AI models released in February 2026
27 releases from 15 labs in February 2026, including Nano Banana 2 (Gemini 3.1 Flash Image), pplx-embed-v1 and pplx-embed-context-v1 and LFM2-24B-A2B. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 27
- Labs
- 15
- Open weights
- 12
February 2026
27 releasesNano Banana 2 (Gemini 3.1 Flash Image)
Nano Banana 2 (Gemini 3.1 Flash Image) is a high-efficiency image generation and editing model that brings Nano Banana Pro's world knowledge and reasoning to Flash-level speed.
- Modalities
- text, image, video → text, image
- Context
- 128K context
- At launch
- API, Cloud platforms, Consumer app
- Supports aspect ratios and resolutions from 512px to 4K
- Replaced Nano Banana Pro as the default image model across the Gemini app's Fast, Thinking and Pro models
- Stable gemini-3.1-flash-image became generally available on May 28, 2026, adding video-to-image generation
pplx-embed-v1 and pplx-embed-context-v1
Two text embedding model families, pplx-embed-v1 for standard dense retrieval and pplx-embed-context-v1 for document-aware chunk embeddings, each in 0.6B and 4B sizes and built by diffusion continued pretraining of Qwen3.
- Modalities
- text → embedding
- Parameters
- 0.6B and 4B
- License
- MIT
- At launch
- API, Open weights
- 32K-token context, native INT8 and binary embeddings, Matryoshka dimensions, no instruction prefixes
- API model IDs pplx-embed-v1-0.6b, pplx-embed-v1-4b, pplx-embed-context-v1-0.6b and pplx-embed-context-v1-4b
- Perplexity's Embeddings API became available the same month (February 2026)
- Weights on Hugging Face under the MIT license
LFM2-24B-A2B
An early checkpoint of the largest LFM2 model, a sparse mixture-of-experts instruct model with 24B total and 2.3B active parameters, designed to fit in 32GB of RAM for laptops, desktops, and cloud deployment.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- MoE: 24B total, 2.3B active
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- 40 layers with 64 experts per MoE block and top-4 routing
- Instruct model without reasoning traces; trained on 17T tokens at release
- Day-zero support in llama.cpp, vLLM, and SGLang
- Weights on Hugging Face; testable on Liquid Playground
Qwen3.5 medium models
The Qwen3.5 medium models are Qwen3.5-122B-A10B, Qwen3.5-35B-A3B, and the dense Qwen3.5-27B, released after the flagship Qwen3.5-397B-A17B. They share the Qwen3.5 hybrid architecture with native vision.
- Modalities
- text, image → text
- Context
- 256K context
- Parameters
- MoE: 122B total, 10B active; MoE: 35B total, 3B active; 27B dense
- License
- Apache-2.0
- At launch
- Open weights
- Native 262,144-token context, extensible to 1,010,000 tokens
- Qwen3.5-35B-A3B also released as a base model
- Weights under the Apache 2.0 license
gpt-realtime-1.5 and gpt-audio-1.5
GPT-Realtime-1.5 is an updated speech-to-speech model for voice agents and customer support in the Realtime API, released with gpt-audio-1.5, an audio-in, audio-out model for the Chat Completions API.
- Modalities
- text, audio, image → text, audio
- At launch
- API
- API model IDs gpt-realtime-1.5 (v1/realtime) and gpt-audio-1.5 (v1/chat/completions)
- gpt-realtime-1.5 lists a 32,000-token context window; gpt-audio-1.5 lists 128,000 tokens
- gpt-realtime-1.5 accepts image input in addition to text and audio
Gemini 3.1 Pro
Gemini 3.1 Pro is an upgraded Gemini 3 Pro that carries the core intelligence behind the February 2026 Deep Think update, released in preview across the Gemini API, Vertex AI, the Gemini app and NotebookLM.
- Modalities
- text, image, audio, video → text
- Context
- 1M context
- At launch
- API, Cloud platforms, Consumer app
- Preview model gemini-3.1-pro-preview, plus a gemini-3.1-pro-preview-customtools endpoint that prioritizes custom tools
- Rolled out with higher limits for Google AI Pro and Ultra subscribers in the Gemini app
- Input token limit of 1,048,576 per the Gemini API model page
Lyria 3
Lyria 3 is a Google DeepMind music generation model that launched in beta in the Gemini app. It creates 30-second tracks with automatically written lyrics and cover art from a text description, photo or video.
- Modalities
- text, image, video → audio
- At launch
- Consumer app, API
- Available to Gemini app users 18+ in eight languages, with higher limits for paid plans
- Also used in YouTube Dream Track for Shorts soundtracks
- Every Gemini-generated track carries a SynthID watermark
- Reached the Gemini API as lyria-3-clip-preview on March 25, 2026
Claude Sonnet 4.6
Claude Sonnet 4.6 is a full upgrade of Anthropic's Sonnet model across coding, computer use, long-context reasoning, agent planning, knowledge work, and design. It became the default model for Free and Pro users on claude.ai.
- Modalities
- text, image → text
- Context
- 1M context
- At launch
- API, Consumer app, Cloud platforms
- API model ID claude-sonnet-4-6
- 1M-token context window launched in beta; out of beta on March 13, 2026
- Supports both adaptive thinking and extended thinking
- Available on all Claude plans, the Claude API, and all major cloud platforms
Recraft V4
Ground-up rebuild of Recraft's image model tuned with designers for visual taste, prompt accuracy and output quality, offered as V4 and a higher-resolution V4 Pro.
- Modalities
- text → image
- At launch
- Consumer app, API
- V4 Pro generates higher-resolution, print-ready images
- V4 Vector and V4 Pro Vector produce editable SVG files
- Available in Recraft and via API
Tiny Aya
Compact multilingual open-weight models from Cohere Labs that are small enough to run locally on consumer hardware and phones, with a 3.35B base covering 70+ languages.
- Modalities
- text → text
- Parameters
- 3.35B
- License
- CC-BY-NC-4.0
- At launch
- Open weights
- TinyAya-Global instruct model supports 67 languages
- Regional variants Earth, Fire and Water published on Hugging Face
- Released with a new multilingual fine-tuning dataset and benchmarks
Qwen3.5
Qwen3.5 is Alibaba's model series that opened with the open-weights Qwen3.5-397B-A17B, a native vision-language model. It uses a hybrid architecture that combines linear attention through Gated Delta Networks with a sparse mixture-of-experts design.
- Modalities
- text, image, video → text
- Context
- 256K context
- Parameters
- MoE: 397B total, 17B active
- License
- Apache-2.0
- At launch
- API, Consumer app, Open weights
- Native 262,144-token context, extensible to 1,010,000 tokens
- Language and dialect support expanded from 119 to 201
- Hosted Qwen3.5-Plus offers a 1M context window by default with built-in tools
- Weights under the Apache 2.0 license
Seed2.0
Seed2.0 is ByteDance Seed's model series for complex real-world tasks, offered as Pro, Lite, and Mini general agent models plus a dedicated Code model, with stronger visual understanding, long-context handling, and multi-step instruction following.
- Modalities
- text, image, video → text
- At launch
- Consumer app, API
- Three general agent sizes (Pro, Lite, Mini) plus a Code model
- Pro and Code models launched in the Doubao app and TRAE
- Full series available through the Volcano Engine API
- Upgraded document, chart, and long-video understanding over Seed1.8
Seedream 5.0 Lite
Seedream 5.0 Lite is an image creation model that brings deeper reasoning and world knowledge to generation and editing, and adds real-time web search for time-sensitive content, building on Seedream 4.0 and 4.5.
- Modalities
- text, image → image
- At launch
- Consumer app, API
- Adds real-time search enhancement to image generation
- Improved knowledge reasoning, editing response, and consistency
- Launched on Jimeng AI and Volcano Ark's experience center, with Doubao in internal testing
GLM-5
GLM-5 is Z.ai's open-weights model for complex systems engineering and long-running agent tasks. It scales to 744B total parameters from 355B in GLM-4.5 and integrates DeepSeek Sparse Attention to reduce deployment cost.
- Modalities
- text → text
- Context
- 200K context
- Parameters
- MoE: 744B total, 40B active
- License
- MIT
- At launch
- API, Open weights
- Pre-training data increased from 23T to 28.5T tokens
- Trained with slime, Z.ai's asynchronous RL infrastructure
- Up to 128K output tokens
- Weights under the MIT license
GPT-5.3-Codex-Spark
GPT-5.3-Codex-Spark is a research preview of a smaller GPT-5.3-Codex designed for real-time coding in Codex, served on Cerebras hardware at more than 1,000 tokens per second.
- Modalities
- text → text
- Context
- 125K context
- At launch
- Consumer app
- Research preview for ChatGPT Pro users in the Codex app, CLI, and VS Code extension
- Text-only at launch with a 128K context window
- Runs on the Cerebras Wafer Scale Engine 3, the first milestone of the OpenAI and Cerebras partnership
- API access limited to a small set of design partners at launch
MiniMax M2.5
MiniMax M2.5 is a language model trained with reinforcement learning across hundreds of thousands of real-world environments, targeting coding, agentic tool use and search, and office work, and completing agentic tasks faster than M2.1.
- Modalities
- text → text
- License
- MiniMax Model License
- At launch
- API, Open weights
- Offered as M2.5 and M2.5-Lightning (highspeed), identical in capability but differing in speed
- M2.5-Lightning runs at a steady 100 tokens per second
- Weights on Hugging Face under the MiniMax Model License dated February 13, 2026
- Trained with the CISPO algorithm plus a process reward mechanism
Seedance 2.0
Seedance 2.0 is ByteDance Seed's video creation model built on a unified multimodal audio-video joint generation architecture, accepting text, image, audio, and video inputs with broad reference and editing capabilities.
- Modalities
- text, image, audio, video → video, audio
- At launch
- Consumer app, API
- Outputs up to 15 seconds of multi-shot video with dual-channel audio
- Supports video editing and extension guided by multimodal references
- Available on Jimeng AI, Doubao, and Volcano Ark; BytePlus offers it as Dreamina Seedance 2.0
Eleven v3 Conversational
Context-aware text-to-speech model built on Eleven v3 and optimized for real-time dialogue, powering Expressive Mode for ElevenAgents with speech that reflects intent and emotion.
- Modalities
- text → audio
- At launch
- API, Enterprise
- Model ID eleven_v3_conversational
- Paired with a new turn-taking system using Scribe v2 Realtime
- Supports 90+ languages in Expressive Mode
Qwen-Image-2.0
Qwen-Image-2.0 is Qwen's next-generation foundational image generation model, unifying image generation and editing in a single mode. It targets professional typography such as infographics, slides, posters, and comics.
- Modalities
- text, image → image
- At launch
- Consumer app
- Supports 1K-token instructions
- Native 2K resolution
- Smaller model size with faster inference than its predecessor; available on Qwen Chat
Claude Opus 4.6
Claude Opus 4.6 is an upgrade to Claude Opus 4.5 that plans more carefully, sustains agentic tasks longer, and works more reliably in large codebases. It introduced adaptive thinking and was the first Opus-class model with 1M-token context.
- Modalities
- text, image → text
- Context
- 1M context
- At launch
- API, Consumer app, Cloud platforms
- API model ID claude-opus-4-6
- 1M-token context window launched in beta; out of beta on March 13, 2026
- 128K max output tokens
- Compaction API launched in beta alongside the model
GPT-5.3-Codex
GPT-5.3-Codex is an agentic coding model that combines GPT-5.2-Codex coding performance with GPT-5.2 reasoning and professional knowledge in one model, running 25% faster and steerable while it works.
- Modalities
- text, image → text
- Context
- 400K context
- At launch
- Consumer app
- Available with paid ChatGPT plans in the Codex app, CLI, IDE extension, and web
- Released to the Responses API as gpt-5.3-codex on February 24, 2026
- First OpenAI model classified as High capability for cybersecurity under its Preparedness Framework
- 400,000-token context window per the API model page
LongCat-Flash-Lite
LongCat-Flash-Lite is a non-thinking mixture-of-experts model that uses an N-gram embedding table to improve parameter efficiency and inference speed, with a focus on tool use and coding.
- Modalities
- text → text
- Context
- 250K context
- Parameters
- MoE: 68.5B total, about 3B active
- License
- MIT
- At launch
- API, Open weights
- 68.5B total parameters with about 3B activated per token
- 256K context length through YaRN
- N-gram embedding table reduces I/O bottlenecks in MoE layers
- Weights on Hugging Face under MIT
Voxtral Mini Transcribe V2
Batch speech-to-text model in the Voxtral Transcribe 2 family that adds speaker diarization, context biasing and word-level timestamps across 13 languages.
- Modalities
- audio → text
- At launch
- API
- API model ID voxtral-mini-2602
- Launched with an audio playground in Mistral Studio
- Deprecated October 31, 2026 in favor of Voxtral Mini Transcribe 3
Voxtral Realtime
Open-weight live transcription model with a streaming architecture that transcribes audio as it arrives, with delay configurable down to under 200 milliseconds.
- Modalities
- audio → text
- Parameters
- 4B
- License
- Apache-2.0
- At launch
- Open weights, API
- Released as Voxtral-Mini-4B-Realtime-2602 on Hugging Face
- Supports 13 languages
- API model ID voxtral-mini-transcribe-realtime-2602
GLM-OCR
GLM-OCR is a compact multimodal OCR model for complex document understanding, built on the GLM-V encoder-decoder architecture with a CogViT visual encoder and a GLM-0.5B decoder. It is designed for fast, high-concurrency deployment.
- Modalities
- text, image → text
- Parameters
- 0.9B
- License
- MIT
- At launch
- API, Open weights
- Only 0.9B parameters, deployable with vLLM, SGLang, and Ollama
- Uses multi-token prediction loss and full-task reinforcement learning
- Handles handwriting, tables, seals, and vertical text
Qwen3-Coder-Next
Qwen3-Coder-Next is an open-weights model designed for coding agents and local development, built on Qwen3-Next-80B-A3B-Base. It was agentically trained at scale with executable task synthesis, environment interaction, and reinforcement learning.
- Modalities
- text → text
- Context
- 256K context
- Parameters
- MoE: 80B total, 3B active
- License
- Apache-2.0
- At launch
- Open weights
- Hybrid attention and mixture-of-experts architecture from Qwen3-Next
- 262,144-token native context
- Weights under the Apache 2.0 license
Step 3.5 Flash
Step 3.5 Flash is StepFun's open-weight sparse mixture-of-experts reasoning model built for fast agentic and coding work, using three-way multi-token prediction and hybrid sliding-window attention for efficient long-context inference.
- Modalities
- text → text
- Context
- 250K context
- Parameters
- MoE: 196B total, 11B active
- License
- Apache-2.0
- At launch
- API, Open weights
- 196B total parameters with about 11B activated per token
- 256K context window using a 3:1 sliding-window to full-attention layer ratio
- 3-way multi-token prediction (MTP-3) for faster decoding
- Weights on Hugging Face and ModelScope under Apache-2.0; pre-training and mid-training checkpoints followed on March 4, 2026
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 February 2026." The AI Engineering, updated Feb 26, 2026. https://theaiengineering.com/releases/2026/02
HTML link
<a href="https://theaiengineering.com/releases/2026/02">AI models released in February 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