AI models released in January 2026
35 releases from 20 labs in January 2026, including Kling 3.0, Genie 3 (Project Genie) and PaddleOCR-VL-1.5. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 35
- Labs
- 20
- Open weights
- 25
January 2026
35 releasesKling 3.0
Kling 3.0 is Kuaishou's model series comprising Video 3.0, Video 3.0 Omni, Image 3.0, and Image 3.0 Omni, unifying generation and editing with native audio, longer clips, and stronger character consistency.
- Modalities
- text, image, video, audio → video, image, audio
- At launch
- Consumer app
- Video generation up to 15 seconds with native multilingual audio
- Video 3.0 Omni adds reference-video based character and voice consistency
- Image 3.0 supports direct 2K and 4K output
- Kuaishou's press release announcing the series followed on February 5, 2026
Genie 3 (Project Genie)
Genie 3 is Google DeepMind's world model that generates explorable, interactive environments in real time. First shown as a limited research preview, it became usable through the Project Genie prototype for Google AI Ultra subscribers.
- Modalities
- text, image → video
- At launch
- Consumer app
- Announced August 5, 2025 as a limited research preview
- Project Genie opened to U.S. Google AI Ultra subscribers aged 18+ on January 29, 2026
- Generations are limited to 60 seconds in Project Genie
- Generates worlds navigable in real time at 720p and 24 frames per second
PaddleOCR-VL-1.5
PaddleOCR-VL-1.5 updates Baidu's 0.9B document-parsing vision-language model for in-the-wild robustness, handling skewed, warped, and photographed pages, and adds text spotting and seal recognition.
- Modalities
- image, text → text
- Parameters
- 0.9B
- License
- Apache-2.0
- At launch
- Open weights
- Supports irregular-shaped (polygon) localization for distorted documents
- Adds automatic cross-page table merging and cross-page heading recognition
- Language coverage adds Tibetan and Bengali
Qwen3-ASR
Qwen3-ASR is an open-weights speech recognition family of Qwen3-ASR-1.7B and Qwen3-ASR-0.6B, released with the non-autoregressive Qwen3-ForcedAligner-0.6B. The models build on the audio understanding of Qwen3-Omni.
- Modalities
- audio → text
- Parameters
- 0.6B, 1.7B
- License
- Apache-2.0
- At launch
- Open weights
- Language identification and recognition for 52 languages and accents
- Qwen3-ForcedAligner-0.6B performs speech forced alignment
- Weights under the Apache 2.0 license
Grok Imagine Video (Imagine API)
The Grok Imagine API brought xAI's video-audio generation and video editing models to developers, supporting image-to-video, text-to-video, and edits such as restyling scenes, adding or removing objects, and controlling motion.
- Modalities
- text, image, video → video, audio
- At launch
- API
- API model ID: grok-imagine-video (text, image, or video input per docs)
- Generates video with native audio
- Docs release notes list a revamped Image Generation API the same day
- Offered on partner platforms alongside the xAI API
DeepSeek-OCR 2
DeepSeek-OCR 2 is DeepSeek's second document OCR vision-language model, introducing a visual encoding approach the lab calls Visual Causal Flow. It follows DeepSeek-OCR and is released with vLLM and Transformers inference code.
- Modalities
- text, image → text
- License
- Apache-2.0
- At launch
- Open weights
- Subtitle: Visual Causal Flow, aimed at more human-like visual encoding
- Weights on Hugging Face under the Apache 2.0 license
- Announced in the DeepSeek-OCR repository on January 27, 2026
Kimi K2.5
Kimi K2.5 builds on Kimi K2 with continued pre-training on about 15T mixed visual and text tokens, making it a native multimodal model. It introduced a self-directed agent swarm that can coordinate up to 100 sub-agents.
- Modalities
- text, image, video → text
- Context
- 250K context
- Parameters
- MoE: 1T total, 32B active
- License
- Modified MIT License
- At launch
- API, Consumer app, Open weights
- Agent swarm runs parallel workflows across up to 1,500 tool calls
- Kimi.com and the Kimi App offer Instant, Thinking, Agent, and Agent Swarm (beta) modes
- Weights and code under the Modified MIT License
MiniMax M2-her
MiniMax-M2-her is a role-play language model for multi-turn character dialogue, built from three years of role-play optimization in MiniMax's Talkie and Xingye apps and offered through the API as a chat model.
- Modalities
- text → text
- At launch
- API, Consumer app
- Designed to keep characters consistent with established world settings and lore
- Uses narration to bridge time or scene changes instead of forcing impossible dialogue
- Listed in the MiniMax API docs as the M2-her chat model
SERA
First release in Ai2's Open Coding Agents family, coding models built on Qwen 3 with a training method for adapting agents to any repository.
- Modalities
- text → text
- Parameters
- 8B to 32B
- License
- Apache-2.0
- At launch
- Open weights
- Trained up to 32K context length
- SERA-14B added on February 3, 2026
- Released under Apache 2.0
HunyuanImage 3.0-Instruct
HunyuanImage 3.0-Instruct adds instruction reasoning to HunyuanImage 3.0 for prompt enhancement, along with image-to-image generation for editing and multi-image fusion. A distilled checkpoint for faster sampling was released alongside it.
- Modalities
- text, image → image
- License
- Tencent Hunyuan Community License Agreement
- At launch
- Open weights
- HunyuanImage-3.0-Instruct-Distil recommends 8-step sampling
- Weights on Hugging Face
Qwen3-Max-Thinking
Qwen3-Max-Thinking is Qwen's flagship reasoning model, built by scaling model parameters and reinforcement learning compute on Qwen3-Max. It adds adaptive tool use for on-demand retrieval and code interpreter calls.
- Modalities
- text → text
- At launch
- API, Consumer app
- Available in Qwen Chat and through the Alibaba Cloud API
- Uses test-time scaling techniques to improve reasoning
- Improves factual knowledge, instruction following, and agent capabilities
Ray3.14
Ray3 update adding native 1080p generation across core Dream Machine workflows, four times faster generation at 720p than Ray3, better prompt adherence and more consistent Modify Video output.
- Modalities
- text, image, video → video
- At launch
- Consumer app
- Native 1080p generation
- Modify Video up to 18 seconds
- Character references not supported at launch
MiniMax Speech 2.8
MiniMax Speech 2.8 is a text-to-speech model that adds native sound tags for breaths, laughter, and filler words, along with higher-fidelity voice cloning and cleaner, studio-grade audio.
- Modalities
- text → audio
- At launch
- API, Consumer app
- Offered as speech-2.8-hd and speech-2.8-turbo
- Voice cloning from about 10 seconds of audio
- Improved cross-language rendering starting with the Mandarin-Japanese pair
Qwen3-TTS
Qwen3-TTS is Qwen's open-weights speech generation family supporting voice cloning, voice design, and natural-language voice control. It is built on the Qwen3-TTS-Tokenizer-12Hz multi-codebook speech encoder.
- Modalities
- text, audio → audio
- Parameters
- 0.6B, 1.7B
- License
- Apache-2.0
- At launch
- Open weights
- Released in 1.7B and 0.6B sizes
- Supports 10 major languages plus various dialects
- Weights under the Apache 2.0 license
LFM2.5-1.2B-Thinking
A 1.2B-parameter reasoning model in the LFM2.5 family that generates thinking traces before answering and runs on-device within 900 MB of memory on a phone.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 1.17B
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- Available on Hugging Face, LEAP, and Liquid Playground
- New launch partners: Qualcomm Technologies, Ollama, FastFlowLM, and Cactus Compute
- 32,768-token context window per model card
GLM-4.7-Flash
GLM-4.7-Flash is a lightweight 30B-A3B mixture-of-experts version of GLM-4.7 for efficient local and API deployment. Z.ai optimized it for high-frequency use cases on its API platform.
- Modalities
- text → text
- Parameters
- MoE: 30B total, 3B active
- License
- MIT
- At launch
- API, Open weights
- Lightweight version of GLM-4.7 offered on the BigModel and Z.ai APIs
- Aimed at coding, writing, translation, and long-text tasks
- Weights on Hugging Face under the MIT license
MiniMax Music 2.5
MiniMax Music 2.5 adds paragraph-level control over song structure and higher-fidelity sound, supporting 14 structural section tags such as Intro, Bridge, Interlude, Build-up, and Hook.
- Modalities
- text → audio
- At launch
- API, Consumer app
- API release notes list it on January 16, 2026; the announcement post is dated January 28, 2026
- Supports more than 100 instruments
- Lyrics can carry structural tags, instrument names, and prompts
FLUX.2 [klein] 4B
Compact FLUX.2 model family member that unifies text-to-image, editing and multi-reference generation with sub-second inference, fully open under Apache 2.0 and able to run in about 13GB of VRAM.
- Modalities
- text, image → image
- Parameters
- 4B
- License
- Apache-2.0
- At launch
- Open weights, API
- Distilled and undistilled Base 4B variants
- FP8 and NVFP4 builds developed with NVIDIA
- Also served through the BFL API
FLUX.2 [klein] 9B
Flagship small FLUX.2 model, a 9B flow model with an 8B Qwen3 text embedder step-distilled to four inference steps for sub-second generation, editing and multi-reference work.
- Modalities
- text, image → image
- Parameters
- 9B
- License
- FLUX Non-Commercial License
- At launch
- Open weights, API
- Undistilled Base 9B variant for fine-tuning and research
- Open weights under the FLUX Non-Commercial License
- Available in the BFL API
TranslateGemma
TranslateGemma is a suite of open translation models built on Gemma 3 in 4B, 12B and 27B sizes. It covers 55 languages and keeps Gemma 3's multimodal ability to translate text found inside images.
- Modalities
- text, image → text
- Context
- 2K context
- Parameters
- 4B, 12B, 27B
- License
- Gemma Terms of Use
- At launch
- Open weights, Cloud platforms
- 4B targets mobile and edge, 12B consumer laptops, 27B a single H100 GPU or TPU
- Evaluated on 55 language pairs, with nearly 500 more pairs in training
- Weights on Kaggle and Hugging Face; deployable in Vertex AI
GLM-Image
GLM-Image is Z.ai's image generation model that combines an autoregressive generator with a diffusion decoder. Z.ai says it was fully trained on domestic Chinese chips and improves text rendering in knowledge-intensive images.
- Modalities
- text, image → image
- Parameters
- 9B autoregressive generator + 7B diffusion decoder
- License
- MIT
- At launch
- API, Open weights
- Autoregressive generator initialized from GLM-4-9B-0414
- 7B single-stream DiT decoder with a Glyph Encoder for accurate in-image text
- Weights on Hugging Face under the MIT license
LongCat-Flash-Thinking-2601
LongCat-Flash-Thinking-2601 is an upgraded LongCat reasoning model that strengthens agentic tool use and search through large-scale multi-environment reinforcement learning and adds a deep thinking mode.
- Modalities
- text → text
- Parameters
- MoE: 560B total, about 27B active
- License
- MIT
- At launch
- API, Open weights
- Trained across environments with more than 60 tool types
- Deep thinking mode combines parallel reasoning with recursive summarization
- A LongCat-Flash-Thinking-ZigZag variant with sparse attention was also released
- Weights on Hugging Face and GitHub under MIT
Step-Audio R1.1
Step-Audio R1.1 (Realtime) is an upgrade to Step-Audio-R1 for interactive spoken dialogue, designed to reason while speaking so that real-time voice conversations keep low latency without giving up reasoning ability.
- Modalities
- audio, text → text, audio
- License
- Apache-2.0
- At launch
- Open weights
- Dual-brain design separates high-level reasoning from speech generation
- Grounds reasoning in acoustic representations rather than transcripts alone
- Inference code and weights released on Hugging Face and ModelScope
MedASR
MedASR is an open automatic speech recognition model from Google Research fine-tuned for medical dictation. Released alongside MedGemma 1.5, it converts clinical speech to text and can produce prompts for MedGemma.
- Modalities
- audio → text
- Parameters
- 105M
- License
- Health AI Developer Foundations Terms of Use
- At launch
- Open weights, Cloud platforms
- Conformer-based model with 105M parameters per its model card
- English-only, trained for medical terminology such as radiology dictation
- Available on Hugging Face and Vertex AI
MedGemma 1.5
MedGemma 1.5 is the next-generation 4B MedGemma model for medical image and text understanding. It adds support for 3D CT and MRI volumes, whole-slide histopathology, longitudinal chest X-rays and medical document extraction.
- Modalities
- text, image → text
- Parameters
- 4B
- License
- Health AI Developer Foundations Terms of Use
- At launch
- Open weights, Cloud platforms
- Accepts multiple slices or patches of high-dimensional scans alongside a text prompt
- Adds anatomical localization in chest X-rays
- Retains MedGemma 1's 2D image and text abilities
- Available on Hugging Face and Vertex AI; free for research and commercial use under HAI-DEF terms
Step3-VL-10B
Step3-VL-10B is a compact open-weight multimodal foundation model from StepFun that targets visual perception, complex reasoning, and human-centric alignment at a 10B-parameter scale.
- Modalities
- text, image → text
- Parameters
- 10B
- License
- Apache-2.0
- At launch
- Open weights
- 10B-parameter vision-language model with a base checkpoint also released
- Optional PaCoRe parallel reasoning mode aggregates 16 parallel rollouts
- FP8 weights and official vLLM support followed the initial release
Niji 7
Niji model tuned for Eastern and anime aesthetics with a major boost in coherency, clearer fine details, closer prompt following and improved text rendering and style reference performance.
- Modalities
- text, image → image
- At launch
- Consumer app
- Developed with Spellbrush
- More literal prompt interpretation than earlier Niji models
- Cleaner, flatter look highlighting line work
Scribe v2
Batch speech-to-text model for transcription, subtitling and captioning at scale, improving on Scribe v1 for long-form audio, pauses and tone changes and adding keyterm prompting.
- Modalities
- audio → text
- At launch
- API, Consumer app
- Keyterm prompting and built-in entity detection
- Automatic multi-language transcription
- Entity redaction and No Verbatim mode added April 2026
Jamba2
Second-generation Jamba family of open models built for enterprise reliability, released as a dense 3B model for devices and a mixture-of-experts Mini model.
- Modalities
- text → text
- Context
- 250K context
- Parameters
- 3B dense; Mini MoE 52B total, 12B active
- License
- Apache-2.0
- At launch
- Open weights, API
- 256K-token context window
- Mini aims at precise question answering without a reasoning model's overhead
- Released under Apache 2.0
Qwen3-VL-Embedding and Qwen3-VL-Reranker
Qwen3-VL-Embedding and Qwen3-VL-Reranker are open-weights multimodal retrieval models built on Qwen3-VL. They handle text, images, screenshots, and video in one framework for embedding and reranking.
- Modalities
- text, image, video → text
- Context
- 32K context
- Parameters
- 2B, 8B
- License
- Apache-2.0
- At launch
- Open weights
- Supports more than 30 languages
- Flexible embedding dimensions and customizable instructions
- Released in 2B and 8B sizes under the Apache 2.0 license
LFM2-2.6B-Transcript
A Liquid Nano fine-tuned from LFM2-2.6B for summarizing 30 to 60 minute meeting transcripts on-device, developed with AMD and showcased at CES running locally across Ryzen AI CPUs, GPUs, and NPUs.
- Modalities
- text → text
- Parameters
- 2.6B
- License
- LFM Open License v1.0
- At launch
- Open weights
- Outputs structured summaries with key points, decisions, and attributed action items
- Under 3 GB RAM usage reported for the AMD deployment
- Available for download on Hugging Face and LEAP
NousCoder-14B
Competitive olympiad programming model from Nous Research, post-trained from Qwen3-14B with reinforcement learning to solve contest-style coding problems and released with open weights under Apache 2.0.
- Modalities
- text → text
- Parameters
- 14B
- License
- Apache-2.0
- At launch
- Open weights
- Post-trained on Qwen3-14B
- Training context extended from 32K to 40K tokens during RL
- Released under Apache 2.0
LFM2.5-1.2B (Base, Instruct, JP)
The first LFM2.5 release, a 1.2B text family on the LFM2 architecture with pretraining extended from 10T to 28T tokens and expanded reinforcement learning, shipped as Base, Instruct, and Japanese-optimized checkpoints.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 1.17B
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- Checkpoints: LFM2.5-1.2B-Base, LFM2.5-1.2B-Instruct, and LFM2.5-1.2B-JP
- Announced alongside LFM2.5-VL-1.6B and LFM2.5-Audio-1.5B
- Open weights on Hugging Face and LEAP; Instruct also on Liquid Playground
- AMD and Nexa AI provide NPU-optimized versions
LFM2.5-Audio-1.5B
A native audio-language model in the LFM2.5 family that accepts and produces both speech and text, introducing a custom LFM-based audio detokenizer that Liquid AI says is 8x faster than LFM2's Mimi detokenizer.
- Modalities
- text, audio → text, audio
- Context
- 32K context
- Parameters
- 1.5B
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- llama.cpp-compatible GGUF checkpoints for CPU inference
- Improved ASR and TTS quality over LFM2-Audio-1.5B per Liquid AI
- Available on Hugging Face, LEAP, and Liquid Playground
LFM2.5-VL-1.6B
A refreshed 1.6B vision-language model on the updated LFM2.5 base backbone, released with the LFM2.5 family and tuned for multi-image comprehension, multilingual visual understanding, and instruction following.
- Modalities
- text, image → text
- Context
- 32K context
- Parameters
- 1.6B
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- Handles prompts in Arabic, Chinese, French, German, Japanese, Korean, and Spanish
- Succeeds LFM2-VL-1.6B
- Available on Hugging Face, LEAP, 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 January 2026." The AI Engineering, updated Jan 31, 2026. https://theaiengineering.com/releases/2026/01
HTML link
<a href="https://theaiengineering.com/releases/2026/01">AI models released in January 2026</a> (The AI Engineering)- Full dataset
- 1,167 releases from 55 labs
- Coverage
- to
- Updated
- License
- CC BY 4.0
- Publisher
- The AI Engineering