AI models released in July 2025
34 releases from 20 labs in July 2025, including Command A Vision, FLUX.1 Krea [dev] and Codestral 25.08. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 34
- Labs
- 20
- Open weights
- 26
July 2025
34 releasesCommand A Vision
Cohere's first Command model that understands images, built for enterprise visual tasks such as charts, diagrams, PDFs, slides and real-world scenes.
- Modalities
- text, image → text
- Context
- 125K context
- Parameters
- 112B
- License
- CC-BY-NC-4.0
- At launch
- API, Open weights
- Deploys privately on two A100 GPUs or one H100 with 4-bit quantization
- Keeps Command A's RAG with citations
- Research weights released as command-a-vision-07-2025
FLUX.1 Krea [dev]
Open-weight text-to-image model developed with Krea AI as the open version of Krea 1, trained to avoid the oversaturated AI look and produce more realistic, diverse images.
- Modalities
- text → image
- License
- FLUX.1 [dev] Non-Commercial License
- At launch
- Open weights, Cloud platforms
- Architecturally compatible with the FLUX.1 [dev] ecosystem
- Weights on Hugging Face, commercial licenses via the BFL Licensing Portal
- Positioned as a base model for downstream fine-tuning
Codestral 25.08
Update to the Codestral fill-in-the-middle model, announced as part of Mistral's coding stack for enterprises with more accepted completions and fewer runaway generations.
- Modalities
- text → text
- Context
- 125K context
- At launch
- API, Enterprise
- Deployable in cloud, VPC or on-premises environments
- Improved chat-mode instruction following
- API model ID codestral-2508
GLM-4.5
GLM-4.5 is Z.ai's open-weights foundation model series for agents, unifying reasoning, coding, and agentic abilities. It comes in GLM-4.5 and the more compact GLM-4.5-Air, both with thinking and non-thinking modes.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 355B total, 32B active; Air 106B total, 12B active
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Pre-trained on 15T tokens of general data, then on code, reasoning, and agent data
- 128K context length
- Weights on Hugging Face and ModelScope under the MIT license
Wan2.2
Wan2.2 is the open video generation family from Alibaba's Wan team, adding mixture-of-experts text-to-video and image-to-video models and a compact 5B text-and-image-to-video model.
- Modalities
- text, image → video
- Parameters
- MoE: A14B (T2V, I2V); 5B (TI2V)
- License
- Apache-2.0
- At launch
- Open weights
- T2V-A14B and I2V-A14B support 480P and 720P
- TI2V-5B uses a high-compression VAE and supports 720P
- Integrated into Diffusers and ComfyUI at launch; Apache 2.0 license
HunyuanWorld 1.0
HunyuanWorld 1.0 is Tencent's open model for generating immersive, explorable 360-degree 3D worlds from text or images, combining panoramic proxy generation, semantic layering, and hierarchical 3D reconstruction.
- Modalities
- text, image → 3d
- License
- Tencent HunyuanWorld-1.0 Community License Agreement
- At launch
- Open weights
- Tencent describes it as the first open-source, simulation-capable immersive 3D world generation model
- Quantized HunyuanWorld-1.0-lite for consumer GPUs added August 15, 2025
Aleph
In-context video model for multi-task editing that adds, removes and transforms objects, generates new angles of a scene and modifies style and lighting on an input video.
- Modalities
- text, image, video → video
- At launch
- Consumer app
- Available to all paid plans from July 31, 2025
- Generates new camera angles from existing footage
- Succeeded by Aleph 2.0 in May 2026
Llama-3.3-Nemotron-Super-49B-v1.5
Significantly upgraded version of Llama-3.3-Nemotron-Super-49B, a reasoning model NVIDIA derived from Meta's Llama-3.3-70B-Instruct using neural architecture search to cut memory needs while keeping accuracy.
- Modalities
- text → text
- Parameters
- 49B
- License
- NVIDIA Open Model License
- At launch
- Open weights
- Trained between November 2024 and July 2025
- Post-trained for reasoning, chat preferences and agentic tasks such as RAG and tool calling
- Part of the Llama Nemotron collection
Qwen3-235B-A22B-Thinking-2507
Qwen3-235B-A22B-Thinking-2507 is an updated thinking-mode version of Qwen3-235B-A22B, released as a separate reasoning model in the Qwen3-2507 series, which split the hybrid Qwen3 model into dedicated instruct and thinking models.
- Modalities
- text → text
- Context
- 256K context
- Parameters
- MoE: 235B total, 22B active
- License
- Apache-2.0
- At launch
- Open weights
- Native 262,144-token context
- Qwen3-30B-A3B-Thinking-2507 followed on July 31 and Qwen3-4B-Thinking-2507 on August 6, 2025
- Weights under the Apache 2.0 license
Step 3
Step 3 is StepFun's multimodal reasoning model built on a mixture-of-experts architecture and co-designed with Multi-Matrix Factorization Attention and Attention-FFN Disaggregation to reduce decoding costs.
- Modalities
- text, image → text
- Context
- 64K context
- Parameters
- MoE: 321B total, 38B active
- License
- Apache-2.0
- At launch
- Open weights
- 321B total parameters (316B language model) with 38B active per token
- Maximum context length of 65,536 tokens
- Model weights were uploaded to Hugging Face on July 31, 2025 under Apache-2.0
Qwen-MT
Qwen-MT (qwen-mt-turbo) is Qwen's translation model on the Qwen API, built on Qwen3 and trained on trillions of multilingual and translation tokens. It uses a lightweight mixture-of-experts architecture for low latency.
- Modalities
- text → text
- At launch
- API
- Translates across 92 major languages and dialects
- Supports terminology intervention, domain prompts, and translation memory
- Available through Alibaba Cloud Model Studio
Seed LiveInterpret 2.0
Seed LiveInterpret 2.0 is an end-to-end simultaneous interpretation model for Chinese-to-English and English-to-Chinese speech, built on a full-duplex framework with real-time voice cloning so translations keep the speaker's voice.
- Modalities
- audio → audio, text
- At launch
- API
- Covers Chinese-to-English and English-to-Chinese interpretation
- Clones the speaker's voice in real time
- Offered on Volcano Engine as Doubao Simultaneous Interpretation 2.0
- Ola Friend earbuds announced as the first hardware to adopt it
Qwen3-Coder
Qwen3-Coder is Alibaba's agentic coding model family, led by Qwen3-Coder-480B-A35B-Instruct. Alibaba also open-sourced Qwen Code, a command-line tool for agentic coding in the terminal, alongside it.
- Modalities
- text → text
- Context
- 256K context
- Parameters
- MoE: 480B total, 35B active
- License
- Apache-2.0
- At launch
- API, Open weights
- Pre-trained on 7.5T tokens with a 70% code ratio
- Native 256K context, extendable to 1M with YaRN
- Qwen3-Coder-30B-A3B-Instruct followed on July 31, 2025
- Weights under the Apache 2.0 license
Qwen3-235B-A22B-Instruct-2507
Qwen3-235B-A22B-Instruct-2507 is an updated non-thinking version of Qwen3-235B-A22B with significant improvements and 256K-token long-context understanding. It began the Qwen3-2507 series of separate instruct and thinking models.
- Modalities
- text → text
- Context
- 256K context
- Parameters
- MoE: 235B total, 22B active
- License
- Apache-2.0
- At launch
- Open weights
- Native 262,144-token context; 1M-token support added August 8, 2025
- Qwen3-30B-A3B-Instruct-2507 followed on July 30 and Qwen3-4B-Instruct-2507 on August 6, 2025
- Weights under the Apache 2.0 license
Seed-X
Seed-X is an open-weights family of 7B multilingual translation models from ByteDance Seed, comprising instruction, reinforcement-learning, and reward models, supporting translation among 28 languages.
- Modalities
- text → text
- Parameters
- 7B
- License
- OpenMDW-1.0
- At launch
- Open weights
- Seed-X-Instruct, Seed-X-PPO, and Seed-X-RM released on Hugging Face
- Uses a Mistral-style architecture
- Quantized 8-bit and 4-bit PPO models followed on July 28, 2025
Act-Two
Next-generation motion capture model succeeding Act-One, with major gains in generation quality and support for head, face, body and hand tracking for character performances.
- Modalities
- video, image → video
- At launch
- Consumer app
- Head, face, body and hand tracking
- Available on paid plans
- Voice changing added to the Act-Two interface in August 2025
Canary-Qwen-2.5B
English speech recognition model built as a speech-augmented language model from a FastConformer Canary encoder and the Qwen3-1.7B language model.
- Modalities
- audio → text
- Parameters
- 2.5B
- License
- CC-BY-4.0
- At launch
- Open weights
- Combines canary-1b-flash with Qwen3-1.7B through LoRA
- Supports English speech-to-text
- Released under CC BY 4.0
LTX-Video 0.9.8
LTX-Video update adding long-shot generation of up to 60 seconds of video, with new 13B and 2B distilled models from a shared base and an IC-LoRA detailer model.
- Modalities
- text, image → video
- Parameters
- 13B and 2B
- License
- LTXV Open Weights License 0.X
- At launch
- Open weights
- Up to 60 seconds of video
- 13B and 2B distilled variants usable in one multiscale pipeline
- Compatible with the official IC-LoRA control models
OpenReasoning-Nemotron
Reasoning models derived from Qwen2.5 base models and post-trained for math, code and science solutions, released in four sizes under CC BY 4.0.
- Modalities
- text → text
- Parameters
- 1.5B, 7B, 14B and 32B
- License
- CC-BY-4.0
- At launch
- Open weights
- 32B model developed from Qwen2.5-32B
- Covers math, code and science reasoning
- Released under CC BY 4.0
CogVideoX-3
CogVideoX-3 is Z.ai's upgraded video generation model on its API, supporting text-to-video and image-to-video. It adds first-and-last-frame generation and improves clarity and the smoothness of large subject motion.
- Modalities
- text, image → video
- At launch
- API
- Adds start and end frame control
- Improved realistic and 3D-style scenes
- Available through the BigModel API platform
Voxtral Small and Voxtral Mini
Mistral's first audio models, open-weight speech understanding models that combine transcription with question answering and summarization, in a 24B production size and a 3B edge size.
- Modalities
- text, audio → text
- Context
- 32K context
- Parameters
- 24B and 3B
- License
- Apache-2.0
- At launch
- Open weights, API
- Handles audio up to 30 minutes for transcription or 40 minutes for understanding
- The API also offers Voxtral Mini Transcribe for transcription
- Released under Apache 2.0
Kimi K2
Kimi K2 is Moonshot AI's open-weights mixture-of-experts model optimized for agentic tasks, released as Kimi-K2-Base for fine-tuning and Kimi-K2-Instruct for chat and agent use. Moonshot trained it with the MuonClip optimizer.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 1T total, 32B active
- License
- Modified MIT License
- At launch
- API, Consumer app, Open weights
- 1T total parameters with 32B activated
- 128K context length at launch
- Offered on the Kimi Open Platform as kimi-k2-0711-preview
- Weights and code under the Modified MIT License
Devstral Medium
Larger agentic coding model released through the API alongside Devstral Small 1.1, built with All Hands AI for software engineering agents.
- Modalities
- text → text
- Context
- 125K context
- At launch
- API, Enterprise
- API model ID devstral-medium-2507
- Available for on-premises deployment and fine-tuning for enterprises
- Deprecated February 27, 2026
Devstral Small 1.1
Upgrade to Devstral Small that keeps the 24B architecture while improving generalization across prompts and agent scaffolds, built with All Hands AI.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 24B
- License
- Apache-2.0
- At launch
- Open weights, API
- Supports both Mistral function calling and XML formats
- Released under Apache 2.0
- API model ID devstral-small-2507
LFM2 (350M, 700M, 1.2B)
Second-generation Liquid Foundation Models using a new hybrid architecture of gated short convolutions and grouped query attention, released as open-weight 350M, 700M, and 1.2B dense checkpoints designed for fast on-device inference.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 350M: 354M; 700M: 742M; 1.2B: 1.17B
- License
- LFM Open License v1.0
- At launch
- API, Consumer app, Open weights
- Architecture: 10 double-gated short convolution blocks and 6 grouped query attention blocks
- Trained on 10T tokens with LFM1-7B as a knowledge-distillation teacher
- Weights on Hugging Face; also on Liquid Playground and OpenRouter
- LFM Open License v1.0 allows free commercial use below a company revenue threshold
Reka Flash 3.1
An update to Reka Flash 3 from an upgraded reinforcement learning stack, aimed at coding and as a base for agentic fine-tuning, released with a 3.5-bit Reka Quant version and the Reka Quant quantization library.
- Modalities
- text → text
- Parameters
- 21B
- License
- Apache-2.0
- At launch
- API, Consumer app, Open weights
- API model name reka-flash-3.1; also available in the Reka Playground
- 3.5-bit Q3_K_S GGUF build quantized with Reka Quant for llama.cpp
- A customized version orchestrates Reka Research; a multimodal version underpins Reka Vision
- Weights on Hugging Face under Apache-2.0
FlexOlmo
Language model training paradigm in which each data owner branches from a shared public model and trains an expert locally on its own data, letting owners opt in or out later.
- Modalities
- text → text
- License
- Apache-2.0
- At launch
- Open weights
- Enables collaboration without pooling sensitive data
- Released as FlexOlmo-7x7B-1T on Hugging Face
- Released under Apache 2.0
Grok 4 and Grok 4 Heavy
Grok 4 is xAI's flagship trained with reinforcement learning at pretraining scale on the 200,000-GPU Colossus cluster, with native tool use and real-time search. Grok 4 Heavy adds parallel test-time compute.
- Modalities
- text, image → text
- Context
- 250K context
- At launch
- Consumer app, API
- API offers a 256,000-token context window with text and vision understanding
- Available to SuperGrok and Premium+ subscribers and through the xAI API
- Grok 4 Heavy offered through a new SuperGrok Heavy tier
- API model grok-4-0709 retired May 15, 2026
MedGemma 27B Multimodal
MedGemma 27B Multimodal extends Google's open medical model collection with a 27B model that accepts images and text, adding longitudinal electronic health record interpretation. It was released together with the MedSigLIP medical image encoder.
- Modalities
- text, image → text
- Parameters
- 27B
- License
- Health AI Developer Foundations Terms of Use
- At launch
- Open weights, Cloud platforms
- Complements the earlier MedGemma 4B Multimodal and 27B text models
- MedSigLIP is a roughly 400M-parameter SigLIP-based encoder for classification and retrieval
- All MedGemma models run on a single GPU
Phi-4-mini-flash-reasoning
Phi-4-mini-flash-reasoning is a 3.8B open Microsoft model for math reasoning on edge and mobile devices, built on the new SambaY decoder-hybrid-decoder architecture that combines Mamba, sliding window attention and gated memory units.
- Modalities
- text → text
- Context
- 64K context
- Parameters
- 3.8B
- License
- MIT
- At launch
- Open weights, API, Cloud platforms
- Uses the SambaY architecture with a Gated Memory Unit
- Supports a 64K token context length
- Available on Azure AI Foundry, NVIDIA API Catalog and Hugging Face
T5Gemma
T5Gemma is a collection of encoder-decoder language models created by adapting pretrained Gemma 2 decoder-only models and further training them with UL2 or PrefixLM objectives. It includes balanced and unbalanced encoder-decoder pairings.
- Modalities
- text → text
- Parameters
- 2B-2B, 9B-9B, 9B-2B, plus T5-sized Small to XL
- License
- Gemma Terms of Use
- At launch
- Open weights, Cloud platforms
- Gemma 2-based configurations include 2B-2B, 9B-9B and an unbalanced 9B-2B
- Pretrained and instruction-tuned checkpoints
- Weights on Hugging Face and Kaggle; inference on Vertex AI
SmolLM3
Fully open 3B model with dual-mode reasoning, six languages and long context, released with an engineering blueprint covering its architecture and data mixtures.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 3B
- License
- Apache-2.0
- At launch
- Open weights
- Trained on 11 trillion tokens
- Trained at 64K context and extends to 128K with YaRN
- Supports English, French, Spanish, German, Italian and Portuguese
Jamba 1.7
Update to Jamba Large and Jamba Mini with better grounding for more complete and accurate answers, plus optimized configurations and quantization.
- Modalities
- text → text
- License
- Jamba Open Model License
- At launch
- Open weights, API, Enterprise
- Available for self-hosted deployment
- Weights on Hugging Face as Jamba Large 1.7 and Jamba Mini 1.7
- Suited to question answering and instruction following
GLM-4.1V-9B-Thinking
GLM-4.1V-9B-Thinking is Z.ai's open-weights vision-language reasoning model, built on GLM-4-9B-0414. It targets video understanding, image question answering, chart reading, and GUI operation with step-by-step reasoning.
- Modalities
- text, image, video → text
- Context
- 64K context
- Parameters
- 9B
- License
- MIT
- At launch
- API, Open weights
- Released with the GLM-4.1V-9B-Base model
- 64K context length support
- Weights under the MIT license; API launched July 2, 2025
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 July 2025." The AI Engineering, updated Jul 31, 2025. https://theaiengineering.com/releases/2025/07
HTML link
<a href="https://theaiengineering.com/releases/2025/07">AI models released in July 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