Open-Weight on Best of AI

ALBERT

Thu, 16 Jul 2026 00:00:00 +0000

ALBERT (A Lite BERT) is a language representation model that Google Research and Toyota Technological Institute published in September 2019 as a follow-up to BERT. It uses two parameter-reduction tricks, factorized embedding parameterization and cross-layer parameter sharing, to cut the parameter count sharply while keeping accuracy competitive with much larger models. The largest configuration, ALBERT-xxlarge at around 223 million parameters, has roughly 18 times fewer parameters than BERT-large yet trains faster and set new state-of-the-art scores on GLUE, SQuAD 2.0, and RACE at the time of release.

Aquila

Thu, 16 Jul 2026 00:00:00 +0000

Aquila is the Beijing Academy of Artificial Intelligence’s first open bilingual language model, released in 2023 as a 7-billion-parameter base model trained on Chinese and English text. BAAI built it on a GPT-style architecture with a custom tokenizer designed to handle Chinese text more efficiently than tokenizers borrowed from English-first models, and it supports a 2,048-token context window. The release included both a base pretrained model and an AquilaChat instruction-tuned variant, along with a code-focused AquilaCode model.

Aquila2

Thu, 16 Jul 2026 00:00:00 +0000

Aquila2 is the Beijing Academy of Artificial Intelligence’s second-generation bilingual model family, released in October 2023 as a follow-up to the original Aquila. The family spans 7B and 34B base models, chat-tuned AquilaChat2 variants, and long-context versions extended to 16,000 tokens through positional-encoding interpolation, plus experimental 70B checkpoints. BAAI reported that Aquila2-34B outperformed other open Chinese-English models of similar size on Chinese-language benchmarks at the time of release.

The models are released under the Apache 2.0 license for the codebase, with model weights governed by BAAI’s own Aquila license agreement that allows commercial use under certain conditions. Aquila2 targeted developers who needed strong bilingual performance without relying on a Western base model like LLaMA, and it remained one of BAAI’s main open releases until the organization shifted more attention toward its FlagEmbedding and other specialized model lines.

Arctic

Thu, 16 Jul 2026 00:00:00 +0000

Arctic is a large language model that Snowflake released in April 2024, built for enterprise work like SQL generation, coding, and instruction following rather than general chat. It uses a mixture-of-experts design with 480 billion total parameters spread across 128 experts, but only about 17 billion are active for any given token, which keeps inference costs down relative to dense models of similar scale. Snowflake released both the base and instruct checkpoints under an Apache 2.0 license, so the weights can be self-hosted or run through Snowflake Cortex. On the Spider text-to-SQL benchmark it scored around 79% accuracy, and Snowflake pitched it as a way to bring capable open models into data warehouses without sending queries to a third-party API. It trades off a shorter 4K context window and weaker general chat ability against strong performance on the narrow enterprise tasks it was built for.

Arctic Embed 2.0

Thu, 16 Jul 2026 00:00:00 +0000

AudioCraft

Thu, 16 Jul 2026 00:00:00 +0000

AudioCraft is a code library that Meta released in August 2023, bundling three generative audio models: MusicGen for text-to-music generation, AudioGen for environmental sound effects, and EnCodec, the neural audio codec both rely on for tokenizing waveforms. MusicGen comes in several sizes (300M, 1.5B, and 3.3B parameters), with variants that can also follow a melody alongside a text prompt, while AudioGen ships as a 1.5B-parameter model trained on sound effects rather than music. Meta released the code and weights under the MIT license, which made AudioCraft one of the first fully open toolkits for text-to-audio generation, in contrast to closed systems from companies like Google. It runs on a single consumer GPU for the smaller checkpoints, and it’s become a common base for open-source music generation projects and research on controllable audio synthesis.

Aya 23

Thu, 16 Jul 2026 00:00:00 +0000

Aya 23 is a family of multilingual language models that Cohere For AI, Cohere’s research lab, released in May 2024, covering 23 languages including Arabic, Chinese, Hindi, Japanese, Russian, and Spanish alongside English. It shipped in two sizes, 8B and 35B parameters, both built on Cohere’s Command architecture and initialized from a pretrained checkpoint before further multilingual training. The 35B version is the one tracked here, with an 8192-token context window aimed at retrieval and instruction-following tasks across languages that are usually underserved by English-centric open models. Cohere released the weights under a CC-BY-NC 4.0 license with an acceptable-use addendum, so it’s free for research and non-commercial use but requires a separate commercial agreement for production deployment. It followed the original Aya project, which had focused on data and evaluation for 101 languages, by narrowing scope to a smaller set of languages in exchange for stronger per-language quality.

Baichuan 2

Thu, 16 Jul 2026 00:00:00 +0000

Baichuan 2 is the second generation of open-weight language models from Baichuan Intelligent Technology, released in September 2023. It came in 7B and 13B parameter sizes, each with base and chat variants, and was trained on 2.6 trillion bilingual Chinese-English tokens, roughly double what the original Baichuan models saw. Baichuan reported that the 7B model gained close to 30% on MMLU over its Baichuan 1 predecessor of the same size, and the family did well on Chinese-language benchmarks like C-Eval and CMMLU relative to other open models available at the time. The weights are distributed under Apache 2.0 for the code plus a separate community license for the model itself, which allows free commercial use as long as the deploying company has under 1 million daily active users and isn’t itself a cloud or software service reselling the model. It has since been superseded by Baichuan 3 and Baichuan 4, which moved to closed, API-only access.

Baichuan-13B

Thu, 16 Jul 2026 00:00:00 +0000

Baichuan-13B is a bilingual Chinese-English language model that Baichuan Intelligent Technology released in July 2023, expanding on the company’s earlier 7B model. It has 13 billion parameters and was trained on about 1.4 trillion tokens, which the company said was 40% more training data than Meta’s LLaMA-13B had used at the time. It came in both a base checkpoint for further fine-tuning and a Baichuan-13B-Chat variant aligned for dialogue, both distributed on Hugging Face and GitHub under a community license that permits commercial use for smaller companies. It predates the Baichuan 2, 3, and 4 model families and is largely superseded by them, but it was one of the earlier open-weight Chinese LLMs to reach 13B scale with a permissive-enough license for commercial experimentation.

BERT

Thu, 16 Jul 2026 00:00:00 +0000

BERT (Bidirectional Encoder Representations from Transformers) is a language model that Google Research published in October 2018, and it changed how the field approached pretraining for NLP tasks. Unlike earlier models that read text left to right, BERT trains on masked-word prediction and next-sentence prediction, letting it build representations that draw on context from both directions at once. Google released two main sizes: BERT-base at 110 million parameters and BERT-large at 340 million, both with a 512-token limit, and made the weights freely available under Apache 2.0 on GitHub and later Hugging Face. Fine-tuned versions of BERT set new records on eleven different NLP benchmarks at launch, including question answering and natural language inference, and it quickly became the standard base model to fine-tune for tasks like classification, named entity recognition, and search ranking. It’s mostly been superseded by newer encoder models like RoBERTa and DeBERTa and by the embedding models that followed it, but it’s still one of the most cited papers in NLP and its architecture underlies a large share of later encoder-based systems.

BGE-large-en

Thu, 16 Jul 2026 00:00:00 +0000

BGE-large-en-v1.5 is a text embedding model from BAAI, the Beijing Academy of Artificial Intelligence, released in September 2023 as an update to the original BGE-large-en model from a few weeks earlier. It has 335 million parameters, produces 1024-dimension embeddings, and handles sequences up to 512 tokens, putting it in the same size class as other BERT-scale encoder models rather than the larger decoder-based embedding models that came later. The v1.5 update mainly fixed a similarity-score distribution issue in the original release and improved retrieval quality when used without a task-specific instruction prefix, which had been a rough edge in v1. BAAI released it under the MIT license, and it became one of the most widely used open embedding models for English retrieval and RAG pipelines, largely because it was free, fast to run, and scored well on the MTEB benchmark relative to its size. It’s since been joined by larger and more capable BAAI models like BGE-M3, but it remains a common default for lightweight self-hosted retrieval.

BGE-M3

Thu, 16 Jul 2026 00:00:00 +0000

BGE-M3 is a text embedding model from BAAI, described in a paper published in February 2024, and the name refers to three things it does at once: multi-linguality across more than 100 languages, multi-granularity handling of inputs up to 8192 tokens, and multi-functionality in that it produces dense vectors, sparse lexical weights, and ColBERT-style token vectors from a single forward pass. It has 568 million parameters, which is modest next to large LLMs but is trained specifically for retrieval rather than generation. That combination let it post state-of-the-art results at the time on the multilingual MIRACL benchmark and the cross-lingual MKQA benchmark, since a single model could substitute for what previously required separate dense and sparse retrieval systems. BAAI released it under the MIT license on Hugging Face, and it’s widely used in retrieval-augmented generation pipelines that need to search across documents in multiple languages or documents longer than the 512-token limit typical of earlier BERT-based embedding models.

BioGPT

Thu, 16 Jul 2026 00:00:00 +0000

BioGPT is a language model that Microsoft Research introduced in October 2022, built specifically for biomedical text rather than adapted from a general-purpose model after the fact. It uses a GPT-2-style decoder architecture trained from scratch on roughly 15 million PubMed abstracts, and it comes in a 347-million-parameter base version and a 1.5-billion-parameter BioGPT-Large version. Because it was trained on biomedical text from the start rather than fine-tuned onto it, it outperformed earlier general-purpose language models on tasks like biomedical relation extraction, question answering over medical literature, and document classification. Microsoft released both checkpoints under the MIT license on Hugging Face and GitHub, making it a common starting point for research groups building biomedical NLP tools without needing to train a domain-specific model from scratch. It predates the wave of general-purpose instruction-tuned chat models and hasn’t been updated since, so it’s mostly used now as a lightweight, self-hostable option for narrow biomedical text tasks rather than as an general assistant.

BitNet b1.58

Thu, 16 Jul 2026 00:00:00 +0000

BitNet b1.58 is a line of research from Microsoft exploring language models built from ternary weights, meaning every weight is restricted to -1, 0, or 1 instead of the usual 16- or 32-bit floating point values. The original paper, “The Era of 1-bit LLMs,” came out in February 2024 and showed that models trained natively this way, rather than quantized after the fact, could match full-precision models like Llama 2 at comparable sizes while cutting memory use and inference energy substantially. Microsoft followed up in April 2025 with BitNet b1.58 2B4T, an openly released 2-billion-parameter checkpoint trained on 4 trillion tokens, whose non-embedding memory footprint runs about 0.4GB, well under similarly sized models like Gemma 3 1B or MiniCPM 2B. The weights and inference code are on GitHub and Hugging Face under the MIT license, and the project is notable less for beating other models on benchmarks than for proving that 1-bit training works at all without a big performance penalty, which matters for running capable models on phones or edge devices.

BLIP

Thu, 16 Jul 2026 00:00:00 +0000

BLIP (Bootstrapping Language-Image Pre-training) is a vision-language model that Salesforce Research introduced in early 2022. It handles both understanding tasks, like image captioning and visual question answering, and retrieval tasks, like matching images to text. The model was trained with a bootstrapping method that cleans up noisy web-scraped captions by generating synthetic ones and filtering out the bad matches, which let it reach strong results without relying purely on massive uncurated datasets. At around 223 million parameters, BLIP is small next to later multimodal systems, but it was influential in showing that caption quality matters as much as caption quantity for pretraining.

BLIP-2

Thu, 16 Jul 2026 00:00:00 +0000

BLIP-2 is a vision-language model that Salesforce Research published in January 2023, and it’s built around a specific efficiency trick: instead of training a large image encoder and language model together from scratch, it freezes both a pretrained image encoder and a pretrained language model and trains only a lightweight connector between them, called a Q-Former. That Q-Former has around 188 million trainable parameters, less than 2% of the total model, which made BLIP-2 far cheaper to train than comparable vision-language models of the time while still doing well on image captioning and visual question answering. Salesforce released several variants pairing the Q-Former with different frozen language models, including versions built on OPT (2.7B and 6.7B parameters) and Flan-T5 (up to 11B parameters), so users could trade off capability against compute cost. The code and weights are on GitHub and Hugging Face under the MIT license as part of Salesforce’s LAVIS library, and the frozen-encoder, lightweight-connector approach it popularized influenced later multimodal models that similarly bolt vision capabilities onto existing language models rather than training everything jointly.

BLOOM

Thu, 16 Jul 2026 00:00:00 +0000

BLOOM is a 176 billion parameter language model built by BigScience, a year-long collaboration of over a thousand researchers coordinated through Hugging Face with compute donated by the French government’s Jean Zay supercomputer. It launched in July 2022 as an answer to closed models like GPT-3, trained on the ROOTS corpus covering 46 natural languages and 13 programming languages, with a deliberate emphasis on underrepresented languages that most large labs ignored at the time. The project published its training data, code, and a full technical report alongside the weights, which was rare for a model at that scale.

BLOOMZ

Thu, 16 Jul 2026 00:00:00 +0000

BLOOMZ is BigScience’s instruction-tuned version of BLOOM, released in late 2022 by fine-tuning the base 176 billion parameter model on xP3, a large collection of multilingual tasks phrased as natural-language instructions. The point of the project was cross-lingual generalization: BLOOMZ was tuned mostly on English instructions but showed it could follow instructions in languages it had never seen paired with instructions during fine-tuning, as long as it had encountered that language during BLOOM’s original pretraining. That made it one of the earlier public demonstrations that instruction-following skills transfer across languages rather than needing separate tuning for each one.

BTLM-3B

Thu, 16 Jul 2026 00:00:00 +0000

BTLM-3B-8K is a 3 billion parameter language model that Cerebras built with the Opentensor Foundation and trained on Cerebras’ wafer-scale CS-2 systems. Announced in the summer of 2023, it was designed for a specific goal: match the quality of 7 billion parameter models while staying small enough to run on phones and low-memory devices. Cerebras reports that BTLM-3B, once quantized, fits in around 3GB of memory and runs on hardware like the iPhone, Google Pixel, and even a Raspberry Pi, while performing in line with contemporary 7B models such as RedPajama-INCITE-7B on common benchmarks.

ByT5

Thu, 16 Jul 2026 00:00:00 +0000

ByT5 is a variant of Google’s T5 text-to-text model that Google Research published in mid-2021, notable for dropping tokenization altogether and operating directly on raw UTF-8 bytes. Standard language models split text into subword tokens using a fixed vocabulary built by a tokenizer, which works well for the languages and scripts that vocabulary was built around but tends to handle rare words, misspellings, and less common languages poorly. ByT5 sidesteps that by treating every input as a sequence of bytes, giving it a vocabulary of only a few hundred possible values instead of tens of thousands of subword tokens.

Canary

Thu, 16 Jul 2026 00:00:00 +0000

Canary is a family of speech models from NVIDIA’s NeMo team, with the flagship Canary-1B released in April 2024. Unlike a typical ASR model that only transcribes audio, Canary handles both speech recognition and speech translation across English, French, German, and Spanish, converting spoken audio into text in the same language or translating it into another. It combines a FastConformer encoder with a transformer decoder, an architecture NVIDIA tuned to get strong accuracy from a comparatively small model, and it topped the Hugging Face Open ASR Leaderboard against competitors several times its size when it launched.

Cerebras-GPT

Thu, 16 Jul 2026 00:00:00 +0000

Cerebras-GPT is a family of language models, ranging from 111 million to 13 billion parameters, that Cerebras trained on its own CS-2 wafer-scale systems and released in March 2023. The point of the project was less about beating benchmarks and more about openness: Cerebras published the weights, training code, and a detailed technical report describing hyperparameters and scaling behavior for every size in the family, following compute-optimal training recipes similar to those behind Chinchilla. That made Cerebras-GPT one of the more fully documented open model families of its time, useful to researchers who wanted a clean, reproducible baseline for studying scaling laws.

Chameleon

Thu, 16 Jul 2026 00:00:00 +0000

Chameleon is a multimodal model from Meta FAIR, described in a May 2024 paper, that mixes text and images from the ground up rather than bolting a vision encoder onto a language model. Most multimodal systems use “late fusion,” where a separate image encoder feeds visual features into a language model that was pretrained on text alone. Chameleon instead quantizes images into discrete tokens and trains on sequences of interleaved text and image tokens from scratch, so the same transformer learns to reason over both modalities and can also generate images as output, not just describe them.

ChatGLM3

Thu, 16 Jul 2026 00:00:00 +0000

ChatGLM3 is the third generation of the ChatGLM dialogue model line, built jointly by Zhipu AI and Tsinghua University’s KEG lab and released in late October 2023. The base ChatGLM3-6B model uses a bilingual Chinese-English design and added stronger support for function calling, code execution, and agent-style tool use compared to its predecessors, along with a more diverse pretraining mix covering more training tokens and better alignment training. At release, Zhipu reported that the 6B base model led other pretrained models under 10 billion parameters across dozens of benchmarks spanning reasoning, math, code, and general knowledge.

CLIP

Thu, 16 Jul 2026 00:00:00 +0000

CLIP (Contrastive Language-Image Pre-training) is a model OpenAI released in January 2021 that learns to match images with text descriptions. It was trained on 400 million image-text pairs scraped from the internet, using a contrastive objective that pulls matching image and text embeddings together while pushing mismatched pairs apart. The largest released variant, built on a ViT-L/14 vision transformer paired with a 63M-parameter text transformer, totals roughly 428 million parameters. Instead of being trained for one narrow task, CLIP learns a general-purpose joint embedding space for images and text, which lets it perform zero-shot image classification by comparing an image against a set of text labels with no task-specific fine-tuning.

CLIP (LAION variant)

Thu, 16 Jul 2026 00:00:00 +0000

This CLIP variant is an open-weight retraining of OpenAI’s original CLIP architecture, produced by LAION and distributed through the open_clip project rather than by OpenAI itself. It uses the ViT-L/14 vision transformer backbone, has about 428 million parameters, and was trained on LAION-2B, the English-language 2 billion image-text pair subset of the larger LAION-5B dataset, using 384 A100 GPUs on the JUWELS supercomputer. Like the original CLIP, it maps images and text into a shared embedding space so you can do zero-shot image classification, image-text retrieval, and similarity search without any task-specific fine-tuning, and it reaches about 75.3 percent zero-shot accuracy on ImageNet-1k. Because the weights are released under an open license and hosted on Hugging Face, it has become a common drop-in replacement for OpenAI’s closed CLIP checkpoints in open-source image search, captioning, and generative art pipelines, including as a text encoder inside several Stable Diffusion variants.

Code Llama

Thu, 16 Jul 2026 00:00:00 +0000

Code Llama is Meta’s code-focused version of Llama 2, released in August 2023. It was trained further on code-heavy data on top of the base Llama 2 weights and shipped in three foundation sizes (7B, 13B, and 34B parameters), each with a base version, a Python-specialized version, and an instruction-tuned version for chat-style coding help. Meta later added a 70B variant. A key change from the base Llama 2 model was extending the context window from 4,096 tokens up to 100,000 tokens, achieved by adjusting the RoPE positional embeddings, which made the model far more useful for reading and editing long files or entire codebases in one pass.

CodeGen2

Thu, 16 Jul 2026 00:00:00 +0000

CodeGen2 is Salesforce Research’s second-generation family of open code models, released in May 2023 as a follow-up to the original CodeGen. It comes in four sizes: 1B, 3.7B, 7B, and 16B parameters. The main improvement over the first CodeGen is infilling support, meaning the model can fill in a missing chunk of code given the surrounding context on both sides, rather than only generating text left to right. This makes it better suited to real editor workflows like autocomplete and code repair, where the model needs to write code in the middle of an existing file.

Codestral

Thu, 16 Jul 2026 00:00:00 +0000

Codestral is Mistral AI’s dedicated code generation model, first released in May 2024 with 22 billion parameters. It’s built for fill-in-the-middle completion and multi-language code generation, and it launched with an 81.1 percent pass@1 score on HumanEval, ahead of larger code models like Code Llama 70B and DeepSeek Coder 33B at the time. The original release used a 32,000-token context window and shipped under the Mistral AI Non-Production License, which limited free use to research and testing and required a separate commercial license for production deployment.

Codestral Mamba

Thu, 16 Jul 2026 00:00:00 +0000

Codestral Mamba is a code generation model Mistral AI released in July 2024, notable mainly for its architecture: instead of the transformer design used by nearly every other code model, it’s built on Mamba, a state-space model architecture. State-space models process sequences with linear rather than quadratic scaling in sequence length, which means inference speed does not degrade as sharply as context grows. Mistral sized it at roughly 7 billion parameters and gave it a 256,000-token context window, with the company reporting reliable in-context retrieval performance even at that length.

CodeT5+

Thu, 16 Jul 2026 00:00:00 +0000

CodeT5+ is a family of open-weight code models built by Salesforce Research, released in May 2023 as a follow-up to the original CodeT5 encoder-decoder model. It comes in several sizes from 220 million up to 16 billion parameters and can run in encoder-only, decoder-only, or full encoder-decoder mode, which lets one architecture cover code understanding tasks like defect detection alongside generation tasks like code completion and text-to-code synthesis. Training combines span denoising, causal language modeling, contrastive learning, and text-code matching objectives, an approach Salesforce says gives it an edge over decoder-only code models of similar size on retrieval-augmented generation tasks. The 16B variant scores around 30.9% pass@1 on HumanEval, and an instruction-tuned version of the same checkpoint pushes past 35%, beating OpenAI’s older code-cushman-001 model on the same benchmark. Weights are hosted on Hugging Face under a BSD-3-Clause license, making CodeT5+ one of the earlier fully open alternatives to Codex-era proprietary code models.

CogVideoX

Thu, 16 Jul 2026 00:00:00 +0000

CogVideoX is Zhipu AI’s open-weight text-to-video model, built with Tsinghua University’s KEG lab and released in August 2024 as a successor to the original CogVideo project. It ships in 2B and 5B parameter variants using a 3D causal VAE for spatial and temporal compression alongside an expert adaptive LayerNorm transformer, and it supports text-to-video, image-to-video, and video continuation from a single frame or prompt. The 5B model generates clips up to 10 seconds long at up to 768p resolution and 16 frames per second, and Zhipu later shipped a CogVideoX-1.5 update in November 2024 with longer, higher-resolution output. Weights are released under Apache 2.0 and hosted on Hugging Face, and Zhipu also serves the models through its own API, so CogVideoX works both as a self-hosted open model and as a hosted service, similar to how Stability AI and Runway position their video generators.

CogView3

Thu, 16 Jul 2026 00:00:00 +0000

CogView3 is a text-to-image model from Zhipu AI and Tsinghua University, described in a March 2024 research paper and open-sourced on Hugging Face in September 2024. It is built around relay diffusion, a cascaded approach where a 3-billion-parameter base stage generates a low-resolution 512x512 image and a second stage adds noise back into that image before running a shorter super-resolution diffusion pass on top of it, rather than conditioning the second stage directly on the low-res output like older cascaded models did. Zhipu reports that CogView3 beats Stable Diffusion XL in human evaluations by a wide margin while needing only about half the inference time, and a distilled version matches that quality using roughly a tenth of SDXL’s inference steps. The model and code are released under Apache 2.0, and Zhipu followed it with CogView-3Plus, a larger diffusion transformer variant in the same family, plus a hosted version through the Zhipu API for developers who don’t want to run it themselves.

Command R Plus

Thu, 16 Jul 2026 00:00:00 +0000

Command R Plus is Cohere’s flagship model from the Command R generation, released in April 2024 with a 128,000 token context window and 104 billion parameters. Cohere built it for retrieval-augmented generation, multi-step tool use, and multilingual tasks, and launched it in partnership with Microsoft Azure as one of the first models in that generation to pair strong RAG performance with agentic tool calling. The model scores 75.7 on MMLU. Cohere published the weights on Hugging Face under a CC-BY-NC 4.0 license with an acceptable use addendum, so researchers can download and run it for non-commercial purposes for free, while commercial use requires a separate license or access through Cohere’s hosted API at $2.50 per million input tokens and $10 per million output tokens. It sits above the smaller Command R model in Cohere’s lineup and was superseded by Command A in 2025.

Command R+

Thu, 16 Jul 2026 00:00:00 +0000

Command R+ is Cohere’s flagship large language model, built for enterprise use cases centered on retrieval-augmented generation, tool use, and multi-step agent workflows. It launched in April 2024 with 104 billion parameters and a 128,000-token context window, then received a refreshed version in August 2024 that cut latency and raised throughput on the same hardware footprint. The model supports 10 languages for business use and ships with citation generation built in, so answers grounded in retrieved documents come with source references attached. Cohere positions it against GPT-4-class models for RAG-heavy deployments rather than as a general chatbot, and it is available both through Cohere’s own API and as open weights on Hugging Face under a non-commercial license.

Command R7B

Thu, 16 Jul 2026 00:00:00 +0000

Command R7B is the smallest model in Cohere’s R series, released in December 2024 as what the company called the final entry in that lineup. At 7 billion parameters it is built to run on cheap hardware, including CPUs and single GPUs, while still handling retrieval-augmented generation, tool use, and reasoning tasks that Cohere’s larger models were designed for. It topped the Hugging Face Open LLM Leaderboard among similarly sized open-weight models at launch, beating Gemma 2 9B, Ministral 8B, and Llama 3.1 8B on average across benchmarks including MMLU, IFEval, and GPQA. The model keeps the same 128,000-token context window as its bigger siblings and is priced far below them, making it a fit for high-volume, latency-sensitive applications rather than frontier-level tasks.

CosyVoice

Thu, 16 Jul 2026 00:00:00 +0000

CosyVoice is a text-to-speech and voice-cloning model from Alibaba’s FunAudioLLM team, released as open weights in July 2024. It generates speech from short reference audio clips, so it can clone a voice from just a few seconds of sample and then speak arbitrary text in that voice across multiple languages, including Chinese, English, Japanese, and Korean. The model builds on supervised semantic speech tokens paired with a language-model-style generation approach, which lets it handle cross-lingual and zero-shot voice cloning without speaker-specific fine-tuning. Alibaba has continued to update the line, later releasing CosyVoice 2 and CosyVoice 3 with larger underlying models and lower latency, and the original release remains available on Hugging Face and through Alibaba Cloud’s Model Studio API under an Apache 2.0 license.

DBRX

Thu, 16 Jul 2026 00:00:00 +0000

DBRX is Databricks’ open-weight large language model, released in March 2024 as the base model behind the instruction-tuned DBRX Instruct. It is a fine-grained mixture-of-experts model with 132 billion total parameters and 36 billion active parameters per token, trained from scratch on 12 trillion tokens of text and code using Databricks’ own Mosaic AI training stack. At release, Databricks reported it outperforming other open models of the time, including Llama 2 70B and Mixtral, on benchmarks such as MMLU and HumanEval, and it was pitched as evidence that an enterprise-focused company could train a competitive foundation model without relying on a big-lab research budget. It has a 32,000-token context window and is released under the Databricks Open Model License, with weights available on Hugging Face for self-hosting or through Databricks’ own serving infrastructure.

DBRX Instruct

Thu, 16 Jul 2026 00:00:00 +0000

DBRX Instruct is the fine-tuned, chat-ready version of Databricks’ DBRX model, released in March 2024 alongside the DBRX base model. It uses a fine-grained mixture-of-experts architecture with 132 billion total parameters, of which only 36 billion are active for any given token, and it was trained on 12 trillion tokens of text and code. Databricks tuned it specifically for instruction following and multi-turn conversation, and at launch the company reported it beating GPT-3.5 on MMLU (73.7% versus 70.0%) and on HumanEval (70.1% versus 48.1%). It supports a 32,000-token context window and is distributed under the Databricks Open Model License, which permits commercial use with some restrictions for very large deployments. Organizations run it through Databricks’ own platform or self-host the weights from Hugging Face.

DeepFloyd IF

Thu, 16 Jul 2026 00:00:00 +0000

DeepFloyd IF is a text-to-image model released in April 2023 by DeepFloyd, a research lab backed by Stability AI. It takes a different approach from Stable Diffusion’s latent-space diffusion: it operates directly in pixel space across a cascade of three modules, first generating a low-resolution 64x64 image and then upsampling it in stages to 1024x1024, which gives it a strong reputation for rendering legible text inside images, a task that latent diffusion models of that era often botched. DeepFloyd trained three sizes of the base module, the largest with 4.3 billion parameters, and the full pipeline was released as a non-commercial research preview under a Stability Community License rather than a fully open license. It never became a mainstream product the way Stable Diffusion did, but its text-rendering results influenced later commercial models that improved on the same weakness.

DeepSeek Coder V2

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek-Coder-V2 is DeepSeek’s second-generation code model, released in June 2024 as a mixture-of-experts model with 236 billion total parameters and 21 billion active per token. It extended the context window to 128,000 tokens, up sharply from the 16,000-token window of the original DeepSeek-Coder, and expanded language coverage from 86 to 338 programming languages. On HumanEval it scored 90.2%, a result DeepSeek billed at the time as closing the gap with closed models like GPT-4 Turbo on coding tasks, and it also improved on general reasoning and math benchmarks compared to its predecessor. Unlike the original DeepSeek-Coder’s more restrictive license, DeepSeek-Coder-V2 is released under the MIT license, making it freely usable for commercial projects, and weights are available on Hugging Face alongside API access through DeepSeek’s own platform.

DeepSeek LLM 67B

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek LLM 67B is DeepSeek’s first dense foundation model, released in November 2023 before the company pivoted toward the mixture-of-experts architectures used in its later V2 and V3 lines. It is a 67-billion-parameter dense transformer trained on 2 trillion tokens of English and Chinese text, and it scored around 71.9 on MMLU, putting it roughly in line with Llama 2 70B on general knowledge benchmarks while beating it on several Chinese-language tasks. The model has a comparatively short 4,096-token context window by later standards, reflecting the norms of late 2023 training runs. It was released with both base and chat variants under a permissive DeepSeek license allowing commercial use, and it served as the foundation DeepSeek built on for its subsequent, more widely used models.

DeepSeek R1

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek V3

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek V4-Pro

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek-Coder-33B

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek-Coder-33B is a code-focused language model that DeepSeek released in November 2023, before the company became widely known for its later reasoning models. It was trained from scratch on 2 trillion tokens split mostly between code and natural language, with a 16,000-token context window and project-level training that lets it reason across multiple files rather than isolated snippets. The instruction-tuned version scored 79.3% pass@1 on HumanEval at release, putting it ahead of most open code models of the time and within striking distance of GPT-3.5 on coding benchmarks. It supports fill-in-the-middle completion for use in code editors and covers a wide range of programming languages. DeepSeek has since superseded it with DeepSeek-Coder-V2, a larger mixture-of-experts model with a longer context window, but the 33B model is still available on Hugging Face for self-hosting.

DeepSeek-Math

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeekMath is a 7-billion-parameter model DeepSeek built specifically for mathematical reasoning, released in February 2024 and initialized from DeepSeek-Coder-Base. It was trained on 120 billion math-related tokens pulled from web data, alongside natural language and code, and scored 51.7% on the competition-level MATH benchmark without relying on external calculators or tools, a result that at the time approached the performance of much larger closed models like Gemini-Ultra and GPT-4. DeepSeek released it in base, instruction-tuned, and reinforcement-learning-tuned variants, with the RL version trained using Group Relative Policy Optimization, an algorithm the company introduced in this paper and later reused for DeepSeek-R1. The model has a 4,096-token context window and is aimed at researchers and developers working on math tutoring, theorem proving, and quantitative reasoning tasks rather than general-purpose chat.

DeepSeek-MoE-16B

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek-MoE-16B, released in January 2024, was DeepSeek’s early attempt at a fine-grained mixture-of-experts architecture, and its design choices carried forward into the company’s later V2 and V3 models. It has 16.4 billion total parameters but only about 2.8 billion active per token, achieved through two techniques described in its paper: splitting experts into smaller, more specialized units, and isolating a set of shared experts that always fire to capture common knowledge. Trained from scratch on 2 trillion English and Chinese tokens, it reached performance comparable to the dense DeepSeek 7B and Llama 2 7B models while using only around 40% of the compute those models require at inference. It has a 4,096-token context window and was released under DeepSeek’s open license for self-hosted use, mainly as a research demonstration of expert specialization rather than a production-focused release.

DeepSeek-R1-Distill-Llama-70B

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek-R1-Distill-Llama-70B is a distilled reasoning model DeepSeek released alongside DeepSeek-R1 in January 2025. Rather than training a reasoning model from scratch, DeepSeek fine-tuned Meta’s Llama 3.3 70B Instruct on reasoning traces generated by the full R1 model, transferring much of R1’s chain-of-thought problem-solving ability into a smaller, cheaper-to-run architecture. It is the largest of six distilled R1 variants DeepSeek published, spanning Qwen and Llama base models from 1.5B up to 70B parameters, and on release it scored 70.0% pass@1 on AIME 2024 and 94.5% on MATH-500, far ahead of general-purpose models like GPT-4o and Claude 3.5 Sonnet on the same math benchmarks. It carries a 128,000-token context window, inherited from its Llama 3.3 base, and is distributed under the Llama 3.3 Community License, with weights on Hugging Face and hosted API access through providers including Together AI.

DeepSeek-R1-Distill-Qwen-32B

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek-V2

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek-VL

Thu, 16 Jul 2026 00:00:00 +0000

DeepSeek-VL2

Thu, 16 Jul 2026 00:00:00 +0000

Devstral

Thu, 16 Jul 2026 00:00:00 +0000

DINOv2

Thu, 16 Jul 2026 00:00:00 +0000

DistilBERT

Thu, 16 Jul 2026 00:00:00 +0000

Dolly 2.0

Thu, 16 Jul 2026 00:00:00 +0000

ELECTRA

Thu, 16 Jul 2026 00:00:00 +0000

ELMo

Thu, 16 Jul 2026 00:00:00 +0000

Emu3

Thu, 16 Jul 2026 00:00:00 +0000

ERNIE 4.5

Thu, 16 Jul 2026 00:00:00 +0000

ESM-2

Thu, 16 Jul 2026 00:00:00 +0000

EXAONE 3.0

Thu, 16 Jul 2026 00:00:00 +0000

EXAONE 3.5

Thu, 16 Jul 2026 00:00:00 +0000

Falcon 180B

Thu, 16 Jul 2026 00:00:00 +0000

Falcon 2

Thu, 16 Jul 2026 00:00:00 +0000

Falcon 3

Thu, 16 Jul 2026 00:00:00 +0000

Falcon 40B

Thu, 16 Jul 2026 00:00:00 +0000

Falcon 7B

Thu, 16 Jul 2026 00:00:00 +0000

Falcon Mamba 7B

Thu, 16 Jul 2026 00:00:00 +0000

Flan-T5

Thu, 16 Jul 2026 00:00:00 +0000

Florence-2

Thu, 16 Jul 2026 00:00:00 +0000

FLUX.1 dev

Thu, 16 Jul 2026 00:00:00 +0000

FLUX.1 Kontext

Thu, 16 Jul 2026 00:00:00 +0000

FLUX.1 Krea

Thu, 16 Jul 2026 00:00:00 +0000

Fuyu-8B

Thu, 16 Jul 2026 00:00:00 +0000

Galactica

Thu, 16 Jul 2026 00:00:00 +0000

Gemma

Thu, 16 Jul 2026 00:00:00 +0000

Gemma 2

Thu, 16 Jul 2026 00:00:00 +0000

Gemma 3

Thu, 16 Jul 2026 00:00:00 +0000

GLM-4

Thu, 16 Jul 2026 00:00:00 +0000

GLM-4.6

Thu, 16 Jul 2026 00:00:00 +0000

GLM-5.2

Thu, 16 Jul 2026 00:00:00 +0000

GPT-2

Thu, 16 Jul 2026 00:00:00 +0000

GPT-J

Thu, 16 Jul 2026 00:00:00 +0000

GPT-Neo 2.7B

Thu, 16 Jul 2026 00:00:00 +0000

GPT-NeoX-20B

Thu, 16 Jul 2026 00:00:00 +0000

GPT-OSS-120B

Thu, 16 Jul 2026 00:00:00 +0000

GPT-OSS-20B

Thu, 16 Jul 2026 00:00:00 +0000

Granite 3.0

Thu, 16 Jul 2026 00:00:00 +0000

Granite Code

Thu, 16 Jul 2026 00:00:00 +0000

Granite Embedding

Thu, 16 Jul 2026 00:00:00 +0000

Granite Guardian

Thu, 16 Jul 2026 00:00:00 +0000

Granite Vision

Thu, 16 Jul 2026 00:00:00 +0000

Grok-1

Thu, 16 Jul 2026 00:00:00 +0000

GTE-large

Thu, 16 Jul 2026 00:00:00 +0000

Hermes 2

Thu, 16 Jul 2026 00:00:00 +0000

Hermes 3

Thu, 16 Jul 2026 00:00:00 +0000

Hunyuan DiT

Thu, 16 Jul 2026 00:00:00 +0000

Hunyuan Large

Thu, 16 Jul 2026 00:00:00 +0000

Hunyuan Video

Thu, 16 Jul 2026 00:00:00 +0000

Hunyuan-A13B

Thu, 16 Jul 2026 00:00:00 +0000

Hunyuan3D

Thu, 16 Jul 2026 00:00:00 +0000

IDEFICS2

Thu, 16 Jul 2026 00:00:00 +0000

InternLM2

Thu, 16 Jul 2026 00:00:00 +0000

InternLM2.5

Thu, 16 Jul 2026 00:00:00 +0000

InternVL2

Thu, 16 Jul 2026 00:00:00 +0000

Jamba 1.5 Large

Thu, 16 Jul 2026 00:00:00 +0000

Jamba 1.5 Mini

Thu, 16 Jul 2026 00:00:00 +0000

Jamba Large 1.6

Thu, 16 Jul 2026 00:00:00 +0000

Jina Embeddings v2

Thu, 16 Jul 2026 00:00:00 +0000

jina-embeddings-v3

Thu, 16 Jul 2026 00:00:00 +0000

Kimi K2

Thu, 16 Jul 2026 00:00:00 +0000

Kimi K2.6

Thu, 16 Jul 2026 00:00:00 +0000

Kimi K2.7 Code

Thu, 16 Jul 2026 00:00:00 +0000

Kimi-Audio

Thu, 16 Jul 2026 00:00:00 +0000

Kimi-VL

Thu, 16 Jul 2026 00:00:00 +0000

Kolors

Thu, 16 Jul 2026 00:00:00 +0000

Kosmos-2

Thu, 16 Jul 2026 00:00:00 +0000

LLaMA

Thu, 16 Jul 2026 00:00:00 +0000

Llama 2

Thu, 16 Jul 2026 00:00:00 +0000

Llama 3

Thu, 16 Jul 2026 00:00:00 +0000

Llama 3.1 405B

Thu, 16 Jul 2026 00:00:00 +0000

Llama 3.1 70B

Thu, 16 Jul 2026 00:00:00 +0000

Llama 3.1 8B

Thu, 16 Jul 2026 00:00:00 +0000

Llama 3.2

Thu, 16 Jul 2026 00:00:00 +0000

Llama 3.3

Thu, 16 Jul 2026 00:00:00 +0000

Llama 4 Maverick

Thu, 16 Jul 2026 00:00:00 +0000

Llama 4 Scout

Thu, 16 Jul 2026 00:00:00 +0000

Llama Guard 2

Thu, 16 Jul 2026 00:00:00 +0000

Llama Guard 3

Thu, 16 Jul 2026 00:00:00 +0000

LLeMA-7B

Thu, 16 Jul 2026 00:00:00 +0000

Mathstral

Thu, 16 Jul 2026 00:00:00 +0000

MiniCPM

Thu, 16 Jul 2026 00:00:00 +0000

MiniCPM-V

Thu, 16 Jul 2026 00:00:00 +0000

MiniMax M3

Thu, 16 Jul 2026 00:00:00 +0000

MiniMax-01

Thu, 16 Jul 2026 00:00:00 +0000

Ministral 8B

Thu, 16 Jul 2026 00:00:00 +0000

Mistral 7B

Thu, 16 Jul 2026 00:00:00 +0000

Mistral NeMo

Thu, 16 Jul 2026 00:00:00 +0000

Mistral Small 3

Thu, 16 Jul 2026 00:00:00 +0000

Mixtral 8x22B

Thu, 16 Jul 2026 00:00:00 +0000

Mixtral 8x7B

Thu, 16 Jul 2026 00:00:00 +0000

MPT-30B

Thu, 16 Jul 2026 00:00:00 +0000

mT5

Thu, 16 Jul 2026 00:00:00 +0000

MusicGen

Thu, 16 Jul 2026 00:00:00 +0000

Nemotron Nano

Thu, 16 Jul 2026 00:00:00 +0000

Nemotron-3 8B

Thu, 16 Jul 2026 00:00:00 +0000

Nemotron-4 340B

Thu, 16 Jul 2026 00:00:00 +0000

NLLB-200

Thu, 16 Jul 2026 00:00:00 +0000

Nomic Embed Text v1

Thu, 16 Jul 2026 00:00:00 +0000

nomic-embed-text-v2

Thu, 16 Jul 2026 00:00:00 +0000

NVIDIA Cosmos

Thu, 16 Jul 2026 00:00:00 +0000

OLMo 2

Thu, 16 Jul 2026 00:00:00 +0000

OLMo 7B

Thu, 16 Jul 2026 00:00:00 +0000

OLMoE

Thu, 16 Jul 2026 00:00:00 +0000

OPT-175B

Thu, 16 Jul 2026 00:00:00 +0000

Orca 2

Thu, 16 Jul 2026 00:00:00 +0000

Orca-Math

Thu, 16 Jul 2026 00:00:00 +0000

Pharia-1

Thu, 16 Jul 2026 00:00:00 +0000

Phi-2

Thu, 16 Jul 2026 00:00:00 +0000

Phi-3

Thu, 16 Jul 2026 00:00:00 +0000

Phi-3-medium

Thu, 16 Jul 2026 00:00:00 +0000

Phi-3-small

Thu, 16 Jul 2026 00:00:00 +0000

Phi-3-vision

Thu, 16 Jul 2026 00:00:00 +0000

Phi-3.5-mini

Thu, 16 Jul 2026 00:00:00 +0000

Phi-4

Thu, 16 Jul 2026 00:00:00 +0000

Phi-4-reasoning

Thu, 16 Jul 2026 00:00:00 +0000

Pixtral 12B

Thu, 16 Jul 2026 00:00:00 +0000

Pixtral Large

Thu, 16 Jul 2026 00:00:00 +0000

Playground v2.5

Thu, 16 Jul 2026 00:00:00 +0000

ProGen

Thu, 16 Jul 2026 00:00:00 +0000

Pythia

Thu, 16 Jul 2026 00:00:00 +0000

Qwen-7B

Thu, 16 Jul 2026 00:00:00 +0000

Qwen-Audio

Thu, 16 Jul 2026 00:00:00 +0000

Qwen-Image

Thu, 16 Jul 2026 00:00:00 +0000

Qwen-VL

Thu, 16 Jul 2026 00:00:00 +0000

Qwen1.5-110B

Thu, 16 Jul 2026 00:00:00 +0000

Qwen2-72B

Thu, 16 Jul 2026 00:00:00 +0000

Qwen2-Audio

Thu, 16 Jul 2026 00:00:00 +0000

Qwen2.5-72B

Thu, 16 Jul 2026 00:00:00 +0000

Qwen2.5-Coder

Thu, 16 Jul 2026 00:00:00 +0000

Qwen2.5-VL

Thu, 16 Jul 2026 00:00:00 +0000

Qwen3-235B-A22B

Thu, 16 Jul 2026 00:00:00 +0000

Qwen3-Coder

Thu, 16 Jul 2026 00:00:00 +0000

QwQ-32B

Thu, 16 Jul 2026 00:00:00 +0000

QwQ-32B-Preview

Thu, 16 Jul 2026 00:00:00 +0000

RedPajama-INCITE

Thu, 16 Jul 2026 00:00:00 +0000

Reka Edge

Thu, 16 Jul 2026 00:00:00 +0000

replit-code-v1.5

Thu, 16 Jul 2026 00:00:00 +0000

RoBERTa

Thu, 16 Jul 2026 00:00:00 +0000

SAM

Thu, 16 Jul 2026 00:00:00 +0000

SAM 2

Thu, 16 Jul 2026 00:00:00 +0000

SantaCoder

Thu, 16 Jul 2026 00:00:00 +0000

Seamless M4T

Thu, 16 Jul 2026 00:00:00 +0000

Seamless M4T v2

Thu, 16 Jul 2026 00:00:00 +0000

Skywork-13B

Thu, 16 Jul 2026 00:00:00 +0000

SmolLM

Thu, 16 Jul 2026 00:00:00 +0000

SmolLM2

Thu, 16 Jul 2026 00:00:00 +0000

Solar 10.7B

Thu, 16 Jul 2026 00:00:00 +0000

Solar Pro 2

Thu, 16 Jul 2026 00:00:00 +0000

Stable Cascade

Thu, 16 Jul 2026 00:00:00 +0000

Stable Diffusion 3.5

Thu, 16 Jul 2026 00:00:00 +0000

Stable Diffusion XL

Thu, 16 Jul 2026 00:00:00 +0000

Stable Video Diffusion

Thu, 16 Jul 2026 00:00:00 +0000

StableLM 2

Thu, 16 Jul 2026 00:00:00 +0000

StableLM 3B

Thu, 16 Jul 2026 00:00:00 +0000

StableLM Zephyr

Thu, 16 Jul 2026 00:00:00 +0000

StarCoder

Thu, 16 Jul 2026 00:00:00 +0000

StarCoder2

Thu, 16 Jul 2026 00:00:00 +0000

T5

Thu, 16 Jul 2026 00:00:00 +0000

Tacotron 2

Thu, 16 Jul 2026 00:00:00 +0000

UL2

Thu, 16 Jul 2026 00:00:00 +0000

Vicuna-13B

Thu, 16 Jul 2026 00:00:00 +0000

Vicuna-33B

Thu, 16 Jul 2026 00:00:00 +0000

Wan2.1

Thu, 16 Jul 2026 00:00:00 +0000

Wav2Vec 2.0

Thu, 16 Jul 2026 00:00:00 +0000

Whisper

Thu, 16 Jul 2026 00:00:00 +0000

Whisper Large v3

Thu, 16 Jul 2026 00:00:00 +0000

WizardCoder

Thu, 16 Jul 2026 00:00:00 +0000

WizardLM-2

Thu, 16 Jul 2026 00:00:00 +0000

XGen-7B

Thu, 16 Jul 2026 00:00:00 +0000

XLNet

Thu, 16 Jul 2026 00:00:00 +0000

Yi-1.5-34B

Thu, 16 Jul 2026 00:00:00 +0000

Yi-34B

Thu, 16 Jul 2026 00:00:00 +0000

Yi-Coder

Thu, 16 Jul 2026 00:00:00 +0000

Yi-VL-34B

Thu, 16 Jul 2026 00:00:00 +0000

Yuan 2.0

Thu, 16 Jul 2026 00:00:00 +0000

Zephyr 141B

Thu, 16 Jul 2026 00:00:00 +0000

Zephyr 7B

Thu, 16 Jul 2026 00:00:00 +0000