Models

NVIDIA's 120B open-weights multimodal model for agentic enterprise work — 1M-token context, reads text, images, video, and PDFs.

A 74B open-weights agent model squeezed into 24GB — 2-bit quantization, built to run locally on RTX hardware.

FLUX 3 Image Klein

BFL's 5.6B open-weights image model — 4K generation and pixel-accurate editing on an 8GB consumer GPU.

xAI's flagship speech-to-speech voice agent model — top of the AA speech index at unchanged $0.08/min pricing.

Directable multi-speaker TTS with word-level editing and single-segment audio repair.

LFM2.5-24B-A2B

24B MoE reasoning model with 2B active parameters, built for on-device agents.

Alibaba's multi-modal speaker toolkit: verification, recognition and diarization (audio-only or audio+video) with challenge-grade models like CAM++ and ERes2NetV2, all Apache 2.0.

Scalable native text-to-3D and image-to-3D via primitive diffusion — PBR-ready GLB assets with smooth geometry and materials.

End-to-end AI that replaces the full weather forecasting pipeline.

ACE-Step

Open-source foundation model for lyrics-to-song music generation.

The first open Arabic LLM covering the full pre-train, SFT, and RLAIF pipeline.

Second-generation open Arabic LLM with full localized pretraining pipeline.

NVIDIA's frontier math post-training suite — 72B instruct plus a record-setting math reward model, non-commercial license.

Math reasoning model whose 72B sibling beat GPT-4o on math benchmarks; CC-BY-NC non-commercial.

Zero-shot anomaly detection that adapts CLIP with lightweight hybrid prompts across industrial and medical domains.

Quality-aware margin loss from CVPR 2022: face recognition that holds up on blurry, low-quality images where others fail.

Adobe Firefly

Commercially-safe image and video generation, trained on licensed content.

Commercially safe image generation, 2K output, trained on licensed content only.

Tiny 126M-parameter multilingual model proving small data works: pretrained on under 1GB of text across 11 African languages.

AfroLM

Multilingual model for 23 African languages trained with self-active learning.

Pure-vision GUI agent from ICML 2025: a 7B model that operates phones and desktops from screenshots alone, no DOM or accessibility tree.

Apple's multimodal autoregressive vision encoders, 300M to 3B.

Open Hindi instruction-tuned LLM from IIT Madras.

A Lite BERT: 18x fewer parameters via sharing, still SOTA on GLUE and SQuAD.

Aleph Alpha Kolibri

Germany's new open 78B MoE for documents and reasoning, Apache 2.0.

Spain's sovereign 40B open model trained on MareNostrum 5.

Yandex's Apache 2.0 open-weights flagship for dialogue, RAG, and agents.

Saudi Arabia's sovereign Arabic LLM — 7B from-scratch + 13B, open via IBM watsonx.

Saudi Arabia's 34B Arabic-centric flagship LLM, closed weights.

Saudi Arabia's national Arabic LLM: a 7B open-weight model for Arabic understanding and generation.

Allegro

Open text-to-video diffusion transformer generating 6-second 720p clips.

LLaMA-2-based translator that beats NLLB-54B and GPT-3.5 using monolingual pretraining plus a small set of human-written parallel data.

Alpaca 7B

Stanford's landmark 2023 instruction-following model, trained for under $600.

Code-generation system that reached median human level in competitive programming.

Gemini-powered competitive programming system at the 85th percentile on Codeforces.

RL-discovered sorting and hashing algorithms that shipped in the C++ standard library.

AlphaEvolve

Gemini-powered evolutionary coding agent that discovers algorithms better than human designs.

The protein-folding breakthrough that solved a 50-year grand challenge in biology.

DeepMind's biomolecular structure predictor: proteins with DNA, RNA, ligands and ions — weights open for academic use.

AlphaGenome

DeepMind's DNA sequence model — 1M base pairs, 11 modalities at once.

AlphaGeometry

Neuro-symbolic IMO geometry solver — 25 of 30 olympiad problems, near human gold-medalist level.

Proteome-wide missense variant pathogenicity scores, adapted from AlphaFold and fine-tuned on population variation.

AlphaProof

AlphaZero-style RL in Lean — the first AI system to reach IMO silver-medal standard.

DeepMind's binder-design system that produced picomolar binders, including the first AI-designed VEGF-A binder.

Transformer-based neural decoder for quantum error correction, published in Nature.

RL system that discovered faster matrix multiplication algorithms, published in Nature October 2022.

Amazon's most capable Bedrock model, built for distillation and long-context work

Speech-to-speech foundation model that listens, reasons and speaks in one model for real-time voice agents on Bedrock.

AWS's production text-to-speech service, now with a generative engine for near-human expressive voices.

AWS's managed face API: detect, compare, search collections, and run face liveness checks at $0.001 an image.

The open toolkit for anything-to-audio: TTS, singing voice conversion and synthesis, text-to-audio and vocoders in one MIT-licensed framework with 10,000+ stars.

A lightweight, fast frame interpolator that builds all-pairs correlation volumes to stay accurate on large motion without transformer-scale cost.

A polished SDXL-based anime model with tag-ordered prompting and strong anatomy out of the box.

Zero-shot industrial anomaly detection and localization from CLIP with two learnable, object-agnostic prompts.

Switzerland's fully open 8B/70B sovereign LLM — weights, code, and training data public.

Apertus 70B

Switzerland's fully open 70B multilingual model, Apache 2.0.

Apertus-8B

Switzerland's fully open 8B multilingual LLM, trained on 15T tokens across 1,000+ languages.

Apple Foundation Models

Apple's 3B on-device LLM — private, offline inference for every app.

Arabic BERT pretrained on 77GB of text, the standard Arabic NLU baseline.

Rime's most expressive TTS: trained on annotated everyday conversation — hesitations, laughter, and <laugh> tags included — now serving Arcana v3.

The additive-angular-margin loss that became the default face-recognition embedding — LFW 99.83%.

Aria

Open multimodal mixture-of-experts model with under 4B active parameters

Mathematical superintelligence with every answer formally verified in Lean 4 — IMO gold, zero hallucinations by construction.

600M-param Conformer ASR trained on 12.5M hours; 99 languages, bundled speech intelligence.

Ai2's open 8B model that writes cited science reports in one forward pass.

Athene-V2

Open 72B post-trained suite from Nexusflow targeting GPT-4o-level chat and agent use.

Text-to-audio generation: describe a soundscape, get the audio.

Latent-diffusion text-to-audio model that hit SOTA training on a single GPU.

AudioLDM 2

Latent diffusion text-to-audio that turns prompts into sound effects, speech, and music

Text-to-audio latent diffusion model for sound effects, speech and music.

AudioLDM2

Open text-to-audio, music, and speech generation from a unified latent diffusion model.

The 2022 paper that reframed audio generation as language modeling.

Research model fusing PaLM-2 with AudioLM to speak and translate speech.

Neural audio watermarking with sample-level detection of AI speech.

A foundation model for open-domain sound separation that pulls any sound out of a mixture from a plain-text description.

Text-to-audio latent diffusion built on Stable Diffusion, with open weights.

Deepgram's multilingual TTS for the enterprise voice stack: 7 languages, 90 voices, streaming over WebSocket — the one-shot counterpart to its Flux agent model.

The largest fully open-source flow-based text-to-image model, MMDiT under Apache 2.0.

Open-source 600M-parameter GAN upscaler that quadruples images in about a quarter second.

xAI's photorealistic image model for Grok — powerful, with notably few content guardrails.

Foundation model for the atmosphere — weather, air quality, cyclones.

A 7B video captioner that writes long, structured, human-readable video descriptions — a simple LLaVA-style design plus frame-wise token merging keeps long videos cheap, and it ships with VDC, a new benchmark built specifically for detailed captioning.

Uses foundation models to orchestrate fleets of robots collecting their own training data in the wild.

Desktop AI video enhancement suite — upscaling to 8K, denoising, colorization, stabilization, and face recovery in one app.

Cohere's 101-language instruction-tuned open model — the multilingual breadth milestone.

Open-weight multilingual models covering 23 languages in depth.

35B multilingual research model serving 23 languages with open weights.

Aya 23 8B

Open weights multilingual model covering 23 languages in depth.

Cohere's open-weight 32B multilingual model covering 23 languages.

Aya Expanse 8B

Multilingual 8B model covering 23 languages — research/non-commercial license.

Cohere's open-weight multilingual 8B covering 23 languages, trained with data arbitrage.

Aya Vision

Open-weight multilingual vision-language models (8B/32B) for research use.

Multilingual 8B vision-language model from Cohere's open research lab, released March 2025.

Microsoft's face detection and verification API — now gated behind limited-access enrollment, with emotion and gender inference retired.

Microsoft's enterprise TTS: 400+ neural voices, 140+ locales, SSML, streaming and batch on Azure.

Babel-9B

Open multilingual LLM covering the top 25 languages, serving over 90 percent of the world.

BAGEL

Unified 7B multimodal model that both understands and generates images

Chinese-English 13B bilingual model from Wang Xiaochuan's Baichuan.

Baichuan's first-generation 13B bilingual chat model under a community license.

Baichuan-M1

Medical-domain LLM trained from scratch on 20 trillion tokens, with an open 14B version.

Baichuan-M2

Open 32B medical LLM that beat GPT-class models on HealthBench; runs on one RTX 4090.

RLHF-tuned 13B bilingual Chinese-English chat model from Baichuan.

Bilingual Chinese-English 7B LLM trained on 2.6 trillion tokens.

IBM's open hybrid Mamba2-transformer model: 9.8B parameters with a fraction of the KV-cache of a dense transformer.

The generalized bandsplit network for cinematic audio — dialogue, music, and effects from a mix, with a query-based sibling for music stems.

Denoising seq2seq pretraining that generalizes BERT, GPT, and more.

The landmark recurrent video super-resolution model that lifted BasicVSR with second-order grid propagation and flow-guided deformable alignment.

Real-time and offline joint beat, downbeat, tempo, and meter tracking.

Self-supervised audio pre-training with acoustic tokenizers; a workhorse audio embedding backbone.

The bidirectional transformer that defined the pre-train/fine-tune era of NLP.

BERT for recommendations — the bidirectional sequential model that learns user taste by filling in masked items.

A 7B Lean 4 tactic prover showing scaled best-first search beats MCTS — 72.95% on miniF2F with no critic network.

BGE Large EN v1.5

The classic English embedding — 10M downloads of MIT-licensed retrieval reliability.

BAAI cross-encoder reranker for scoring query-document relevance in RAG pipelines.

568M-parameter multilingual cross-encoder reranker — the RAG quality upgrade.

BGE-M3

The world's most-downloaded embedding model — 32.8M downloads of multilingual, multi-function retrieval.

bge-reranker-v2-m3

BAAI's 568M multilingual cross-encoder reranker, the default second stage for open RAG.

India's sovereign 17B MoE — 22 Indian languages, open weights.

Poland's open-source Polish-first LLM family, from 7B to 11B instruct models.

Poland's open sovereign LLM with a dedicated Polish tokenizer, trained on national supercomputers.

Poland's flagship open LLM — 11B params, Polish-first tokenizer, beats 70B models on Polish benchmarks.

Efficient Polish mid-size model from the Bielik v3 generation.

Poland's first open Polish LLM, released April 2024 by the SpeakLeash community.

Distilled 7.5B Polish model from the Bielik v3 generation.

Google's sparse-attention transformer for 4K sequences — with theoretical guarantees.

Bilibili Index-Translate

Open Apache 2.0 translation models for 150 languages — 2B laptop-sized up to a 35B MoE, with a free API.

BioEmu-1

Deep-learning emulator of protein structural ensembles, orders of magnitude faster than MD.

GPT-style model pretrained from scratch on 15M PubMed abstracts, MIT licensed.

A biomedical CLIP trained on 15 million figure-caption pairs from the scientific literature.

2.7B biomedical language model trained purely on PubMed literature.

BiomedParse

Foundation model unifying detection, recognition, and segmentation across 9 medical imaging modalities.

Open medical LLM — Mistral 7B further pretrained on PubMed Central, Apache 2.0.

Microsoft's native 1-bit LLM: 2B parameters in ternary weights that run on a CPU with ~0.4 GB of memory.

The first native 1-bit LLM — ternary weights, CPU inference, full-precision-class quality.

Microsoft's 2.4B-parameter native 1-bit LLM: ternary weights trained from scratch on 4 trillion tokens.

Microsoft's native 1-bit LLM: 2B ternary-weight parameters trained on 4T tokens, MIT licensed, runs on CPU via bitnet.cpp.

Facebook AI's 2020 open-domain chatbot that blended empathy, knowledge, and persona.

Efficient vision-language pretraining by bridging frozen encoders with a Q-Former.

BLIP-3

Salesforce's open multimodal family (xGen-MM) — vision-language models you can actually download.

Fully open unified model for image understanding and generation.

Korean-English bilingual LLM built on Llama 3 with vocabulary expansion.

176B open multilingual model built by 1,000+ researchers across 46 languages; landmark RAIL-licensed release.

50B-parameter finance LLM trained on 363B tokens of Bloomberg data; research paper only, no public weights.

Instruction-tuned BLOOM — crosslingual multitask finetuning that unlocked zero-shot generalization in 46 languages.

Sovereign Indian TTS — 50+ native voices across 10 languages at ₹12 per 10,000 characters.

Boltz-1

Fully open-source biomolecular structure predictor matching AlphaFold3's accuracy.

Boltz-2

Biomolecular co-folding model predicting structure and binding affinity

Bonsai 2 27B

27B-class reasoning, coding, and vision compressed to 5.9GB via ternary weights.

AI-composed functional music engineered for focus, sleep, and relaxation, backed by the company's own neuroscience research.

MediaTek's open 7B model optimized for Traditional Chinese and English.

Commercial-safe object removal API trained on licensed data, erasing masked regions and inpainting them with context-aware detail.

Plug-and-play diffusion inpainting adapter that injects masked-image features into any pretrained diffusion model via a dual-branch design.

ByteDance's band-split Transformer that won the Sound Demixing Challenge 2023 — the architecture behind today's best stem splitters.

The band-split RNN that started the modern separation era — split the spectrum by musical knowledge, model bands and time alternately.

India's voice model for Indian languages: sub-250ms streaming TTS across 11 languages with Hinglish code-switching and native name pronunciation — ₹20 per 10K characters.

Multi-object tracking by associating every detection box, including the low-confidence ones everyone else throws away.

C4AI Command A

Cohere's 111B open-weights model for agentic and multilingual enterprise work.

Pioneering Brazilian Portuguese instruction-tuned LLM based on OpenLLaMA 3B.

Reverse-complement-equivariant DNA modeling with Mamba, reaching 131k base pairs without attention.

CyberAgent's Apache 2.0 Japanese-English bilingual chat model.

Fully open vision-centric multimodal family and training cookbook

Brazilian Portuguese 7B LLM with extended training on Portuguese web text.

Ultralight 182M multilingual speech-to-text and translation model.

Canary-1B

Multilingual ASR plus speech translation in 25 languages — commercial-friendly open weights

Canary-1B

Multilingual speech recognition and translation under open terms.

1B multilingual speech model for transcription and translation across 25 European languages.

ASR and LLM hybrid that transcribes audio then reasons over the text.

2.5B-parameter ASR-LLM hybrid with 5.63% word error rate in English.

Cartesia Sonic

Ultra-low-latency commercial TTS API — expressive voices in under 90ms, built on state-space models.

Cartesia Sonic 3

Streaming TTS on state-space models — 42 languages, natural laughter, sub-100ms streaming.

Cendol

Open instruction-tuned LLM collection for Indonesian and regional languages, 300M to 13B.

Multimodal foundation model for molecular structure prediction.

Generative AI for de novo antibody design with experimental validation.

Meta's early-fusion token-based multimodal foundation model family.

3B PaliGemma-based specialist for chart reasoning in the wild: QA, summarization and fact-checking from chart images.

The 6.2B bilingual chat model that kicked off China's open LLM wave.

Zhipu's second-generation bilingual 6B chat model with 32K context.

Bilingual Chinese-English 6B dialogue model with native function calling and code execution; open weights.

Open-source LLaMA2-based LLM that treats music as a second language via ABC notation: composes, answers theory questions, and chats.

A multimodal LLM that chats about 3D human pose — estimation and generation from text or images.

Chatterbox

MIT-licensed open TTS family — 23 languages, emotion control, beats ElevenLabs in evals.

Resemble AI's 110M-parameter edge TTS: Turbo's single-step architecture shrunk to run 3x real-time on an 8-core CPU — zero-shot cloning and [laugh]-style tags included.

Chatterbox Turbo

MIT-licensed 350M-param open TTS — one-step mel decoder, paralinguistic tags, PerTh watermarked.

ChatTTS

Conversational TTS for Chinese and English — laughter, pauses, and dialogue dynamics from text.

ChemBERTa-2

Chemical foundation model: RoBERTa pretrained on 77M molecules.

70B model on 1.4T tokens that rewrote LLM scaling laws; never publicly released.

NVIDIA's domain-adapted chip-design LLMs: LLaMA 2 continued-pretrained on 20B+ tokens of chip data, matching GPT-4 at 70B.

Chirp 3 HD

Google's Gemini-powered HD text-to-speech — 30 voices, 50+ languages

Multilingual vision-language model for nine Indian regional languages.

Chronos

Amazon's language-modeling framework that tokenizes time series for probabilistic forecasting.

120M Apache-2.0 time-series foundation model with native multivariate and covariate support.

The production Chronos: patch-based direct forecasting, 250x faster than the original.

Multilingual music information retrieval across 101 languages, for ABC notation and MIDI.

Anthropic's 100K-context model that pioneered the long-context era.

Claude 3 Haiku

Anthropic's fastest, most affordable Claude 3 model, built for high-volume enterprise workloads.

Claude 3 Opus

Anthropic's most capable 2024 model, the benchmark for careful reasoning.

Anthropic's June 2024 workhorse — beat Opus while running faster, then pioneered computer use.

Claude 3.7 Sonnet

The first hybrid reasoning model: instant answers or extended thinking in one model

Claude Fable 5.1

Anthropic's public flagship — Mythos-class intelligence with safeguards, $10/$50.

Claude Haiku 4.5

Near-frontier coding speed and quality at one-third the cost of Sonnet

Claude Haiku 5.5

Anthropic's cheapest, fastest small model ever — $0.10/M input, with a 1M-token context.

Claude Opus 4

Anthropic's flagship for the hardest tasks, priced for teams that need every last point

Claude Opus 4.1

Anthropic's top coding model of 2025: 74.5% SWE-bench Verified, hybrid reasoning.

Claude Opus 4.5

The first model to break 80% on SWE-bench Verified — Anthropic's 2025 flagship for coding, agents, and computer use

Anthropic's first 1M-context Opus — adaptive thinking for long agentic coding runs.

Anthropic's April 2026 flagship: sharper vision, xhigh reasoning effort, and steadier long-horizon agents.

Claude Opus 4.8

The refined pre-5.0 flagship — 1M context, Fast Mode, and dynamic agent workflows.

Claude Opus 5

Flagship Opus at near-Fable capability for roughly half the price, API-only.

Anthropic's flagship general-purpose model, leading agentic coding benchmarks at 40% less cost than Opus 5.

Claude Sonnet 4

Anthropic's workhorse frontier model: near-flagship capability at a mid-tier price

Claude Sonnet 4.5

Anthropic's frontier coding model: 77.2% SWE-bench Verified, 30-hour agentic runs.

Claude Sonnet 4.6

Mid-tier Claude model with near-flagship intelligence at Sonnet pricing.

Claude Sonnet 5.5

Anthropic's fast mid-tier frontier model — 30%+ faster than Sonnet 5 at the same per-token price.

The open speech-processing studio: pretrained FRCRN denoisers, MossFormer separators, super-resolution, and audio-visual target-speaker extraction in one pip package.

Clef

Cloudflare's open-weight decision model — typed probabilities instead of prose, 27B, Apache 2.0.

Clef Flash

The 9B low-latency Clef — sub-40ms decisions at $0.038 per million tokens.

Clef Omni

Clef goes multimodal — one decision call scores text, images, audio, and video.

Open medical LLM with dialogue-based knowledge encoding, 13B and 70B variants.

T5-large further pretrained on MIMIC clinical notes for medical text generation.

OpenAI's 2021 vision-language breakthrough — zero-shot image classification from 400M image-text pairs.

Conformer-based metric GAN that pushed VCTK/DEMAND denoising to 3.41 PESQ by training the generator against a learned quality score.

The accuracy-first on-device VAD: Picovoice's engine that reports 98.9% true-positive rate at 5% false-positive rate with ~5% CPU on a Raspberry Pi Zero — fully local, no audio ever leaves the device.

Image-text foundation model training contrastive and captioning losses jointly.

Meta's code-specialized Llama 2 family — the template every later open coder competed against.

Meta's 70B code specialist — 53% HumanEval, near GPT-4 class at release.

Ant Group's multi-task fine-tuned CodeLlama 34B: 74.4% HumanEval, beating GPT-4 in the authors' tests.

CodeGeeX4

Open 9B multilingual code model on the GLM-4 backbone.

Multilingual 9B code model with function calling, code interpreter, and web search.

Google's open code models on the Gemma base: 2B infill specialist and 7B instruct.

CodeGemma 7B

Google's Gemma-licensed code model — lightweight completion and generation for everyday coding.

CodeLlama 34B

Meta's pioneering code LLM — the model that proved open code models could compete.

CodeLlama 7B Instruct

Instruction-tuned 7B code model built on Llama 2, trained on 500B code tokens.

CodeLlama-34B-Instruct

Meta's 34B instruction-tuned code model for serious code work.

The 7B predecessor to Qwen2.5-Coder, under the Tongyi Qianwen license.

The SIGMOD 2024 open text-to-SQL model family: 1B to 15B, built specifically for schema-aware SQL generation.

7B multilingual code foundation model with VS Code and JetBrains plugins.

Codestral 22B

Mistral's 22B code specialist — research/non-production license.

Codestral 22B

Mistral's 22B fluent-code model — 80+ languages under the non-production MNPL license.

Codestral 25.01

Mistral's FIM-specialist code model: 256K context, 80+ languages, 2x faster.

Mistral's enterprise code model: 80+ languages, low-latency fill-in-the-middle, at the heart of the Mistral coding stack.

Codestral 25.08

Mistral's August 2025 code model refresh with 256K context and FIM completion.

Codestral Embed

Mistral's code-tuned embedding model that beat OpenAI and Cohere on retrieval benchmarks.

7B Mamba2 code model with a 256K context window under Apache 2.0.

Identifier-aware encoder-decoder model for code understanding and generation.

Salesforce's encoder-decoder code LLM family (220M to 16B) for understanding and generation, with an instruction-tuned 16B flagship.

Componentized VLA that splits cognition from action: a VLM reasons, a diffusion transformer acts.

18B vision-language model built for GUI understanding and navigation.

Cogito v2

Open-weight hybrid reasoning model family with self-improving machine intuition.

Cognition's most capable SWE model — reinforcement learning stacked on RL, served at 1,000 tokens per second.

Zhipu's 5B open text-to-video model — 10s clips, the first commercial-grade open video generator.

CogVideoX-5B

Zhipu AI and Tsinghua's 5B open-weights text-to-video diffusion model with an expert transformer

CogVideoX1.5-5B

Zhipu's 5B open video model — text-to-video and image-to-video from the CogVideo lineage.

Open-source text-to-image model with Chinese character generation.

CogVLM

Open vision-language model with a trainable visual expert module for deep fusion.

CogVLM2

High-resolution open VLM aimed at GPT-4V-level image understanding.

111B enterprise reasoning model with token-budgeted thinking, open weights.

Cohere's first vision model: 112B, 128K context, open weights for research (CC-BY-NC).

Cohere Embed 4

Cohere's multimodal embedding model — vectorize text, images, and whole PDFs into one space.

1024-dim managed embedding with task-specific input types

Cohere Embed v3

Cohere's enterprise embedding API — multilingual with a compression endpoint, API-only.

Cohere Embed v4

Cohere's multimodal embedding model: text and images in one vector space, 128K context.

Enterprise reranker with 100+ languages for precision retrieval.

The default commercial cross-encoder reranker for two-stage retrieval

Cohere's 32K-context reranker in Pro and Fast flavors — built so AI agents can rank whole documents, not fragments.

Cohere Transcribe

Cohere's first audio model — a 2B open-weights transcriber that took #1 on the Open ASR Leaderboard.

ColBERTv2

Late-interaction retrieval: token-level matching with MaxSim scoring

ColPali

Vision-language document retriever that indexes pages as images, no OCR needed.

Visual document retrieval that treats every page as an image.

Reference-free quality estimation for machine translation: score any translation using only the source text.

Command A

Cohere's 111B-parameter enterprise LLM built for RAG, tool use and agentic work on just two GPUs.

Cohere's open-weight enterprise flagship — a sparse MoE built for RAG, tool use, and sovereign deployment.

35B open-weights model built for RAG and multilingual generation

Command R+

128K-context RAG specialist with native tool use — non-commercial weights.

Command R+ (08-2024)

Cohere's 104B enterprise LLM, August 2024 refresh.

Command R7B

7B multilingual RAG and tool-use model for research use.

Free, open-source, self-hosted face recognition server: one docker command gives you a full REST API for detection, recognition, age, and gender.

The fully-convolutional time-domain separator that beat ideal masking on speech — the foundation every modern separator builds on.

2B Apache-2.0 agent model tuned for Huawei Ascend NPUs.

The community-carried successor to Mozilla DeepSpeech: trainable, streamable, small-footprint STT — now unmaintained.

Open omnimodal world model — reasons, generates, and acts in one pass.

Cosmos Predict-1

NVIDIA's open world-foundation models that predict physics-aware video for robots and AVs.

Cosmos Reason

Open reasoning vision-language model for robots and physical AI.

CosyVoice

Multilingual zero-shot TTS with voice cloning, open weights.

Zero-shot multilingual voice cloning TTS across 9 languages and 18 Chinese dialects

Alibaba's open zero-shot TTS with natural-language instruction control over speaking style.

Half-billion-parameter multilingual TTS, RL-tuned for naturalness.

The CNN that made neural pitch tracking mainstream: feed it raw 16 kHz audio and it returns a confidence-scored F0 estimate every 10 ms, trained to beat pYIN and its classical predecessors.

Single image to textured 3D mesh in about 10 seconds — open MIT feed-forward reconstruction.

Tiny 1.3B truly bilingual French-English model with fully open data and code.

CSM

Sesame's open-weight conversational speech model that turns text into natural, context-aware dialogue audio.

1.6B conditional transformer with control codes for steerable generation.

Continuous 3D perception with persistent state from streaming images.

Object-level memory reading for video segmentation: remembers the object, not just pixels.

Pioneering Japanese GPT-NeoX model released under CC BY-SA 4.0.

Sony's densely-connected multidilated DenseNet — the 2020 MUSDB18 benchmark that DenseNet-era separation is measured against.