AI models released in December 2023
14 releases from 9 labs in December 2023, including Nous-Hermes-2-Yi-34B, Midjourney V6 and CogAgent. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 14
- Labs
- 9
- Open weights
- 7
December 2023
14 releasesNous-Hermes-2-Yi-34B
Hermes 2 model from Nous Research, a fine-tune of the Yi 34B base model, released openly on Hugging Face under Apache 2.0 in December 2023.
- Modalities
- text → text
- Parameters
- 34B
- License
- Apache-2.0
- At launch
- Open weights
- Listed by Nous Research as Hermes on Yi
- Built on the Yi 34B base model
- Published under Apache 2.0
Midjourney V6
Sixth-generation Midjourney image model with improved prompt accuracy for longer inputs, better coherence and knowledge, and advanced image prompting and remixing.
- Modalities
- text, image → image
- At launch
- Consumer app
- Default model from February 14 to July 30, 2024
- Supports Character Reference with --cref
- Released in December 2023, ahead of becoming the default in February 2024
CogAgent
CogAgent is Z.ai's open visual language model for GUI agents, improved from CogVLM. CogAgent-18B pairs 11B visual parameters with 7B language parameters and understands images at 1120 by 1120 resolution.
- Modalities
- text, image → text
- Parameters
- 18B
- License
- CogVLM License
- At launch
- Open weights
- Adds GUI agent abilities on top of CogVLM
- Supports high-resolution 1120 by 1120 image understanding
- Accepted as a CVPR 2024 Highlight
Imagen 2
Imagen 2 is Google's second-generation text-to-image model, launched on Vertex AI with improved photorealism, text rendering in images and logo generation. It also added long-form captioning and visual question answering on images.
- Modalities
- text → image
- At launch
- Cloud platforms
- Generally available on Vertex AI for allowlisted customers
- Prompts in six additional languages in preview
- Integrated with SynthID digital watermarking for allowlisted customers
Phi-2
Phi-2 is a 2.7B Microsoft Research base language model trained on 1.4T tokens of synthetic textbook-quality and filtered web data, scaled up from Phi-1.5 and offered as a platform for research.
- Modalities
- text → text
- Context
- 2K context
- Parameters
- 2.7B
- License
- MIT
- At launch
- Open weights, Cloud platforms
- Base model without RLHF or instruction tuning
- Trained for 14 days on 96 A100 GPUs
- Released in the Azure AI Studio model catalog
Mistral 7B v0.2
Minor release of Mistral 7B whose instruct version served the mistral-tiny endpoint at the beta launch of la Plateforme, with the context window extended to 32K tokens.
- Modalities
- text → text
- Context
- 32K context
- License
- Apache-2.0
- At launch
- Open weights, API
- Context window extended from 8K to 32K tokens
- Mistral 7B Instruct v0.2 served the mistral-tiny endpoint, later renamed open-mistral-7b
- Released under Apache 2.0
Mistral Embed
General-purpose text embedding model launched as the mistral-embed endpoint on la Plateforme, producing 1024-dimensional vectors designed with retrieval in mind.
- Modalities
- text → embedding
- Context
- 8K context
- At launch
- API
- Outputs 1024-dimensional embeddings
- Launched alongside the first la Plateforme chat endpoints
- API model ID mistral-embed-2312
Mistral Medium
Mistral AI's first proprietary model, a prototype served as the highest-quality mistral-medium endpoint when la Plateforme opened in beta with three chat endpoints and an embedding endpoint.
- Modalities
- text → text
- Context
- 32K context
- At launch
- API
- Launched in beta with the first la Plateforme endpoints
- Covers English, French, Italian, German, Spanish and code
- Later dated as mistral-medium-2312
Mixtral 8x7B
Sparse mixture-of-experts model in which a router picks two of eight expert groups per token, released with open weights under Apache 2.0 alongside an instruct version.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- MoE: 46.7B total, 12.9B active
- License
- Apache-2.0
- At launch
- Open weights, API
- Handles English, French, Italian, German and Spanish
- Served the original mistral-small endpoint on la Plateforme
- Uses 12.9B of its 46.7B parameters per token
Llama Guard
Llama Guard is a 7B Llama 2 based safety classifier for filtering LLM inputs and outputs, released as part of Meta's Purple Llama trust and safety project.
- Modalities
- text → text
- Parameters
- 7B
- License
- Llama 2 Community License
- At launch
- Open weights
- Instruction-tuned from Llama 2 7B
- Covers six harm categories in its original taxonomy
- Launched with CyberSec Eval cybersecurity benchmarks under Purple Llama
StableLM Zephyr 3B
3B-parameter chat model extending Stable LM 3B-4e1t, preference-tuned with a Zephyr-style pipeline for instruction following and question answering on edge devices.
- Modalities
- text → text
- Parameters
- 3B
- License
- Stability AI Community License
- At launch
- Open weights
- Supervised fine-tuning followed by preference tuning
- Launched under a non-commercial license
- Built on Stable LM 3B-4e1t
Gemini 1.0 Nano
Gemini 1.0 Nano was the most efficient model in Google's first Gemini generation, designed for on-device tasks. Pixel 8 Pro was the first smartphone engineered to run it.
- Modalities
- text → text
- At launch
- Consumer app
- Powered Summarize in the Recorder app and Smart Reply in Gboard on Pixel 8 Pro
- Offered to Android developers through AICore in Android 14, starting with an early preview on Pixel 8 Pro
- One of three Gemini 1.0 sizes alongside Ultra and Pro
Gemini 1.0 Pro
Gemini 1.0 Pro was the mid-sized model in Google's first Gemini generation, built for scaling across a wide range of tasks. A fine-tuned version began powering Bard on Dec 6, 2023, and developers gained API access on Dec 13, 2023.
- Modalities
- text, image → text
- Context
- 32K context
- At launch
- API, Cloud platforms, Consumer app
- Bard switched to a fine-tuned Gemini Pro on Dec 6, 2023 in English in more than 170 countries and territories
- Gemini API and Vertex AI access began Dec 13, 2023 with a 32K context window for text
- A separate Gemini Pro Vision endpoint accepted text and images with text output
- Supported 38 languages at API launch
Skylark2
Skylark2 was the second generation of ByteDance's Skylark language models on Volcano Engine's Ark platform, launched as Skylark2-pro-4k and followed by 32K, lite, role-play, and turbo variants before Doubao models replaced it.
- Modalities
- text → text
- At launch
- API
- Skylark2-pro-4k V1.0 released December 5, 2023
- Skylark2-pro-32k V1.0 followed on December 25, 2023
- Retired on August 20, 2024 in favor of Doubao models
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 December 2023." The AI Engineering, updated Dec 25, 2023. https://theaiengineering.com/releases/2023/12
HTML link
<a href="https://theaiengineering.com/releases/2023/12">AI models released in December 2023</a> (The AI Engineering)- Full dataset
- 1,167 releases from 55 labs
- Coverage
- to
- Updated
- License
- CC BY 4.0
- Publisher
- The AI Engineering