through the Global Model Catalog with a 1M token context window. It is processed outside the EU, in a region we do not control
• Our provider lists it as a preview model with limited capacity. It accepts text only
• logit_bias and any repetition_penalty other than 1.0 return an error on this model. See OpenAI compatibility

September 30, 2026

Upcoming Deprecation: MiniMax-M2.7

• MiniMax-M2.7 will be deprecated on September 30, 2026 at 6:00 PM CEST. After the cutover, requests to MiniMax-M2.7 return an error
• This notice is shorter than our standard two to three weeks. We apologise for the short notice
• Recommended replacements: gpt-oss-120b (reasoning, function calling, 128k context) or gemma-4-31B-it (vision-capable, 128k context). Both are EU-hosted and live now
• See deprecations for migration guidance, or contact support@infercom.ai

Documentation: agentic coding guides withdrawn

• The agentic coding guides are withdrawn for now. Their URLs redirect to OpenAI compatibility, Responses API or Anthropic compatibility
• Code examples across the docs now use gpt-oss-120b

September 18, 2026

Correction: logit_bias works on gpt-oss-120b only

• On September 11 we listed logit_bias as working on DeepSeek-V3.2 and Meta-Llama-3.3-70B-Instruct. That was wrong. Our test counted failed requests as changed output
• logit_bias returns HTTP 500 on DeepSeek-V3.1, DeepSeek-V3.2 and Meta-Llama-3.3-70B-Instruct, and on some MiniMax-M2.7 requests. It is ignored on gemma-4-31B-it and on the other MiniMax-M2.7 requests
• Action: send logit_bias only to gpt-oss-120b. The OpenAI SDKs retry a 500 twice

Clarification: how repetition_penalty and seed behave

• repetition_penalty applies on /v1/chat/completions only. /v1/responses and /v1/messages accept it and ignore it
• It also penalises tokens in your prompt. A long system prompt, such as a voice agent’s, can lose words it relies on. It is not a drop-in replacement for frequency_penalty
• Higher values lengthen reasoning rather than always truncating. On gpt-oss-120b a one-sentence answer used 104 completion tokens at 1.0 and 244 at 1.3. With a tight max_tokens that can truncate the answer. 1.8 still returns no content. Stay at or below 1.2
• seed pins MiniMax-M2.7 on each of its two deployments, so the same seed returns one of two texts
• DeepSeek-V3.1 added to the parameter table: n, logprobs and seed work; repetition_penalty has no effect
• /v1/responses echoes frequency_penalty and presence_penalty in the response. They are still not applied

Advance notice: stricter validation for penalty parameters

• An upcoming platform release will return an error for frequency_penalty or presence_penalty set to anything other than 0. Today these values are accepted and have no effect. We will add the date here once it is set
• The same release will return an error for repetition_penalty on /v1/responses and /v1/messages, where it is ignored today
• Action now: remove frequency_penalty and presence_penalty from your requests, or set them to 0, and send repetition_penalty to /v1/chat/completions only. Nothing changes for requests that already do this

September 11, 2026

Correction: five features we listed as unsupported actually work

• n, logprobs and top_logprobs work on every model. n: 2 returns two choices. logprobs: true with top_logprobs: 3 returns populated token-level logprobs with alternatives
• seed and logit_bias work on gpt-oss-120b, DeepSeek-V3.2 and Meta-Llama-3.3-70B-Instruct. On those three a seed pins the response at temperature 1.0: the same seed returns the same text every time, and a different seed returns different text. Without a seed the same prompt varies on nearly every request
• seed on MiniMax-M2.7 narrows the output without pinning it. A seed leaves two variants rather than one, where the same prompt without a seed varies on every request. It has no effect at all on gemma-4-31B-it
• These arrived with a platform release in May 2026 and our documentation was not updated. The unsupported list now contains only the two parameters that really are unsupported
• A parameter that has no effect on a model is still accepted. Nothing in the response says it was dropped. See Supported, but not on every model

Documentation correction: frequency_penalty does nothing, repetition_penalty is the one that works

• frequency_penalty and presence_penalty are accepted and ignored on every model. On all five models at temperature 0, both return output byte-identical to sending nothing. frequency_penalty: 99 is outside the documented range, is not rejected, and is also ignored. presence_penalty: 99 does return a 400
• The API reference described both with full OpenAI semantics and no disclaimer. It now states that they are accepted for compatibility and ignored
• repetition_penalty is the parameter that applies. It was absent from the API reference and is now documented there and on the OpenAI compatibility page, next to top_k. Accepted range is 1 to 2; values outside it return a 400
• It is honoured on gpt-oss-120b, MiniMax-M2.7 and gemma-4-31B-it, and ignored on DeepSeek-V3.2 and Meta-Llama-3.3-70B-Instruct
• High values cut the answer short. On gpt-oss-120b at temperature 0, 1.3 truncates the response with finish_reason: "length", and 1.8 returns no content at all, the whole budget having gone to reasoning. Start at 1.05 and stay at or below 1.2
• Action: if you set frequency_penalty to stop a repetition loop, it is not doing anything. Fix temperature first, then use repetition_penalty

September 10, 2026

Documentation correction: a retired model returns a plain-text 410

• Requesting a deprecated model on /v1/chat/completions returns HTTP 410 with content-type: text/plain; charset=utf-8. The body is not JSON. There is no error object and no request_id, so parsing it raises a decode error
• Official SDKs survive it. The OpenAI Python SDK raises openai.APIStatusError with status_code 410 and the sentence in .body. Clients that call .json() on the raw response themselves do not
• The same condition returns JSON on /v1/responses (envelope, type and code both internal_error) and /v1/messages (Anthropic shape, not_found_error)
• The error codes page listed two responses that do not use the envelope. There are three. The page, the defensive parsing helper, and the OpenAPI reference now cover all three
• model_deprecated is documented as the 410 code, but no endpoint returns it today. The row now says so. A fix is in progress with our platform provider
• Do not retry a 410. The condition is permanent. Change the model ID. See API error codes

September 9, 2026

Documentation correction: top_k is silently disabled at low temperature

• top_k has no effect unless temperature is 0.001 or above. Below that threshold, including when you omit temperature entirely, top_k is forced to 1 and decoding is greedy. A request carrying top_k: 40 and no temperature returns argmax, and nothing in the response says the value was ignored
• On gemma-4-31B-it, top_k: 40 without temperature returns the same text on every request. The same request with temperature: 1.0 varies on every request
• Documented the defaults applied when you omit a sampling parameter: temperature 0.0, top_p 1.0, top_k 1048576, repetition_penalty 1.0
• top_k must be 1 or greater. -1 and 0 are rejected with a 400 on /v1/chat/completions and /v1/messages. They are accepted on /v1/responses, which does not validate the lower bound, but they do not mean “unrestricted” anywhere. The API reference stated a maximum of 100 for top_k; no upper bound is enforced
• The byte-identical output you get when you omit temperature is not a pinned seed. It is the forced top_k of 1. Do not rely on it across releases
• Documented why MiniMax-M2.7 samples while the other models decode greedily: on the MiniMax bundles only, unset sampling parameters are backfilled from the checkpoint’s own generation_config.json
• Action: set temperature explicitly before you tune top_k. See How top_k interacts with temperature

Platform maintenance release: vision tool calling, reasoning token counts, error envelope

• Vision requests now support tools and structured output. gemma-4-31B-it honours tools, tool_choice and response_format on requests that carry an image. The constraint is applied during generation. See Vision
• Reasoning token counts are reported on every EU model. Read usage.completion_tokens_details.reasoning_tokens to budget and bill reasoning separately from the visible answer. It was previously null on gemma-4-31B-it. See Counting reasoning tokens
• Malformed image data returns HTTP 400, not 500. This is a permanent error. Do not retry it
• Errors use a consistent JSON envelope, with two exceptions. The response carries a request_id you should quote in every support report. A malformed image and an unhonoured forced tool_choice still return a flat shape where error is a string. See API error codes
• Streaming tool calls on gpt-oss-120b are clean. No reasoning scaffolding leaks into delta.content. Note that the same model adds a channel field to the delta on streams with no tools - ignore delta fields you do not recognise. See Streaming tool calls
• Documentation corrections: the audio page listed stream_options as an accepted parameter. The endpoint rejects it with HTTP 400. response_format honours only json and text; verbose_json, srt and vtt are accepted but return plain text. See Audio

August 18, 2026

Documentation correction: temperature has no default

• The API reference stated a default temperature of 0.7 on /v1/chat/completions and /v1/responses. That default does not exist. The claim is removed from the OpenAPI spec and from every page that repeated it
• Omit temperature and gemma-4-31B-it, gpt-oss-120b, DeepSeek-V3.2, and Meta-Llama-3.3-70B-Instruct decode greedily. MiniMax-M2.7 samples. The same applies on /v1/chat/completions, /v1/responses, and /v1/messages
• Greedy decoding is the setting most likely to send a reasoning model into a non-terminating repetition loop, so requests that set no sampling parameters were exposed to the failure the docs warned about
• /v1/messages differs from the Anthropic API, which documents a temperature default of 1.0. Code ported from the Anthropic SDK switches to greedy decoding unless you set the parameter
• Action: set temperature explicitly on every request. See If you omit temperature

July 16, 2026

Deprecation: MiniMax-M2.5 retired

• MiniMax-M2.5 is deprecated effective July 16, 2026, 9:00 PM CEST. After the cutover, requests to MiniMax-M2.5 return an error
• Migrate to MiniMax-M2.7 (192k context) or gemma-4-31B-it (cost-efficient, vision-capable)
• Model overview, rate limits, and deprecations updated

July 3, 2026

Upcoming Deprecation: MiniMax-M2.5

• MiniMax-M2.5 will be deprecated on July 16, 2026 at 9:00 PM CEST
• Recommended replacements: MiniMax-M2.7 (complex reasoning, 192k context) or gemma-4-31B-it (cost-efficient, vision-capable)
• Both replacement models are already live - migrate before July 16th to avoid service disruption
• See deprecations for migration guidance

June 20, 2026

Model Update: Gemma 4 replaces Gemma 3

• gemma-3-12b-it has been removed and replaced by gemma-4-31B-it in a single switchover. Requests to gemma-3-12b-it are no longer served - update your integrations to use the gemma-4-31B-it model ID
• Gemma 4 is EU-hosted on Infercom’s sovereign infrastructure with full data sovereignty, offering a 128k context window and improved performance while maintaining vision support
• Vision guide updated with gemma-4-31B-it examples
• Rate limits and deprecations updated

June 5, 2026

New Models: Whisper-Large-v3 and E5-Mistral-7B-Instruct

• Whisper-Large-v3 is now available for audio transcription and translation - OpenAI’s state-of-the-art ASR model, EU-hosted with full data sovereignty
• E5-Mistral-7B-Instruct is now available for generating embeddings - high-quality vector representations for RAG, search, and classification workflows
• Both models run on EU sovereign infrastructure with no data leaving EU jurisdiction
• Rate limits updated for both models
• Model overview updated with new embedding and audio model sections

May 24, 2026

Anthropic SDK Compatibility

• New /v1/messages endpoint for Anthropic SDK compatibility
• Use the Anthropic Python SDK with Infercom models - just change the base URL
• Supports streaming, system prompts, multi-turn conversations, and tool use
• Documentation

May 22, 2026

New Model: MiniMax-M2.7

• MiniMax-M2.7 is now available as an EU-hosted model with 192K context window
• Recommended for agentic coding workflows - replaces M2.5 as the default recommendation
• MiniMax-M2.5 (160K context) remains available but will be deprecated in a future release
• Rate limits updated for MiniMax-M2.7

May 15, 2026

Responses API

• New POST /v1/responses endpoint for agentic workflows - supports function tools, streaming, and reasoning effort control
• Compatible with OpenAI Responses API standard
• Supported models: MiniMax-M2.7, MiniMax-M2.5, gpt-oss-120b
• Documentation

Model Update

• gemma-3-12b-it moved to EU sovereign infrastructure - first vision-capable model with full EU data sovereignty
• Vision guide

New Integration

• Codex CLI now supported via Responses API
• Updated Cline and OpenCode guides with Plan/Execute configuration pattern

April 24, 2026

Model Catalog Updates

• DeepSeek-V3.1 moved to Global Model Catalog: Now served via proxy (US region) instead of EU-hosted infrastructure
• gemma-3-12b-it added: Google’s Gemma 3 12B model now available via Global Model Catalog (Japan). This is the first vision-capable model on Infercom — see Vision guide
• DeepSeek-V3.2 context length corrected: Updated from 8k to 32k tokens
• EU-hosted models: MiniMax-M2.5 and gpt-oss-120b remain as the sovereign EU options

Documentation Updates

• Updated Vision guide with gemma-3-12b-it examples and sovereignty notice
• Added Capabilities column to model overview tables

April 6, 2026

New Documentation: Agentic Coding

• Added Agentic Coding section with guides for 9 AI coding tools
• Terminal tools: Aider, OpenCode, Goose, Warp
• IDE tools: Cursor, Cline, Continue, Windsurf
• Advanced: Claude Code via ccproxy
• Tool comparison guide - Recommendations for choosing the right tool

April 3, 2026

Global Model Catalog Streamlined

• Reduced Global Model Catalog to 2 models: Meta-Llama-3.3-70B-Instruct (128k) and DeepSeek-V3.2 (8k)
• 8 models deprecated: Qwen3-32B, Qwen3-235B, DeepSeek-R1-0528, DeepSeek-R1-Distill-Llama-70B, DeepSeek-V3-0324, DeepSeek-V3.1-Terminus, Llama-4-Maverick-17B-128E-Instruct, Meta-Llama-3.1-8B-Instruct
• See deprecations for migration guidance
• Fixed MiniMax-M2.5 context length to 160k (was incorrectly listed as 164k)

March 12, 2026

Model Update: MiniMax M2.5 replaces Meta-Llama-3.3-70B-Instruct

• MiniMax M2.5 is now available as an EU-hosted model on Infercom’s sovereign infrastructure in Germany
• Meta-Llama-3.3-70B-Instruct has been deprecated and removed from the platform. See deprecations for migration guidance
• Rate limits updated for MiniMax M2.5

February 27, 2026

Global Model Catalog Launch

• 9 new models available through the Global Model Catalog, including DeepSeek-R1-0528, Qwen3-235B, Llama-4-Maverick, and more
• EU-hosted and globally-routed models are now clearly separated in the model overview
• Rate limits updated for all 12 models across Free and Developer tiers
• API sn_metadata now returns is_external and region fields for all models

February 14, 2026

New Documentation

• Added Performance & Latency guide covering connection pooling, response performance metadata, streaming optimization, and best practices for minimizing latency

February 5, 2026

Documentation for Global Model Catalog

• Added documentation for the Global Model Catalog, covering EU-hosted and globally-routed models
• Documented how to identify model regions via the API (sn_metadata.region) and the Playground
• Updated API reference with sn_metadata schema and ?verbose=true query parameter
• Reviewed and qualified sovereignty claims across documentation to reflect Global Model Catalog

January 23, 2026

Model Update

• Updated DeepSeek model from DeepSeek-V3-0324-cb to DeepSeek-V3.1 with expanded 128k context window
• The previous model has been deprecated and is listed on the deprecations page
• Updated rate limits for all models

January 13, 2026

New

• Hugging Face Spaces Chat Demo - Interactive chat with model selection and real-time performance metrics

Improvements

• Documentation cleanup and link fixes

December 30, 2025

Updated the model catalog to reflect currently available models on Infercom Inference Service. The model list and rate limits documentation have been updated accordingly.

November 17, 2025

We are pleased to announce the launch of Infercom’s documentation portal. This comprehensive documentation has been created based on SambaNova’s documentation (dated October 7, 2025) and adapted for Infercom’s EU sovereign AI inference platform.

Key features

• Complete API reference documentation with OpenAI-compatible endpoints.
• Developer guides for integrating with Infercom’s inference platform.
• Model catalog and configuration documentation for all supported models.
• Platform architecture and deployment guides.
• Usage examples in Python and TypeScript, powered by OpenAI-compatible SDKs.

All documentation will be continuously updated to reflect Infercom-specific features, European data sovereignty capabilities, and platform enhancements as they become available.