AI models released in October 2025
32 releases from 19 labs in October 2025, including MiniMax Music 2.0, Kimi Linear and MiniMax Speech 2.6. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 32
- Labs
- 19
- Open weights
- 18
October 2025
32 releasesMiniMax Music 2.0
MiniMax Music 2.0 is a music generation model with more lifelike vocals across singing styles and finer instrumental control, generating structurally complete songs with verses, choruses, and bridges of up to five minutes.
- Modalities
- text → audio
- At launch
- API, Consumer app
- Supports male-female duets and a cappella
- Covers genres including pop, jazz, blues, rock, and folk
Kimi Linear
Kimi Linear is a hybrid linear attention architecture built on Kimi Delta Attention, a refined version of Gated DeltaNet. Moonshot released 48B base and instruct checkpoints trained on 5.7T tokens.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 48B total, 3B active
- License
- MIT
- At launch
- Open weights
- Reduces KV cache needs by up to 75%
- Up to 6 times higher decoding throughput at contexts up to 1M tokens
- KDA kernel open-sourced in the flash-linear-attention library
MiniMax Speech 2.6
MiniMax Speech 2.6 is a text-to-speech model tuned for voice agents, with end-to-end latency under 250 milliseconds, direct reading of URLs, emails, phone numbers, dates, and amounts, and more natural prosody.
- Modalities
- text → audio
- At launch
- API, Consumer app
- Offered as speech-2.6-hd and speech-2.6-turbo
- Introduces Fluent LoRA for more fluent cloned voices
- Voice cloning across the 40+ languages the model supports
gpt-oss-safeguard-120b and gpt-oss-safeguard-20b
gpt-oss-safeguard is a research preview of open-weight safety reasoning models, fine-tuned from gpt-oss, that classify messages, completions, and chats against a developer-provided policy interpreted at inference time.
- Modalities
- text → text
- Parameters
- 120b: 117B total, 5.1B active; 20b: 21B total, 3.6B active
- License
- Apache-2.0
- At launch
- Open weights
- Released in two sizes, 120b and 20b, with weights on Hugging Face
- Same Apache 2.0 license as gpt-oss
- Provides chain-of-thought that developers can review to see how decisions were reached
- Developed with ROOST, which launched a model community alongside the release
Amazon Nova Multimodal Embeddings
Amazon Nova Multimodal Embeddings is a single embedding model that maps text, documents, images, video and audio into one semantic space for cross-modal retrieval, agentic RAG and semantic search.
- Modalities
- text, image, video, audio → embedding
- Context
- 8K context
- At launch
- API
- Supports text, documents, images, video and audio inputs
- Four output dimension sizes trained with Matryoshka Representation Learning
- Up to 8K tokens of context, with text in up to 200 languages
Granite 4.0 Nano
Smallest Granite 4.0 models, released in roughly 350M and 1B sizes as both hybrid H and conventional versions for edge and on-device use.
- Modalities
- text → text
- Parameters
- 350M to 1.6B
- License
- Apache-2.0
- At launch
- Open weights
- 1B models support 128K tokens; 350M models support 32K
- Base and instruct checkpoints released
- Released under Apache 2.0
Kaleido-14B-S2V
Kaleido is Z.ai's open-weights multi-subject reference video generation model, which generates video that keeps the identity of several reference subjects consistent. Z.ai released the Kaleido-14B-S2V checkpoints after open-sourcing the code.
- Modalities
- text, image → video
- Parameters
- 14B
- License
- MIT
- At launch
- Open weights
- Training and inference code released October 22, 2025
- Checkpoints released October 28, 2025
- Weights under the MIT license
LFM2-ColBERT-350M
A multilingual late-interaction retriever built on the LFM2 backbone that encodes queries and documents at the token level, supporting cross-lingual search where documents in one language are retrieved with queries in another.
- Modalities
- text → embedding
- Parameters
- 353M
- License
- LFM Open License v1.0
- At launch
- Open weights
- Document length 512 tokens and query length 32 tokens per model card
- Positioned as a drop-in replacement in existing RAG pipelines
- Available on Hugging Face with an interactive demo
- Succeeded by LFM2.5-ColBERT-350M in 2026
MiniMax Hailuo 2.3
MiniMax Hailuo 2.3 is a video generation model building on Hailuo 02, with improved body movement, stylization, and facial micro-expressions, and better response to motion instructions.
- Modalities
- text, image → video
- At launch
- API, Consumer app
- Launched with Hailuo 2.3 Fast, a faster image-to-video variant
- Broader support for anime, illustration, ink wash, and game CG styles
- Available on the Hailuo AI website and app and the MiniMax API
Nemotron Nano 12B v2 VL
Multimodal document intelligence model that reads multiple document images such as invoices, receipts and manuals for question answering and extraction.
- Modalities
- text, image → text
- Context
- 125K context
- Parameters
- 12.6B
- License
- NVIDIA Open Model License
- At launch
- Open weights
- 128K tokens of combined input and output
- Handles multiple document images per request
- Released under the NVIDIA Open Model License
MiniMax M2
MiniMax M2 is an open-weight mixture-of-experts model built for coding and agentic workflows, designed for end-to-end development in tools such as Claude Code and Cursor and for long chains of tool calls.
- Modalities
- text → text
- Parameters
- MoE: 230B total, 10B active
- License
- Modified-MIT
- At launch
- API, Open weights
- 230B total parameters with 10B active
- Weights on Hugging Face under a modified MIT license
- Uses interleaved thinking during tool use
- Open-sourced and launched on the MiniMax API the same day
LongCat-Video
LongCat-Video is Meituan's 13.6B-parameter foundational video generation model, covering text-to-video, image-to-video, and video continuation in a single model, released with a technical report.
- Modalities
- text, image, video → video
- Parameters
- 13.6B
- License
- MIT
- At launch
- Open weights
- 13.6B parameters
- Single model for text-to-video, image-to-video, and video continuation
- Weights released under MIT
LTX-2
Unified audio-video foundation model that generates synchronized video and sound in one process, with native 4K at up to 50 fps and clips up to 10 seconds at launch.
- Modalities
- text, image, video, audio → video, audio
- Parameters
- 19B
- License
- LTX-2 Community License
- At launch
- Open weights, API, Consumer app
- Full open weights and code released January 6, 2026
- 19B-parameter checkpoints with FP8 and FP4 variants
- Multi-keyframe conditioning and LoRA fine-tuning
Seed3D 1.0
Seed3D 1.0 is a DiT-based 3D generation model that creates high-fidelity, simulation-ready 3D assets with detailed textures and materials from a single image, aimed at world simulators and embodied AI.
- Modalities
- image → 3d
- At launch
- API
- Geometry generation model has 1.5B parameters
- Volcano Engine describes outputs as meshes with PBR materials
- API released alongside the technical report
- Extensible to scene generation
HunyuanWorld-Mirror (HunyuanWorld 1.1)
HunyuanWorld-Mirror, released as HunyuanWorld 1.1, is an open feed-forward model for 3D geometric prediction that creates 3D worlds from videos or multi-view images, optionally using camera pose, intrinsics, and depth priors.
- Modalities
- image, video → 3d
- License
- Tencent HunyuanWorld-Mirror Community License Agreement
- At launch
- Open weights
- Outputs point clouds, multi-view depths, camera parameters, surface normals, and 3D Gaussians
- Training and evaluation code followed on November 7, 2025
LFM2-VL-3B
A 3B-parameter vision-language model built on the LFM2-2.6B backbone with a SigLIP2 400M NaFlex encoder, adding more capacity to the LFM2-VL family and extending visual understanding to ten languages.
- Modalities
- text, image → text
- Context
- 32K context
- Parameters
- 3B
- License
- LFM Open License v1.0
- At launch
- Open weights
- Processes images at native resolution with variable aspect ratios
- Adjustable vision tokens per image to balance speed and quality
- Available on Hugging Face under the LFM Open License and on LEAP
olmOCR 2
Updated OCR model for English digitized print documents, built on Qwen2.5-VL-7B and trained with unit test rewards for document correctness.
- Modalities
- text, image → text
- Parameters
- 7B
- License
- Apache-2.0
- At launch
- Open weights
- Released as olmOCR-2-7B-1025
- Fine-tuned on olmOCR-mix-1025 with 270,000 PDF pages
- Released under Apache 2.0
DeepSeek-OCR
DeepSeek-OCR is an open-weights vision-language model that explores contexts optical compression, studying the role of vision encoders from an LLM-centric viewpoint. It converts document images to text and markdown.
- Modalities
- text, image → text
- License
- MIT
- At launch
- Open weights
- Released on GitHub and Hugging Face on October 20, 2025
- Supported in upstream vLLM from October 23, 2025
- Weights under the MIT License
PaddleOCR-VL
PaddleOCR-VL is Baidu's open document-parsing system built around PaddleOCR-VL-0.9B, a compact vision-language model that pairs a NaViT-style dynamic-resolution visual encoder with the ERNIE-4.5-0.3B language model.
- Modalities
- image, text → text
- Parameters
- 0.9B
- License
- Apache-2.0
- At launch
- Open weights
- Supports 109 languages
- Recognizes text, tables, formulas, and charts, including handwriting and historical documents
- Uses PP-DocLayoutV2 for layout analysis and reading order, with Markdown and JSON output
Claude Haiku 4.5
Claude Haiku 4.5 is Anthropic's small model, which Anthropic says matches Claude Sonnet 4 on coding at lower cost and higher speed, aimed at real-time, low-latency work and multi-agent setups that run subtasks in parallel.
- Modalities
- text, image → text
- Context
- 200K context
- At launch
- API, Consumer app, Cloud platforms
- API model ID claude-haiku-4-5 (snapshot claude-haiku-4-5-20251001)
- 200K-token context window and 64K max output tokens
- Available in the Claude apps, Claude Code, the Claude API, Amazon Bedrock, and Google Cloud Vertex AI
Doubao-Seed-1.6-lite
Volcano Engine upgraded Doubao-Seed-1.6 with four selectable thinking lengths (Minimal, Low, Medium, High) and released Doubao-Seed-1.6-lite, a smaller and faster deep-thinking model for enterprise workloads.
- Modalities
- text → text
- At launch
- API
- Adds Minimal, Low, Medium, and High thinking-length settings
- Volcano Engine calls it the first model in China with native graded thinking-length control
- Doubao-Seed-1.6-lite released the same day
Doubao-Seed-1.6-Rerank
Doubao-Seed-1.6-Rerank is an all-modality reranking model that reorders retrieved text, image, and video results, integrated into Volcano Engine's VikingDB vector database alongside Doubao-Seed-1.6-Embedding for end-to-end retrieval.
- Modalities
- text, image, video → text
- At launch
- API
- Reranks mixed text, image, and video candidates
- Understands negation and comparison logic in queries
- Produces stable relevance scores suited to threshold filtering
Seed-TTS 2.0 and Seed-ICL 2.0
Seed-TTS 2.0 and Seed-ICL 2.0 are Doubao's speech synthesis and voice cloning models, introducing a query-response conversational synthesis approach that uses dialogue context and natural-language instructions to control tone, emotion, and speaking rate.
- Modalities
- text → audio
- At launch
- API
- Takes the user's query and background information as context for delivery
- Speaking rate, emotion, and style can be steered with text instructions
- Available in the Doubao Speech console on Volcano Engine
Veo 3.1
Veo 3.1 is Google DeepMind's upgrade to Veo 3, bringing richer generated audio, stronger prompt adherence for image-to-video and more narrative control. It launched with a Veo 3.1 Fast variant across Google's apps and APIs.
- Modalities
- text, image → video, audio
- At launch
- Consumer app, API, Cloud platforms
- Flow features Ingredients to Video, Frames to Video and Extend now include audio
- Veo 3.1 and Veo 3.1 Fast entered public preview in the Gemini API on launch day
- Generally available on Vertex AI from November 17, 2025
- Also available in the Gemini app
MAI-Image-1
MAI-Image-1 is Microsoft AI's first image generation model developed entirely in-house. It later became available in Bing Image Creator and in Copilot Audio Expressions Story Mode.
- Modalities
- text → image
- At launch
- Consumer app
- Microsoft's first in-house text-to-image model
- Added to Bing Image Creator and Copilot on November 4, 2025
- Available in countries that can access Bing Image Creator and Copilot Labs
Jamba Reasoning 3B
Compact open-source reasoning model built on AI21's hybrid SSM-Transformer architecture, designed to run on phones and computers for long-context reasoning.
- Modalities
- text → text
- Context
- 250K context
- Parameters
- 3B
- License
- Apache-2.0
- At launch
- Open weights
- Handles 256K tokens and can stretch to 1M
- Runs locally via llama.cpp or LM Studio
- Available on Hugging Face and Kaggle
Gemini 2.5 Computer Use
The Gemini 2.5 Computer Use model is a specialized model built on Gemini 2.5 Pro's visual understanding and reasoning that powers agents able to interact with user interfaces in browsers and on mobile.
- Modalities
- image, text → text
- Context
- 125K context
- At launch
- API, Cloud platforms
- Public preview in the Gemini API on Google AI Studio and Vertex AI
- Model page lists a 128,000 input token limit and 64,000 output tokens
- Computer Use later became a tool in Gemini 3 and 3.5 models, reaching general availability on Oct 30, 2026
LFM2-8B-A1B
Liquid AI's first on-device mixture-of-experts model, an early checkpoint that adds sparse MoE feed-forward layers to the LFM2 backbone, with 8.3B total and 1.5B active parameters per token.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- MoE: 8.3B total, 1.5B active
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- 32 experts per MoE block with top-4 routing; 18 convolution and 6 GQA blocks
- Pre-trained on about 12T tokens with more code than LFM2-2.6B
- Weights and GGUF quants on Hugging Face; testable on Liquid Playground
- Later replaced by LFM2.5-8B-A1B (May 2026)
GPT Image 1 Mini
GPT Image 1 Mini is a cost-efficient version of GPT Image 1 for image generation and editing, released in the API at OpenAI DevDay 2025 alongside new audio and video models.
- Modalities
- text, image → image
- At launch
- API
- API model ID gpt-image-1-mini on the image generation and edit endpoints
- Released at OpenAI DevDay on October 6, 2025
- Accepts text and image input and outputs images, per the API model page
gpt-realtime-mini and gpt-audio-mini
gpt-realtime-mini and gpt-audio-mini are cost-efficient versions of gpt-realtime and gpt-audio for speech-to-speech applications, released at OpenAI DevDay 2025 for the Realtime API and Chat Completions API respectively.
- Modalities
- text, audio, image → text, audio
- At launch
- API
- gpt-realtime-mini runs in the Realtime API; gpt-audio-mini in Chat Completions
- Context: 32,000 tokens for gpt-realtime-mini and 128,000 for gpt-audio-mini
- Image input is supported by gpt-realtime-mini only, per the API model pages
- Updated snapshots dated 2025-12-15 followed in December 2025
Granite 4.0
Fourth-generation Granite language models that add hybrid H variants, including mixture-of-experts models, alongside a conventional dense Micro model, all with 128K-token context.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- H-Small 32B (9B active), H-Tiny 7B (1B active), H-Micro 3B, Micro 3B
- License
- Apache-2.0
- At launch
- Open weights
- H-Small activates 9B of 32B parameters
- Base and instruct checkpoints released
- Released under Apache 2.0
LFM2-Audio-1.5B
An end-to-end audio and text foundation model that extends LFM2-1.2B to accept and generate both speech and text, covering conversational chat, speech recognition, text-to-speech, and audio classification in one 1.5B-parameter model.
- Modalities
- text, audio → text, audio
- Context
- 32K context
- Parameters
- 1.5B
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- Continuous embeddings for audio input; discrete audio tokens for output
- Tokenizer-free input: raw waveform chunked into segments of about 80 ms
- Supported language per model card: English
- Available on Hugging Face and Liquid Playground
Cite and reuse this data
These release records are free to reuse under CC BY 4.0, including commercially. Credit The AI Engineering and link to this page's URL.
Dates are the labs' official announcement dates, and every record links to its source. List fields in the CSV are separated by " | ".
Citation
The AI Engineering Editors. "AI models released in October 2025." The AI Engineering, updated Oct 31, 2025. https://theaiengineering.com/releases/2025/10
HTML link
<a href="https://theaiengineering.com/releases/2025/10">AI models released in October 2025</a> (The AI Engineering)- Full dataset
- 1,167 releases from 55 labs
- Coverage
- to
- Updated
- License
- CC BY 4.0
- Publisher
- The AI Engineering