AI models released in March 2026
42 releases from 22 labs in March 2026, including LFM2.5-350M, Veo 3.1 Lite and LongCat-AudioDiT. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 42
- Labs
- 22
- Open weights
- 21
March 2026
42 releasesLFM2.5-350M
An updated 350M-parameter model with pretraining extended from 10T to 28T tokens and large-scale reinforcement learning, aimed at tool use, data extraction, and structured outputs on CPUs, NPUs, and GPUs.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 350M
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- Released as LFM2.5-350M-Base and post-trained LFM2.5-350M
- Available on Hugging Face, LEAP, and Liquid Playground
- New ecosystem partners at launch include Zetic, RunAnywhere, and Mirai
Veo 3.1 Lite
Veo 3.1 Lite is the most cost-effective model in Google's Veo 3.1 video generation lineup, aimed at high-volume developer use. It supports text-to-video and image-to-video at 720p or 1080p in landscape or portrait.
- Modalities
- text, image → video
- At launch
- API, Cloud platforms
- Clip durations of 4, 6 or 8 seconds
- 16:9 and 9:16 aspect ratios
- Paid preview in the Gemini API and Google AI Studio (veo-3.1-lite-generate-preview)
- Public preview on Vertex AI from April 2, 2026
LongCat-AudioDiT
LongCat-AudioDiT is a non-autoregressive diffusion text-to-speech model that operates directly in a waveform latent space, aimed at high-fidelity zero-shot voice cloning.
- Modalities
- text, audio → audio
- License
- MIT
- At launch
- Open weights
- Released in 1B and 3.5B sizes
- Diffusion TTS in the waveform latent space instead of mel-spectrograms
- Code and weights released under MIT
Qwen3.5-Omni
Qwen3.5-Omni is Qwen's omni-modal model series in Plus, Flash, and Light sizes, using a Thinker-Talker design for text and speech output. It adds semantic interruption, voice cloning, and coding from audio-visual instructions.
- Modalities
- text, image, audio, video → text, audio
- Context
- 250K context
- At launch
- API
- Supports 256K long-context input
- Available through the Offline API and the Realtime API
- Talker generates contextual speech from the Thinker's outputs
Granite 4.0 3B Vision
Vision-language model from IBM designed for enterprise-grade document data extraction, focused on specialized, complex extraction tasks that ultracompact models often struggle with.
- Modalities
- text, image → text
- License
- Apache-2.0
- At launch
- Open weights
- Focused on document data extraction
- Predecessor to Granite Vision 4.1
- Released under Apache 2.0
SAM 3.1
SAM 3.1 updates Segment Anything Model 3 with Object Multiplex, a shared-memory approach to joint multi-object tracking that makes real-time video detection and tracking significantly faster.
- Modalities
- text, image, video → image, video
- License
- SAM License
- At launch
- Open weights, API
- Drop-in replacement checkpoint for SAM 3, published on Hugging Face
- Faster multi-object video tracking without sacrificing accuracy
- Offered as sam-3.1 in the Meta Model API
Cohere Transcribe
Cohere's first transcription model, an automatic speech recognition model that converts audio to text in 14 languages through the Audio Transcriptions API.
- Modalities
- audio → text
- License
- Apache-2.0
- At launch
- API, Open weights
- Model name cohere-transcribe-03-2026
- Languages include English, German, French, Japanese, Korean, Chinese and Arabic
- Released under Apache 2.0
Gemini 3.1 Flash Live
Gemini 3.1 Flash Live is Google's audio-to-audio model for real-time dialogue and voice-first applications, offered in preview in the Live API and powering Gemini Live and Search Live.
- Modalities
- text, image, audio → text, audio
- Context
- 128K context
- At launch
- API, Consumer app
- Preview model gemini-3.1-flash-live-preview in the Gemini Live API via Google AI Studio
- Its multilingual support enabled Search Live's expansion to more than 200 countries and territories
- Input token limit of 131,072 per the Gemini API model page
Suno v5.5
Suno's most expressive model at its release, launched alongside Voices, Custom Models and My Taste features for capturing a creator's voice and personal sound.
- Modalities
- text, audio → audio
- At launch
- Consumer app
- Voices available to Pro and Premier subscribers
- Custom Models for a personalized sound
- My Taste personalization
LongCat-Next
LongCat-Next is a native multimodal model that processes text, vision, and audio under a single autoregressive objective, handling both understanding and generation in one model with discrete tokenizers.
- Modalities
- text, image, audio → text, image, audio
- License
- MIT
- At launch
- Open weights
- Described by Meituan as an A3B-size model (about 3B activated parameters)
- Discrete Native Autoregression (DiNA) paradigm for unified modalities
- Model weights and tokenizers released under MIT
Lyria 3 Pro
Lyria 3 Pro is Google DeepMind's higher-tier Lyria 3 music model for full-length songs. It generates tracks up to three minutes long and understands song structure, so users can prompt for intros, verses, choruses and bridges.
- Modalities
- text → audio
- At launch
- Consumer app, API, Cloud platforms
- Tracks up to 3 minutes, versus 30 seconds for Lyria 3 in the Gemini app
- Public preview on Vertex AI; available in the Gemini API and AI Studio as lyria-3-pro-preview
- Longer generations in the Gemini app for paid subscribers, plus Google Vids and ProducerAI
- All outputs carry a SynthID watermark
MolmoWeb
Open visual web agent built on Molmo 2 that navigates websites, fills out forms and retrieves information in a browser, designed for self-hosting.
- Modalities
- text, image → text
- Parameters
- 4B and 8B
- License
- Apache-2.0
- At launch
- Open weights
- Released with weights, training data and evaluation tools
- Training data includes 36K human task trajectories
- Released under Apache 2.0
Voxtral TTS
Mistral's first text-to-speech model, a 4B model for expressive multilingual speech in nine languages with zero-shot voice adaptation from a short voice prompt.
- Modalities
- text, audio → audio
- Parameters
- 4B
- License
- CC-BY-NC-4.0
- At launch
- Open weights, API
- Streaming with about 90 ms time to first audio
- Weights with reference voices released under CC BY-NC 4.0
- API model ID voxtral-mini-tts-2603
LongCat-Flash-Prover
LongCat-Flash-Prover is a 560B-parameter mixture-of-experts model for formal reasoning in Lean 4, using agentic tool-integrated reasoning for auto-formalization, proof sketching, and theorem proving.
- Modalities
- text → text
- Parameters
- MoE: 560B total
- License
- MIT
- At launch
- Open weights
- Targets auto-formalization, lemma-style sketching, and whole-proof generation
- Trained with the Hierarchical Importance Sampling Policy Optimization (HisPO) algorithm
- Weights released under MIT
MAI-Image-2
MAI-Image-2 is Microsoft AI's second in-house text-to-image model, focused on photorealism, reliable in-image text and detailed scene generation. It launched in MAI Playground and began rolling out to Copilot and Bing Image Creator.
- Modalities
- text → image
- At launch
- API, Consumer app
- Available in MAI Playground at launch
- API access first for select customers, then in Microsoft Foundry from April 2, 2026
- Rolling out to Copilot and Bing Image Creator
MiMo-V2-Omni
MiMo-V2-Omni is Xiaomi's omnimodal agentic foundation model, fusing dedicated image, video, and audio encoders into a shared backbone so it can perceive several modalities at once and act on them.
- Modalities
- text, image, video, audio → text
- Context
- 250K context
- At launch
- API, Consumer app
- Supports up to 256K tokens of context
- Trained to connect perception with next actions for agent use
- Available through the MiMo API Platform
MiMo-V2-Pro
MiMo-V2-Pro is Xiaomi's flagship foundation model for agentic workloads, scaled to more than 1T total parameters and designed to orchestrate complex workflows and production engineering tasks.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: over 1T total, 42B active
- At launch
- API, Consumer app
- Hybrid attention with a 7:1 ratio of sliding-window to global attention
- Supports up to 1M tokens of context
- Lightweight multi-token prediction layer for faster generation
- Released alongside MiMo-V2-Omni and MiMo-V2-TTS
MiMo-V2-TTS
MiMo-V2-TTS is Xiaomi's large-scale speech synthesis model, built on an in-house audio tokenizer and multi-codebook speech-text modeling and pretrained on more than 100 million hours of speech data.
- Modalities
- text → audio
- At launch
- API, Consumer app
- Supports style control, singing, and voice cloning
- Refined with multi-dimensional reinforcement learning
- Launch demo covers Northeastern Mandarin, Sichuanese, and Cantonese speech
MiniMax M2.7
MiniMax M2.7 is a language model for software engineering and professional office work that MiniMax describes as its first model deeply involved in its own development, building the agent scaffolding, skills, and memory used in its RL experiments.
- Modalities
- text → text
- License
- Non-Commercial License
- At launch
- API, Open weights
- Released on the API as MiniMax-M2.7 and M2.7-highspeed
- Weights open-sourced on April 12, 2026
- Weights license permits non-commercial use; commercial use requires written authorization
- Uses Agent Teams, complex Skills, and dynamic tool search
Qianfan-OCR
Qianfan-OCR is a 4B end-to-end document intelligence model from Baidu's Qianfan team that unifies document parsing, layout analysis, table and formula recognition, chart understanding, and key information extraction in one model.
- Modalities
- image, text → text
- Parameters
- 4B
- License
- Apache-2.0
- At launch
- Open weights
- Optional Layout-as-Thought phase reasons about bounding boxes, element types, and reading order before output
- Language model based on Qwen3-4B with 32K context, extendable to 131K
- Supports 192 languages
- Trained on 1,024 Kunlun P800 chips
GPT-5.4 mini and GPT-5.4 nano
GPT-5.4 mini and nano are smaller GPT-5.4 models for high-volume workloads; mini targets coding, subagents, and computer use at more than twice GPT-5 mini's speed, while nano handles classification, extraction, and ranking.
- Modalities
- text, image → text
- Context
- 400K context
- At launch
- API, Consumer app
- API model IDs gpt-5.4-mini and gpt-5.4-nano; both list a 400,000-token context window
- Mini is in the API, Codex, and ChatGPT (Free and Go via Thinking; rate-limit fallback for others)
- Nano is API-only and supports compaction but not tool search or computer use
- Mini supports tool search, built-in computer use, and compaction
Midjourney V8 (Alpha)
Early V8 image model opened to the community on alpha.midjourney.com, with better adherence to detailed directions, improved text rendering, more coherent images and roughly five times faster generation.
- Modalities
- text, image → image
- At launch
- Consumer app
- New --hd mode renders natively at 2K
- Backward compatible with V7 personalization profiles, moodboards and srefs
- Scheduled for deprecation two weeks after V8.1 became the default
Leanstral
Mistral's first open-source code agent for the Lean 4 proof assistant, a highly sparse model built for formal proof engineering in realistic repositories rather than isolated math problems.
- Modalities
- text → text
- Context
- 250K context
- Parameters
- MoE: 119B total, 6.5B active
- License
- Apache-2.0
- At launch
- Open weights, API
- Available in an agent mode within Mistral Vibe and through an API endpoint
- Announced with a new FLTEval evaluation suite for proof engineering
- API model ID labs-leanstral-2603
Mistral Small 4
Hybrid model that unifies the reasoning of Magistral, the vision of Pixtral and the agentic coding of Devstral in one open-weight model with configurable reasoning effort.
- Modalities
- text, image → text
- Context
- 250K context
- Parameters
- MoE: 119B total, 6B active
- License
- Apache-2.0
- At launch
- Open weights, API
- 128 experts with 4 active per token
- Released as Mistral joined the NVIDIA Nemotron Coalition as a founding member
- API model ID mistral-small-2603
Nemotron 3 Nano 4B
Compact member of the Nemotron 3 family with about 4B parameters and a long context window, for local and edge deployments.
- Modalities
- text → text
- Context
- 262K context
- Parameters
- 3.97B
- License
- NVIDIA Nemotron Open Model License
- At launch
- Open weights
- Context length up to 262K tokens
- Released under the NVIDIA Nemotron Open Model License
- Smaller sibling of Nemotron 3 Nano 30B
GLM-5-Turbo
GLM-5-Turbo is a GLM-5-based model that Z.ai optimized for OpenClaw-style agent workloads. It strengthens calls to external tools and skills and is more stable on long, multi-step tasks.
- Modalities
- text → text
- Context
- 200K context
- At launch
- API
- Improved decomposition of complex instructions and multi-agent coordination
- Up to 128K output tokens
- Available through the BigModel and Z.ai APIs
Mistral Moderation 2
Second-generation Mistral moderation model with a 128K-token context window and jailbreak detection, built for complex multilingual content including long multi-turn conversations.
- Modalities
- text → text
- Context
- 125K context
- At launch
- API
- API model ID mistral-moderation-2603
- Released with custom guardrails for Agents and Conversations
- Mistral's model card lists March 1, 2026
LongCat-Flash-Omni-2603
LongCat-Flash-Omni-2603 is an upgraded version of LongCat-Flash-Omni, an end-to-end omni-modal interaction model with more natural spoken dialogue and stronger perception across vision, speech, and text.
- Modalities
- text, image, audio, video → text, audio
- At launch
- API, Consumer app
- Available in LongCat Chat and on the API as model=LongCat-Flash-Omni-2603
- Adds native voice function calling from audio commands
- Retired from the LongCat API on May 29, 2026
Nemotron 3 Super
Mid-size Nemotron 3 model built on a hybrid latent mixture-of-experts architecture with Mamba-2, MoE and attention layers, aimed at agentic workflows and long-context reasoning.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 120B total, 12B active
- License
- NVIDIA Nemotron Open Model License
- At launch
- Open weights
- Adds multi-token prediction layers
- Supports up to 1M tokens of context
- Trained between December 2025 and March 2026
Reka Edge (reka-edge-2603)
A new version of Reka's 7B vision-language model for physical AI and edge devices, pairing a ConvNeXt V2 vision encoder with a transformer trained from scratch and using 64 tokens per image tile.
- Modalities
- text, image, video → text
- Context
- 16K context
- Parameters
- 7B (657M vision encoder, 6.4B transformer)
- License
- Business Source License 1.1 (Reka Edge 2603 License)
- At launch
- API, Consumer app, Open weights
- API model ID reka-edge-2603 with a 16,384-token context per Reka docs
- Targets image understanding, video analysis, object detection and agentic tool use
- Open weights on Hugging Face; the license permits commercial use below an annual revenue threshold
- Runs on NVIDIA Jetson, Apple Silicon Macs, PCs and Snapdragon devices; vLLM plugin available
Gemini Embedding 2
Gemini Embedding 2 is Google's first natively multimodal embedding model, mapping text, images, video, audio and documents into a single embedding space for multimodal retrieval and classification.
- Modalities
- text, image, video, audio → embedding
- Context
- 8K context
- At launch
- API, Cloud platforms
- Released in public preview in the Gemini API and Vertex AI; generally available on Apr 22, 2026
- Accepts interleaved multimodal input and captures semantic intent across over 100 languages
- Default 3072-dimension output with Matryoshka Representation Learning for smaller sizes
- Input token limit of 8,192 per the Gemini API model page
Grok 4.20 and Grok 4.20 Multi-agent
Grok 4.20 is a model family on the xAI API with reasoning and non-reasoning variants, plus a Multi-agent version in which several agents collaborate on research tasks and a leader agent synthesizes the answer.
- Modalities
- text, image → text
- Context
- 1M context
- At launch
- API
- API IDs: grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning, grok-4.20-multi-agent-0309
- 1M-token context window with text and image input
- Multi-agent reasoning effort sets the number of agents (4 or 16)
Runway Characters (GWM-1 Avatars)
Real-time video agent API for fully custom conversational characters built on the GWM-1 world model, generating expressive avatars from a single image with control over voice and personality.
- Modalities
- image, audio, text → video, audio
- At launch
- API, Enterprise
- Built on the GWM-1 general world model announced December 2025
- Zero fine-tuning; any visual style from one image
- Available via the Runway API with a web demo
Granite 4.0 1B Speech
Compact speech-language model for multilingual automatic speech recognition and bidirectional speech translation, with half the parameters of Granite Speech 3.3 2B.
- Modalities
- audio → text
- License
- Apache-2.0
- At launch
- Open weights
- Built for resource-constrained devices
- Supports Japanese and keyword-biased recognition
- Released under Apache 2.0
GPT-5.4 (Thinking and Pro)
GPT-5.4 is a frontier model for professional work that folds in GPT-5.3-Codex coding capabilities, adds native computer use and tool search in the API, and ships in ChatGPT as GPT-5.4 Thinking alongside GPT-5.4 Pro.
- Modalities
- text, image → text
- Context
- 1.1M context
- At launch
- Consumer app, API, Enterprise
- API model IDs gpt-5.4 and gpt-5.4-pro (Pro in the Responses API only)
- 1,050,000-token context window per the API model page, with native compaction support
- GPT-5.4 Thinking for Plus, Team, and Pro, replacing GPT-5.2 Thinking; GPT-5.4 Pro for Pro and Enterprise
- Also rolled out in Codex, with experimental 1M-token context support
LTX-2.3
Upgraded LTX-2 model with a 22B-parameter architecture, sharper details from a new VAE, cleaner audio, stronger image-to-video and better prompt understanding, offered as Pro and Fast API variants.
- Modalities
- text, image, video, audio → video, audio
- Parameters
- 22B
- License
- LTX-2 Community License
- At launch
- Open weights, API, Consumer app
- 22B parameters, up from 19B
- Up to 4K at 50 fps and up to 20 seconds per clip
- Native 9:16 portrait video
Olmo Hybrid
Fully open 7B model family that mixes transformer attention with linear recurrent layers, released with a controlled comparison against Olmo 3 7B.
- Modalities
- text → text
- Parameters
- 7B
- License
- Apache-2.0
- At launch
- Open weights
- Combines transformers and linear RNNs
- Developed through 1B and 7B scale experiments
- Released under Apache 2.0
MiniMax Music 2.5+
Music 2.5+ extends MiniMax Music 2.5 with instrumental music generation, producing tracks without vocals in styles from classical orchestration and modern electronic to ambient sound and natural soundscapes.
- Modalities
- text → audio
- At launch
- API, Consumer app
- Listed in the API release notes as Music2.5+
- Targets uses such as meditation, sleep aids, advertising, and game scoring
- Style switching requires no additional tuning
Phi-4-reasoning-vision-15B
Phi-4-reasoning-vision-15B is Microsoft's open-weight multimodal reasoning model, pairing a SigLIP-2 vision encoder with the Phi-4-Reasoning backbone. It reasons step by step for math and science and answers perception tasks directly.
- Modalities
- text, image → text
- Context
- 16K context
- Parameters
- 15B
- License
- MIT
- At launch
- Open weights, API
- 15B parameters with a 16,384-token context length
- Built on the Phi-4-Reasoning backbone with a SigLIP-2 vision encoder
- Available on Microsoft Foundry, Hugging Face and GitHub
Gemini 3.1 Flash-Lite
Gemini 3.1 Flash-Lite is the first Flash-Lite model in the Gemini 3 series, Google's fastest and most cost-efficient Gemini 3 model for high-volume workloads such as translation and content moderation.
- Modalities
- text, image, audio, video → text
- Context
- 1M context
- At launch
- API, Cloud platforms
- Preview in the Gemini API via Google AI Studio and in Vertex AI at launch
- Stable gemini-3.1-flash-lite became generally available on May 7, 2026
- Input token limit of 1,048,576 per the Gemini API model page
GPT-5.3 Instant
GPT-5.3 Instant is an update to ChatGPT's most-used model, aimed at fewer unnecessary refusals and caveats, better-synthesized answers from web search, a more direct conversational style, and stronger writing.
- Modalities
- text, image → text
- Context
- 125K context
- At launch
- Consumer app, API
- Available to all ChatGPT users and in the API as gpt-5.3-chat-latest
- Replaced GPT-5.2 Instant, which was retired from ChatGPT on June 3, 2026
- 128,000-token context window per the gpt-5.3-chat-latest API model page
- A March 16, 2026 update improved follow-up tone and reduced teaser-style phrasing
Qwen3.5 small models
The Qwen3.5 small models are four compact open-weights models, Qwen3.5-9B, 4B, 2B, and 0.8B, completing the Qwen3.5 lineup. Like the rest of the series, they are causal language models with a vision encoder.
- Modalities
- text, image → text
- Context
- 256K context
- Parameters
- 0.8B, 2B, 4B, 9B
- License
- Apache-2.0
- At launch
- Open weights
- Native 262,144-token context, extensible to 1,010,000 tokens
- Base checkpoints released alongside the instruct models
- Weights on Hugging Face and ModelScope under the Apache 2.0 license
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 March 2026." The AI Engineering, updated Mar 31, 2026. https://theaiengineering.com/releases/2026/03
HTML link
<a href="https://theaiengineering.com/releases/2026/03">AI models released in March 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