Every AI model release, by provider.
Frontier LLMs, reasoning models, open weights, and the image, video and audio models that matter, from 104 labs across the US, China, Europe, Asia-Pacific and the Middle East. Every entry links to the provider's own announcement or changelog. Updated twice a day.
- 420releases tracked
- 104providers
- 5regions
- 218open-weight
Just shipped.
-
Thinking Machines Lab Inkling 2 Mini
Looped MoE (26.4B total, 3.4B active; middle layers run 3x) with image input, ~20GB NVFP4/MXFP8 checkpoint for laptops. Apache 2.0 weights on HF; on Tinker.
-
Baseten bison-1 and groundhog-1
Gemma 4-based decision models (bison-1 31B dense; groundhog-1 26B MoE, ~4B active), text+image in; bison-1 tops AA Decision Index v2.1 (76). $0.10 and $0.035/M tokens.
-
Tencent Hunyuan WeMM-Video (2B, 4B, 9B and 35B-A3B)
WeChat Vision omni-modal family: text, image, audio and video in, text out, with thinking and non-thinking modes. ~1.2T training tokens. Apache 2.0 weights on HF.
-
OVHcloud (OVHai LLM) Olala v1.0 mini 7A1B
OVH's first LLM: 6.8B total / ~1B active hybrid Mamba-3/attention MoE, trained from scratch in Europe; four reasoning levels and tool use. Apache 2.0; base and SFT checkpoints too.
-
Microsoft AI Aion Instruct
2.3B on-device small language model, now GA as open weights on Hugging Face and Edge's built-in model (replacing Phi-4-mini) behind the Prompt API origin trial.
-
RadixArk DaisyRose-27B
RadixArk's first model: Qwen3.8-27B post-trained with RL only on ~1,200 agent tasks via Miles. Terminal Bench 4.1: 12.0 vs 5.6 base (self-reported). Apache 2.0 weights.
-
Google DeepMind Gemini 4 Neon
Sits between Hydrogen and Carbon; near-Carbon coding and reasoning at faster speed, four thinking levels. $3/$15 per 1M tokens; replaces Argon in the Gemini app.
-
Cloudflare Clairic and Clairic-flash
Redesigned Clairic (27B) leaves preview and new 9B Clairic-flash joins it: typed extraction models that now read images and PDFs. Open weights on Hugging Face; on Workers AI.
-
SenseTime (SenseNova) SenseNova-Loop-9B
Official release after the Oct 31 preview: ~9B looped-transformer multimodal model that runs its middle 24 layers twice. Apache 2.0, gated download.
-
OpenAI GPT-6.2 Astra and GPT-6.2 Sol
New frontier Astra ($8/$40 per M tokens, 20% below GPT-6 Astra) and everyday Sol (keeps 6.1 Sol pricing). In ChatGPT Work, Codex and the API.
Every lab, newest first.
1 release latest Nov 20, 2026
-
bison-1 and groundhog-1
Gemma 4-based decision models (bison-1 31B dense; groundhog-1 26B MoE, ~4B active), text+image in; bison-1 tops AA Decision Index v2.1 (76). $0.10 and $0.035/M tokens.
2 releases latest Nov 20, 2026
-
Inkling 2 Mini
Looped MoE (26.4B total, 3.4B active; middle layers run 3x) with image input, ~20GB NVFP4/MXFP8 checkpoint for laptops. Apache 2.0 weights on HF; on Tinker.
-
Inkling 2 Small
594B MoE (17B active) with image input and variable effort. Launched on Tinker; Apache 2.0 weights released on Hugging Face Nov 12, with API access via NVIDIA, Fireworks, Together and others.
14 releases latest Nov 19, 2026
-
Aion Instruct
2.3B on-device small language model, now GA as open weights on Hugging Face and Edge's built-in model (replacing Phi-4-mini) behind the Prompt API origin trial.
-
MAI-Cyber-2-Flash, MAI-Realtime-1, MAI-Image-2.7-Flash and Decision 1.1
Rest of Microsoft's Ignite MAI lineup: a cyber-defense model, its first speech-to-speech model (from $0.01/audio min), a fast image model and a routing decision model.
-
MAI-Code-2 and MAI-Code-2-Flash
Everyday enterprise coding models announced at Ignite. Both in GitHub Copilot; Code 2 Flash also in limited-access preview in Microsoft Foundry.
-
Mage-VL-Cowork
4B computer-use agent combining GUI control and bash in one vision-language policy; 66.0% on OSWorld-Verified, best at 4B. Apache-2.0 weights.
-
MAI-Image-2.7
Image editing model launched in preview; No. 3 on Arena Edit and No. 2 on Artificial Analysis image editing at launch.
-
Microsoft-Decision-1
Fast decision-scoring model (post-trained Qwen3.5-9B) for routing, classification and workflow control; on Microsoft Foundry and OpenRouter.
-
MAI-Transcribe-2-Streaming, MAI-Voice-2.1 and 2.1-Flash
-
MAI-Image-2.6 and MAI-Image-2.6-Flash
-
MAI-Thinking-1
Public preview in Microsoft Foundry.
-
MAI-Code-1.1-Flash
Small-tier coding model in GitHub Copilot; adds native vision over MAI-Code-1-Flash at a 73% lower list price.
-
MAI-Code-1-Flash, MAI-Image-2.5, MAI-Transcribe-1.5, MAI-Voice-2
Seven-model MAI family announced at Build 2026.
-
MAI-Image-1
-
MAI-Voice-1 and MAI-1-preview
First in-house MAI models.
-
Phi-4-reasoning
1 release latest Nov 19, 2026
-
Olala v1.0 mini 7A1B
OVH's first LLM: 6.8B total / ~1B active hybrid Mamba-3/attention MoE, trained from scratch in Europe; four reasoning levels and tool use. Apache 2.0; base and SFT checkpoints too.
14 releases latest Nov 19, 2026
-
WeMM-Video (2B, 4B, 9B and 35B-A3B)
WeChat Vision omni-modal family: text, image, audio and video in, text out, with thinking and non-thinking modes. ~1.2T training tokens. Apache 2.0 weights on HF.
-
Hy4 flash
Cheapest Hy model yet: 220B MoE (64B Engram) with ~12B active, text and image input, 1M context. Launched Nov 16 on TokenHub and Yuanbao; open weights (Hy4 flash, FP8, Base) followed Nov 18.
-
Hy Realtime
Real-time voice model in the Hunyuan line with agentic tool calling (web search) and singing. Live in Yuanbao voice calls and Tencent customer service; Tencent Cloud API coming soon.
-
Hy4
770B MoE flagship (49B active) with vision and computer use, 1M context; about $0.90/$2.70 per 1M tokens. Open weights (Hy4, FP8, Base; Tencent Hy License) followed Nov 20.
-
Hy-SimulMT
30B-A3B real-time simultaneous speech translation model for Chinese and English in both directions, with real-time voice cloning; FP8 and GGUF builds, also in the Hy Translate app.
-
Hy-Nomas
Apache-2.0 interactive video world model: generates a world segment by segment, steered by multi-turn text instructions and camera control. Ships with Hy-Nomas-Bench.
-
WeVisOne-VL (2B and 4B)
WeChat Vision model built on Qwen3-VL that unifies detection, referring localization, counting, pointing and OCR with referring segmentation and alpha matting. Apache-2.0 code.
-
Hy-Memory-30B-A3B
MoE long-term-memory writer (~3B active) that replaces the general LLM inside Mem0, MemOS and Hy-Memory; Apache 2.0.
-
Youtu-Parsing-Omni
5B omni-modal parsing model: documents, images, charts, geometry, audio and video → structured JSON; SOTA OmniDocBench v1.6 (96.96).
-
Hy4 preview
-
Hy3
295B MoE, Apache 2.0.
-
Hy3 preview
-
HunyuanImage 3.0
-
Hunyuan-T1
4 releases latest Nov 18, 2026
-
Clairic and Clairic-flash
Redesigned Clairic (27B) leaves preview and new 9B Clairic-flash joins it: typed extraction models that now read images and PDFs. Open weights on Hugging Face; on Workers AI.
-
Clairic
Preview of Cloudflare’s first System One extraction model (@cf/cloudflare/clairic on Workers AI), returning typed values with confidence; open weights planned.
-
Clef-omni
Multimodal decision model (Qwen3-Omni 30B-A3B backbone) scoring text, image, audio and video in one call; Apache 2.0 on Workers AI.
-
Clef and Clef-flash
Cloudflare's first trained models: 27B and 9B decision models (Qwen backbones) returning typed answer probabilities; Jev-API compatible, Apache 2.0, on Workers AI.
36 releases latest Nov 18, 2026
-
Gemini 4 Neon
Sits between Hydrogen and Carbon; near-Carbon coding and reasoning at faster speed, four thinking levels. $3/$15 per 1M tokens; replaces Argon in the Gemini app.
-
Gemini 4 Hydrogen
Fast Gemini 4 model for coding and agentic work, replacing Gemini 3.8 Flash. Four thinking levels up to xhigh; $1.50/$7.50 per 1M tokens across the full 1M context.
-
Gemini Decision 1
Decision model for the new Decisions API: calibrated yes/no, single-select and rubric probabilities over text, images, audio, video and PDFs with zero output tokens, up to 1,000 questions per request.
-
AlphaGenome 2
Conditional DNA sequence model predicting variant effects per cell type, trained on 107M+ single-cell profiles. Non-commercial research via AlphaGenome API.
-
Gemini Robotics 2.5
DeepMind's newest VLA robot-control model: 30-100% higher task success and 20-30% faster than Gemini Robotics 2 over 200+ tasks; 96% in a Mercedes-Benz kitting pilot. Partners only.
-
Gemini Omni 1.2 Flash
GA video generation/editing with native 15-second clips, multi-turn conversational edits, stronger reference adherence, and 30% lower pricing vs Omni 1.1.
-
Gemini 4 Carbon
Frontier model for cybersecurity and long-horizon tasks; first offered to cyber defenders via Fairwind. GA on Nov 10 in the Gemini API ($2/$10 per 1M tokens intro) and Gemini app.
-
Genie 4
Real-time interactive world model with native audio and dialogue; research preview for trusted testers, coming to Project Genie and Gemini API early access.
-
Gemini Nano Banana 2.1
GA update to Nano Banana 2 (gemini-nano-banana-2.1): better visual quality, prompt adherence, text rendering and new panoramic aspect ratios up to 4K.
-
EmbeddingGemma 2
740M Apache-2.0 on-device embedder built on Gemma 4; maps text, code, images, audio and video into one space, 8K-token context.
-
Gemini 4 Argon
Initially limited to trusted cyber defenders; generally available via the Gemini API and Gemini app on Oct 13.
-
Gemini 3.8 Flash TTS and Flash-Lite TTS
-
Lyria 3.5
Full-length song generation.
-
Gemini 3.8 Flash
-
Gemini 3.7 Flash
-
Gemini Robotics ER 2
Embodied-reasoning VLM on Gemini 3.5 Flash that acts as a robot's high-level brain: video reasoning, task planning, tool use and multi-robot coordination. Gemini API preview; Google Cloud from Oct 27.
-
Gemini 3.6 Flash and Gemini 3.5 Flash-Lite
-
Gemini Omni Flash (preview)
Conversational video generation and editing.
-
Gemini 3.5 Flash
-
Gemma 4
Moved to Apache 2.0.
-
Lyria 3
Music generation.
-
Gemini 3.1 Flash-Lite (preview)
-
Gemini 3.1 Flash Image (Nano Banana 2)
-
Gemini 3.1 Pro (preview)
-
Gemini 3 Flash
Preview.
-
Gemini 3 Pro Image (Nano Banana Pro)
-
Gemini 3 Pro
-
Veo 3.1 and Veo 3.1 Fast
-
Gemini 2.5 Flash Image (Nano Banana)
-
Gemini 2.5 Pro and Flash (GA), 2.5 Flash-Lite (preview)
-
Veo 3 and Imagen 4
Announced at Google I/O 2025.
-
Gemma 3n
On-device preview.
-
Gemini 2.5 Flash (preview)
-
Gemini 2.5 Pro (experimental)
-
Gemma 3
-
Gemini 2.0 Flash (GA), 2.0 Pro and 2.0 Flash-Lite
1 release latest Nov 18, 2026
-
DaisyRose-27B
RadixArk's first model: Qwen3.8-27B post-trained with RL only on ~1,200 agent tasks via Miles. Terminal Bench 4.1: 12.0 vs 5.6 base (self-reported). Apache 2.0 weights.
3 releases latest Nov 17, 2026
-
Kling Image 4.0
Kling's new image model: native 8K output, up to 2,000 characters of text per image in 7 languages, and sets of up to 10 consistent images from one prompt.
-
Kling 4.0 Flash
Lighter Kling 4.0 variant, opened to all users and the API after early access for Ultra subscribers: 3-20 s 720p text-, image- and omni-reference-to-video.
-
Kling 4.0
Global launch: native 30-second generation, up to 15 multimodal references and 10 keyframes; API for developers coming soon.
27 releases latest Nov 17, 2026
-
GPT-6.2 Astra and GPT-6.2 Sol
New frontier Astra ($8/$40 per M tokens, 20% below GPT-6 Astra) and everyday Sol (keeps 6.1 Sol pricing). In ChatGPT Work, Codex and the API.
-
GPT-6 Cyber
Cybersecurity-specialized GPT-6 Astra variant for authorized vulnerability discovery and exploit validation. Daybreak Red gated access; Critical cyber preparedness designation.
-
GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare (Images 2.6)
50% faster generation, animated GIF and SVG output, markup-guided edits and a new complexity control. Live in ChatGPT, Codex and the API.
-
GPT-Rosalind-6
Life-sciences reasoning model on GPT-6 Astra weights; trusted access only, in Work and Codex (API where approved). Rated High capability in biology.
-
GPT-6.1 Terra
Fast, efficient successor to GPT-5.6 Terra focused on visual reasoning and computer use; in ChatGPT Work and Codex and via the API as gpt-6.1-terra at $1/$5 per million input/output tokens.
-
GPT-Live-1.1 and GPT-Live-1.1 Quip
Full-duplex voice models succeeding GPT-Live-1 and GPT-Live-1 mini; new defaults for ChatGPT Voice, Codex and Dots.
-
GPT-6.1 Sol
-
GPT-6 Sol and GPT-6 Luna
-
GPT-6 Astra
Initially limited to select organizations.
-
GPT-5.6 Sol
Limited preview from June 26, GA July 9.
-
GPT-5.5
API availability followed on April 24.
-
GPT-5.4 mini
-
GPT-5.4 (Thinking, Pro)
-
GPT-5.3-Codex
-
GPT-5.2
-
GPT-5.1
-
Sora 2
-
GPT-5-Codex
-
gpt-realtime
Speech-to-speech model.
-
GPT-5
-
gpt-oss-120b and gpt-oss-20b
First open-weight OpenAI LLMs since GPT-2.
-
o3-pro
-
gpt-image-1
-
o3 and o4-mini
-
GPT-4.1 (incl. mini, nano)
-
GPT-4.5
Research preview.
-
o3-mini
3 releases latest Nov 17, 2026
-
SenseNova-Loop-9B
Official release after the Oct 31 preview: ~9B looped-transformer multimodal model that runs its middle 24 layers twice. Apache 2.0, gated download.
-
SenseNova 6.8 Flash
General release of SenseTime’s multimodal agent model for office work (long-term memory, cross-session tasks) and design output (web, SVG, Blender, animated PPT). Via SenseNova Studio and API.
-
SenseNova-Loop-9B-Preview
Looped multimodal model: 24 decoder layers run twice with shared weights (56 layer passes from 32 layers), trained up to 256K tokens. Apache-2.0.
19 releases latest Nov 16, 2026
-
Qwen4-Max
Official Qwen4 flagship: 2.4T-param MoE (71B active), native vision, 1M context. On QwenCloud API, Token Plan, QwenWork and QwenStudio; smaller open Qwen4 models promised.
-
Qwen4-Max-Preview
First Qwen4 flagship preview: text, image and video input, thinking on by default. Token Plan only until Nov 25, then retired; also in Qwen App and QwenStudio.
-
Qwen-Audio-3.1-Realtime-Flash
Real-time duplex speech model with text and audio in and out, function calling, web search and voice cloning; Flash tier of Qwen-Audio-3.1 Realtime on Alibaba Cloud Model Studio.
-
RynnUnify
DAMO Academy 16B embodied model pairing an 8B reasoner with an 8B video generator for interleaved video-language planning; Agentic Reasoner and World-Action Model checkpoints too.
-
Qwen-Image-2.1-Turbo
8-step accelerated checkpoint of Qwen-Image-2.1 (same 7B architecture); Pro and Turbo APIs live on Alibaba Cloud Model Studio. Qwen Research License.
-
Qwen-Image-2.1
Unified text-to-image and editing model with a 7B generation component and native transparent (RGBA) output. Qwen Research License.
-
Qwen3.8-Flash-Next
Open-weight 125B multimodal MoE (6B active, plus 51B n-gram embeddings) previewing the Qwen4 architecture; served as Qwen3.8-Flash on Qwen Cloud.
-
Qwen3.8-Max
Open weights followed on August 12.
-
Qwen3.7-Max
-
Qwen3.6 (open-weight)
-
Qwen3.6-Plus
-
Qwen3.5
-
Qwen3-Max-Thinking
-
Qwen-Image
-
Qwen3-Coder
-
Qwen3
-
Qwen2.5-Omni
-
QwQ-32B
-
Qwen2.5-Max
2 releases latest Nov 16, 2026
-
CloudRobo-WM1.2
Ascend-native embodied world model that predicts action outcomes and generates robot actions. Weights, training framework and Ascend docs open-sourced; tops RoboTwin 2.0 per Huawei.
-
openPangu-2.1-Pro
Successor to openPangu-2.0-Pro with upgraded agent skills, coding (HarmonyOS ArkTS, Ascend operators) and throughput. Live on Huawei Cloud MaaS; weights not yet on Hugging Face.
5 releases latest Nov 16, 2026
-
d1.2 and d1.2-ultra
Omnimodal decision models adding audio input and five new answer types (values, spans, boxes, estimates, multi-choice); d1.2-ultra tops AA Decision Index v2.1. $0.02/$0.042 per 1M input.
-
LFM3 (LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B and d1-Omni-2B)
Six ~2B open-weight models on one tri-hybrid LFM3-1.7B backbone (text, vision, audio, omni, decision); 22T-token pretraining; text model has 256K context.
-
d1-ultra
Multimodal decision model returning calibrated probabilities for noul/Choice/Score without generating tokens; API-only at $0.042 per 1M input tokens.
-
LFM2.5-24B-A2B
24B MoE reasoning model with ~2B active params built for on-device agents; largest and final LFM2.5 model; base and post-trained weights on Hugging Face.
-
d1-3B and d1-omni-600M
Open edge decision models built on LFM2.5 backbones; d1-3B takes text and images, experimental d1-omni-600M adds audio.
3 releases latest Nov 16, 2026
-
Sage S2.1 Flash
321B scientific multimodal agent model (MIT, BF16 + FP8) that reads molecules, crystals and proteins via Sage SciPerception; 90.3 on MLE-bench Lite.
-
Atria Dawn (open weights)
Full release of the agentic model built on the 744B MoE GLM-5.2; MIT weights plus FP8 now on Hugging Face after an API-first launch. 512K context, up from 256K in Preview.
-
Atria Dawn Preview
744B MoE agentic model built on GLM-5.2, MIT license. Technical report followed on Sep 14.
2 releases latest Nov 13, 2026
-
HeyGen Decision 1 and Decision 1 Turbo
Decision models that answer yes/no, choice and score questions about text, images or video with probabilities, not text. 128K context (Turbo preview: 256K).
-
HeyGen Voice
HeyGen's first in-house voice model; debuted #1 on Artificial Analysis' voice leaderboard. A lower-latency HeyGen Voice 1 Turbo followed in the API.
6 releases latest Nov 13, 2026
-
Duo v1 (Duo-1.0 and Duo-1.2)
Preview full-duplex audiovisual world model from Ant Research: real-time conversation with jointly generated video and audio. Causal, distilled and quantized weights.
-
Ling-3.2-flash
Native multimodal MoE (560B total, 25B active) with image/text/video input; focuses on visual understanding, long-horizon agents, 3D and medical multimodal. Free API for first 2 weeks.
-
Ling-3.1-flash-1019
Upgraded Ling-3.1-flash checkpoint (560B total, 25B active MoE, 1M context) tuned for reasoning and agentic work; AA Intelligence Index 44 vs 41. MIT weights on Hugging Face Oct 27.
-
Ming-3.0-tts (0.5b and tiny) and Ming-3.0-asr-tiny
MIT-licensed Ming-3.0 speech family: dense 0.5b and Ling V3 MoE tiny (7.9B total, 1.3B active) TTS with voice cloning and paralinguistic tags, plus a context-aware Chinese/English ASR model.
-
Ling-3.0-flash-Sante
MIT healthcare research model built on Ling-3.0-flash: 124B total, 5.1B active hybrid-linear MoE trained to 256K tokens. Not a medical device.
-
Ling-3.1-flash
Ant InclusionAI 560B MoE (25B active) native hybrid-reasoning model with a 1M-token context window; MIT-licensed weights published Oct 13, 2026 at huggingface.co/inclusionAI/Ling-3.1-flash.
1 release latest Nov 12, 2026
-
Edge0-145B and Edge0-120B
On-device latent sparse MoE models (145B total, about 10B active) built on Qwen3.8-Flash-Next; 29GB peak memory on a laptop. Edge0 Community Model License.
9 releases latest Nov 12, 2026
-
ERNIE Wizshot Micro-Expression
Expressive talking-character video model that processes text, images and speech together, with scripts up to 4,000 Chinese characters. API on Qianfan; unveiled at Baidu World 2026.
-
ERNIE 5.5 Flash Preview
First post-trained model on ERNIE 5.5 Flash Base: 280B total, 14B active. API on Qianfan; Apache 2.0 weights promised soon.
-
ERNIE-5.5-Flash-Base (open weights)
280B-A14B MoE pre-training checkpoint (33.5T tokens), no post-training; FP8 and BF16 weights. Post-trained ERNIE 5.5 Flash to follow.
-
Apollo World
Agent-driven multimodal world model for Apollo Go robotaxi sim; up to ~100k virtual km/day; unveiled at Baidu AI Day Apollo session in Shenzhen.
-
Unlimited-OCR-2
MIT-licensed document-parsing VLM whose sparse sliding-window attention keeps decoding cost near constant, so it can parse documents hundreds of pages long; on HF, ModelScope and Baidu Cloud.
-
ERNIE 5.1
-
ERNIE 5.0
2.4T-parameter natively omni-modal model.
-
ERNIE 4.5 open-source family
-
ERNIE 4.5 and ERNIE X1
11 releases latest Nov 12, 2026
-
Seedance 2.5 flash
Faster, cheaper Seedance 2.5: about 2x speed, ~34% lower 720p price, clips up to 30s at 480p/720p with up to 10 images, 5 videos and 5 audio refs. Volcano Ark API from Nov 15.
-
Seedream 5.5
Image generation and editing model with pixel-level preservation across 100-round multi-turn edits, native RGBA transparent output, and better small-text rendering; Pro version to follow.
-
Cola2
Open continuous-latent hierarchical LM (~9B) from ByteDance Seed; 346B source tokens, 60.10 eight-task avg (paper; DiT unconverged).
-
OmniHuman 2.0
Audio-driven talking-avatar video generation from a single photo, offered through the BytePlus API.
-
Seedream 5.0 Pro
-
Seed 2.1 (incl. Seed 2.1 Pro)
Agent-focused successor to Seed 2.0 with stronger coding, computer use and video understanding; in Doubao and Volcano Engine.
-
Seed 2.0 (Doubao-Seed-2.0 Pro, Lite, Mini, Code)
-
Seedream 5.0 Lite
-
Seedance 2.0
Unified audio-video generation.
-
Seedream 4.0
-
Seed-OSS-36B
5 releases latest Nov 12, 2026
-
Granite 4.1 LLMON Preview (3B and 8B)
Experimental Granite 4.1 variants trained on LLMON markup that separates instructions from data; IBM reports near-zero prompt-injection success on OPI.
-
Granite Switch 4.1 (8B and 30B)
Composes a Granite 4.1 base with Decision and Guardian aLoRA adapters and a speech encoder behind one endpoint; Apache 2.0.
-
Granite 4.2
-
Granite 4.1
-
Granite 4.0
13 releases latest Nov 12, 2026
-
Nemotron-Labs-Sigma (3B and 8B)
Research continuous diffusion language models that denoise text in a 16-D embedding space and decode in parallel blocks; 14 checkpoints (base, instruct, distilled). Nemotron Open Model License.
-
Cosmos 3 Nano (Transfer-Auto, Sim-Auto, Sim-Transfer and Sim-Bimanual)
Four 16B open-weight Cosmos 3 Nano checkpoints for interactive world simulation (multi-view driving, LiDAR, simulator-to-RGB, bimanual robot); OpenMDW1.1. HF collection also dated Oct 20.
-
Nemotron 3.5 Transcribe
Promptable open speech model (30B MoE, ~4B active): transcription in 40 languages, translation, speaker labels and audio Q&A; first among open models on Open ASR (English).
-
Nemotron 3.5 Bolt E4B 4GB
Open small model for local conversational agents, built on Gemma 4 E4B with Nemotron post-training; runs in about 4GB (Q4_K_M), 35 languages, configurable reasoning.
-
Nemotron 3.5 Embed (8B and 1B), Rerank 8B and Parse
Open Nemotron Retriever models for multimodal (text and image) embedding, reranking and 631M document parsing; Rerank cuts agent token cost about 14%. On Hugging Face, NIM and OpenRouter.
-
Kumo Relational
Pre-trained relational foundation model that predicts churn, demand or fraud on enterprise databases in-context, no training. Beats supervised baselines on 41 tasks; on Hugging Face and NIM.
-
Nemotron Labs Dory1.0-4B
Experimental looped (latent-recurrent) 4B model: a shared block is re-applied to trade inference compute for accuracy from one checkpoint. OpenMDW-1.1 license; BF16, MXFP8, NVFP4 and GGUF.
-
Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B 24GB
Super VL: open 120B-A12B vision-language model, 1M context. Bolt: 74B-A9B local agent model in 24GB via native 2-bit. Bolt E4B 4GB due Oct 23.
-
Kumo Tabular (Small, Medium and Large)
Open tabular foundation models (28M–215M) for in-context classification/regression; OpenMDW-1.1; #1 on TabArena (ELO 1950).
-
Nemotron 3.5 Lightning (30B-A3B)
Open 30B MoE with 3B active, built for high-volume low-latency agents; MTP speculative decoding, NVFP4 and BF16 checkpoints.
-
Nemotron 3 Ultra
-
Nemotron 3 Super
-
Nemotron 3 Nano
4 releases latest Nov 12, 2026
-
pplx-decider-v1.3 and pplx-decider-flash-v1.3
Open-weight decision models on Qwen3.8-27B and Qwen3.5-4B (Flash). Adds Locate, which extracts verbatim text spans with character offsets. Apache 2.0.
-
pplx-decider-v1.2 and pplx-decider-flash-v1.2
Open-weight multimodal decision models: 27B (Qwen3.8-27B, Decision Index 68.09) and 4B Flash (Qwen3.5-4B); $0.02 and $0.01/M input tokens. Apache 2.0.
-
pplx-embed-v2-late (0.6B and 9B)
MIT-licensed late-interaction multimodal retrieval models for text, images and pages, sharing one embedding space across sizes.
-
pplx-decider-v1.1-27b
Open-weight multimodal decision model (Qwen3.8-27B backbone); Decision Index 61.56, ahead of Jev; $0.02/M input tokens.
1 release latest Nov 12, 2026
-
SOV-Transcribe-1
Open-weight speech-to-text (26.6B total, 4.6B active) on a Whisper large-v3 encoder and Gemma 4 26B-A4B. 7 languages, up to 30 min audio, instruction-controlled style. Apache 2.0.
2 releases latest Nov 12, 2026
-
Bielik-One (11B v3.1, 11B v3.0, 7B v3.0 and 4B)
Four Bielik decision models that return calibrated probabilities for choice, yes/no, score and multi-label questions instead of generating text. Shipped as merged weights and LoRA adapters for vLLM.
-
Bielik 11B v3.1
Polish open model, now 56 languages (up from 32) after 500B more tokens, with function calling, structured output and reasoning mode. Apache 2.0, with ACK Cyfronet AGH.
1 release latest Nov 11, 2026
-
Praxis-1
27B MoE diffusion world action model for robots, predicting video and actions together. Tops NVIDIA RoboLab-120 at 46.2%. Apache 2.0 weights plus fine-tuning code.
1 release latest Nov 11, 2026
-
BotPsy
0.8B open-weight on-device agent model (Apache 2.0), built on Qwen3.5-0.8B for tool calling.
4 releases latest Nov 10, 2026
-
Scribe v3 and Scribe v3 Live
Speech-to-text: v3 has 27% fewer word errors than Scribe v2 across 102 languages; v3 Live is the realtime variant with diarization and smarter turn detection.
-
SFX v3
Sound effects model that follows multi-sound prompts more reliably and handles music and vocals better than SFX v2. Limited access in the API.
-
Eleven Realtime v1 Mini and v1 Max
First ElevenLabs speech-to-speech voice-agent models; Max reasons before responding (98.8% Big Bench Audio), Mini is the low-cost default. 90+ languages, OpenAI Realtime API compatible.
-
Eleven v4 and Eleven v4 Turbo
Most expressive ElevenLabs TTS family; Turbo ~100ms median latency for agents; 90+ languages; Artificial Analysis #1 voice.
2 releases latest Nov 10, 2026
-
Live 3
Full-duplex voice model for agents that listens while speaking, learns turn-taking from real calls and can react in 186 ms. English only.
-
Fish Audio Drama 3
Directable, emotive text-to-speech in 83 languages; 35B MoE (~3B active); weights under Fish Audio Community License v1.0.
2 releases latest Nov 10, 2026
-
PAN 2
Open world model: an LLM plans events from a scene plus an action, a from-scratch video diffusion model renders them. Apache 2.0 weights and training code.
-
Jais 3 and Jais Speech
Arabic-native fleet of six open models (0.9B to 375B) built on K2 Horizon, plus Jais Speech ASR; with Inception42 (G42).
14 releases latest Nov 10, 2026
-
MiniMax H3 Image
Open-weight 33.1B DiT for text-to-image and editing in one checkpoint; native ~16 MP output, up to 16 reference images, 6-step sampling. MiniMax-H3-Image community license.
-
MiniMax M3.1 Flash (open weights)
Weights for the Oct 19 M3.1 Flash: ~428B-total / 23B-active multimodal MoE (text, image, video in), 1M context, MiniMax community license.
-
MiniMax M3.1 Flash
Multimodal Flash model with tunable thinking depth; $0.15/$0.60 per 1M tokens. Weights followed on Oct 23 (see the separate open-weights entry).
-
MiniMax H3
Open multimodal video model.
-
MiniMax Music-3.0
Open weights (~2B) published Aug 7, 2026 at huggingface.co/MiniMaxAI/MiniMax-Music3.
-
MiniMax-M3
-
MiniMax-M2.7
-
MiniMax-M2.5
-
MiniMax-M2.1
-
MiniMax Hailuo 2.3 and 2.3-Fast
-
MiniMax-M2
-
MiniMax Hailuo 02
1080p, up to 10-second clips.
-
MiniMax-M1
-
MiniMax-Text-01 and MiniMax-VL-01
18 releases latest Nov 9, 2026
-
Claude Fable 5.2 and Claude Mythos 5.2
Same model; Fable is generally available with cyber/bio safeguards, Mythos for vetted orgs. Up to 30% faster than 5.1 at $8/$40 per 1M tokens; FrontierCode 2.0 49.1%.
-
Claude Haiku 5.5
Anthropic’s cheapest, fastest small model; first Haiku with an adjustable effort setting, about 75% cheaper to run than Haiku 4.5.
-
Claude Sonnet 5.5
-
Claude Opus 5.5
-
Claude Fable 5.1 and Claude Mythos 5.1
-
Claude Opus 5
$5 / $25 per million input / output tokens.
-
Claude Sonnet 5
-
Claude Fable 5 and Claude Mythos 5
-
Claude Opus 4.8
-
Claude Opus 4.7
-
Claude Sonnet 4.6
-
Claude Opus 4.6
-
Claude Opus 4.5
-
Claude Haiku 4.5
-
Claude Sonnet 4.5
-
Claude Opus 4.1
-
Claude Opus 4 and Claude Sonnet 4
-
Claude 3.7 Sonnet
First hybrid reasoning Claude.
5 releases latest Nov 9, 2026
-
K-EXAONE 2.5
Further-trained K-EXAONE 2.0 validation model, free API via docs.exaone.ai; SWE-Bench Pro 59.2, BrowseComp 78.0, Claw-Eval 77.3 (LG-reported).
-
K-EXAONE 2.0
750B total / 37B active MoE.
-
EXAONE 4.5
33B vision-language model.
-
EXAONE 4.0
-
EXAONE Deep
10 releases latest Nov 9, 2026
-
Muse Ember
Meta's smallest, cheapest Muse model for high-volume agent subtasks; natively takes text, images, audio and video. $0.10/$0.50 per 1M tokens, up to 512K output; Artificial Analysis Index 36.
-
Muse Spark Open
Most capable US open-weight model to date; 816B total / 42B active MoE, natively multimodal, 1M context; weights on Hugging Face; runs on two Apple M5 Ultra Mac Studios via MLX.
-
Muse Wonder and Muse Wonder Full
Meta's most capable models, built for agentic tasks with Active Memory and Active Focus. Wonder on Meta Model API and Muse Code; Wonder Full for Muse Code subscribers.
-
Muse Spark 1.3
Improved agentic and coding vs 1.2; max reasoning on Muse Code and Meta Model API.
-
Muse Glimmer
30B open-weight model for local use.
-
Muse Spark 1.2
Released alongside the Muse Code agent.
-
Muse Spark 1.1
-
Muse Image
Meta Superintelligence Labs' first image-generation model, in Meta AI, Instagram and WhatsApp.
-
Muse Spark
First model from Meta Superintelligence Labs.
-
Llama 4 Scout and Llama 4 Maverick
1 release latest Nov 9, 2026
-
ZONOS3-Mini
Apache-2.0 zero-shot voice-cloning TTS: sparse MoE with 100M active / 800M total params, real-time on-device (WebGPU, GGML), with accent, age, emotion and pace controls.
1 release latest Nov 6, 2026
-
Kanana-2-v-decide
Open-weight 3.67B multimodal decision model (choice/noul/score, no text generation); Kanana Open License. Press release Nov 6.
1 release latest Nov 6, 2026
-
Arete-27B
Open-weight SFT of Qwen3.8-27B to reduce CCP-framed answers; Hermes Bench ~65.6%, Terminal-Bench-2.1 70%.
1 release latest Nov 5, 2026
-
PLaMo 3 Translate (0.8B, 2B, 8B and 31B)
Open-weight translation LLMs for 53 languages that run offline on phones, laptops and servers; built on PLaMo 3, with glossary and context control. PLaMo Community License.
1 release latest Nov 4, 2026
-
SWE-2 Flash
Coding model mid-trained and RL-tuned on Thinking Machines Inkling-Small (276B-A12B). In Devin Desktop and CLI at one-fifth the price and 3x the speed of SWE-2; weights gated.
1 release latest Nov 4, 2026
-
Lion-ER-35B-A3B (and Lion-ER-4B, Lion-ER-9B)
Open Apache-2.0 embodied-reasoning MoE (~35B/3B active) fine-tuned from Qwen3.6-35B-A3B; avg@4 65.29 on ten visual/embodied benches.
1 release latest Nov 4, 2026
-
Ripple-65B Preview
Dense 65B model pretrained from scratch on 12T tokens, Apache-2.0 with BF16 and NVFP4 builds. Reports 80.52 on TerminalBench 2.1; also on Together AI, GMI Cloud and OpenRouter.
1 release latest Nov 4, 2026
-
GroundingPI
Open 4B visual grounding foundation model (points/boxes); 73.68 avg across 34 perception benches vs GPT-6 Astra 71.54 (team eval).
1 release latest Nov 3, 2026
-
Endeavor 1.1
Frontier multimodal (image+text) with EU inference option and inspectable thinking traces; access by request, managed or private deploy.
2 releases latest Nov 3, 2026
-
Kandinsky Avatar 1.0 (S2V 1.0 Lite and Pro)
Open speech-to-video avatar models (2B Lite, 21B Pro, plus distilled variants): talking video from one image, a speech track and a prompt. MIT.
-
Kandinsky 6.0 Video (Lite and Pro)
Open-sourced 3B Lite and 29B Pro text/image-to-audio-video diffusion models: 5s clips with synced 44 kHz audio and lip-sync, upscaling to 1080p. MIT.
4 releases latest Nov 3, 2026
-
Sarvam Prism
Decision model (LoRA fine-tune of Sarvam 105B) returning calibrated option probabilities in one pass; 44 ms median on one B200. Apache 2.0 weights.
-
Bulbul V4 Flash
Expressive text-to-speech for Indian languages built for real-time voice agents; under 250 ms time-to-first-audio.
-
Sarvam 30B and Sarvam 105B
Weights open-sourced March 6.
-
Sarvam-M
Built on Mistral Small.
1 release latest Nov 2, 2026
-
Hydrogen-1.0
Open-weight science model for chemistry, materials and life sciences, post-trained from GLM-5.3 (~753B MoE). 1M context, MIT license, NVFP4/MXFP4 builds; API via NeuroSpark.
3 releases latest Nov 2, 2026
-
LongCat-2.5
Native multimodal MoE for long-horizon agentic coding: 1.6T total, ~48B active. On longcat.ai and API; weights (plus W4A4 quant) released Nov 9 on Hugging Face under MIT.
-
LongCat-2.0
Weights open-sourced July 6.
-
LongCat-Flash-Chat
1 release latest Nov 2, 2026
-
Drex for Finance
Finance-specialized Drex decision model (choice/noul/score, no free text); 77.3 on Drex Finance Index vs 63.6 for Jev 1.14.
2 releases latest Oct 31, 2026
-
MiniCPM-OCR 1.0 and MiniCPM-OCR Jev
First OCR models in the MiniCPM family: a 1.3B edge document parser and a 2B Jev model that scores OCR output quality. Open weights on GitHub, Hugging Face and ModelScope.
-
MiniCPM-V 4.7 (35B-A3B and 1B)
Open-weight on-device vision-language models; 35B-A3B is the series first MoE, running at 13+ tok/s in 3 GB on iPhone 17 Pro Max. Apache 2.0.
2 releases latest Oct 29, 2026
-
Canva Brush
Canva Research vision-language model for design editing; powers Canva AI edits for Pro users, up to 69% faster and 46% cheaper per edit than the previous design agent.
-
Canva Craft
Canva Research's first in-house image model: plans a layout, then renders up to 2K in 30+ aspect ratios, with text in 55 languages. In Dream Lab and the public API.
23 releases latest Oct 29, 2026
-
Mistral Large 4 (open weights)
Improved Large 4 (Le Chonk) with weights released: ~1T-parameter multimodal MoE (~52B active), three reasoning modes, Mistral Ecosystem Open-Weight license.
-
Voxtral 3 (Small Transcribe 3, Mini Transcribe 3, Mini Transcribe Realtime 3 and Speech 3)
New speech-to-text (50+ languages, diarization, context biasing) and TTS (35+ languages) family plus a voice agent SDK; Mini Transcribe 3 is 4B, $0.0035/min.
-
Timestral (5M, 28M, 123M and 209M)
Time-series foundation models for zero-shot probabilistic forecasting; up to 32,768-step context and 1,024-step horizon.
-
Voxtral Mini 4B Realtime Arabic
Streaming Arabic speech-to-text for MSA and dialects; fine-tuned from Voxtral-Mini-4B-Realtime-2602. Apache 2.0.
-
Mistral Large 4 (preview)
1T-parameter natively multimodal model (49B active), Mistral's largest yet; public preview API on Mistral Studio, weights promised by end of October.
-
Mistral OCR 4.1
-
Mistral Medium 3.5
Open weights under a Modified MIT license.
-
Voxtral TTS
-
Mistral Small 4
Unifies instruct, reasoning and coding.
-
Voxtral Mini Transcribe 2 and Transcribe Realtime
-
Devstral 2 and Devstral Small 2
-
Mistral Large 3 and Ministral 3 (3B, 8B, 14B)
Apache 2.0.
-
Magistral Medium 1.2 and Small 1.2
-
Mistral Medium 3.1
-
Voxtral Small and Voxtral Mini
First Mistral audio models.
-
Devstral Medium and Devstral Small 1.1
-
Magistral Medium and Magistral Small
First Mistral reasoning models (Small is open-weight).
-
Devstral Small
-
Mistral Medium 3
-
Mistral Small 3.1
-
Mistral OCR
-
Mistral Saba
Arabic and South Asian languages.
-
Mistral Small 3
1 release latest Oct 29, 2026
-
Pika H3 Flex
Post-trained MiniMax H3 video model: 6-step generation with synced audio at 480p/768p, up to 12 references, 4K upscaling and audio stems via API. Open weights.
1 release latest Oct 28, 2026
-
Clover-35B-A3B-Preview
Apache-licensed block diffusion LLM (35B MoE, ~3B active) built from Qwen3.6-35B-A3B; self-reported 73.6 SWE-bench Verified and 680 tok/s on one GB300.
1 release latest Oct 28, 2026
-
Prism
Enterprise speech-to-speech model with an inspectable text checkpoint before audio; 0.57s to first audio and 56.6 on AA τ-Voice (self-run); fine-tunable.
1 release latest Oct 24, 2026
-
Nanbeige4.3-33B
Apache-2.0 dense 33B model for agentic coding and cowork tasks, released with a Base version; the teacher model for the Nanbeige4.3 series. Self-reported 80.9 on Terminal-Bench 2.1.
7 releases latest Oct 23, 2026
-
Kimi K3.1 Pro and K3.1 Pro Highspeed
Post-trained Kimi K3 (2.8T MoE, 104B active) with gains in coding, 3D creation and finance; Highspeed runs 2-3x faster. Full weights released Nov 4 under the Kimi K3.1 Pro License.
-
Kimi K3
Full weights released July 27 under the Kimi K3 License.
-
Kimi K2.7 Code
-
Kimi K2.6
-
Kimi K2.5
-
Kimi K2 Thinking
-
Kimi K2
1 release latest Oct 22, 2026
-
Carbon-X
Open 14.8B Apache-2.0 model that reads DNA and answers in English (variant effects, regulatory regions); built on Qwen3-14B, runs on one GPU.
2 releases latest Oct 22, 2026
-
Spark X2.6
293B-A30B MoE for coding and agents, launched at iFlytek's 1024 festival; trained on domestic Chinese compute. Flagship is closed; sibling Spark-X2.6-4B is Apache-2.0 on Hugging Face.
-
iFlytek Spark X2
Trained entirely on domestic Chinese compute.
7 releases latest Oct 21, 2026
-
Rerank 5 Pro and Rerank 5 Fast
Enterprise reranking models (~2B Pro, ~500M Fast) with new Context Focus passage extraction and calibrated relevance scores; 100+ languages. Fast beats Rerank 4 Pro.
-
Tiny Aya L2-Thinker
3.35B open research model that reasons in the prompt language across 60 languages; CC-BY-NC with Cohere Labs AUP.
-
Embed 5 Pro and Embed 5 Fast
Multimodal, 100+ language embedding family sharing one embedding space; Pro for quality, Fast for latency.
-
Command A+
218B total / 25B active MoE, Apache 2.0.
-
Command A Reasoning
-
Command A Vision
-
Command A
1 release latest Oct 21, 2026
-
Flint (40B-A2.5B)
Recursive's first model: Apache-2.0 MoE coding LLM, 39.8B total/2.5B active, 262K context, runs on laptops. Reports 74.2% SWE-bench Verified; built largely by its Eureka Machine AI researcher.
1 release latest Oct 20, 2026
-
Fibo 2.0 (Dev and Edit Turbo)
8B open-weight model unifying text-to-image and image editing, trained on 100% licensed data; #1 open-weights image editing model on Artificial Analysis. Gated weights on Hugging Face.
1 release latest Oct 20, 2026
-
Sugar Glider 120B
120B model on the Cerbearus constant-memory architecture (O(1) memory per token); 73.3 on 13-task RULER at 1M tokens. i14 cloud, VPC or on-prem via waitlist.
1 release latest Oct 20, 2026
-
FastH3-Omni
Open-weight accelerated MiniMax H3 built with FastVideo (UCSD Hao AI Lab); image/video/audio references to 5-15 s video with synced audio, 5.6-11.9x faster.
1 release latest Oct 20, 2026
-
Ornith-1.5-122B
MIT-licensed 122B MoE coding and agentic model that serves on a single 80GB GPU; 84.5 SWE-bench Verified, 81.2 Terminal-Bench 2.1. FP8, NVFP4, GGUF and MLX builds.
1 release latest Oct 19, 2026
-
FLUX 3 Max
Text/image-to-video model post-trained by fal Research with Black Forest Labs on FLUX 3; 1,139 Elo in the preliminary Artificial Analysis image-to-video arena, ~11s per 10s clip.
1 release latest Oct 19, 2026
-
Mercury 3
Diffusion LLM, 40x larger than Mercury 1, running at 1,000+ tokens/s; $0.20/$0.70 per 1M tokens (50% off at launch).
13 releases latest Oct 17, 2026
-
DeepSeek-V4.1-Flash-1017
Update to V4.1-Flash with faster task completion and shorter outputs; DeepSWE v1.1 77.0, Terminal-Bench 4.0 42.3. MIT-licensed weights on Hugging Face.
-
DeepSeek-V4.1-Flash
Natively multimodal; first model in a new architecture family.
-
DeepSeek-V4-Flash-Vision-Exp
Experimental vision model; weights followed on Hugging Face.
-
DeepSeek-V4-Pro (GA)
-
DeepSeek-V4-Flash (official)
Public beta of the re-post-trained V4-Flash.
-
DeepSeek-V4 Preview (V4-Pro and V4-Flash)
-
DeepSeek-V3.2 and V3.2-Speciale
-
DeepSeek-V3.2-Exp
Introduced DeepSeek Sparse Attention.
-
DeepSeek-V3.1-Terminus
-
DeepSeek-V3.1
Hybrid thinking / non-thinking model.
-
DeepSeek-R1-0528
-
DeepSeek-V3-0324
-
DeepSeek-R1
1 release latest Oct 16, 2026
-
Glimpse
Research-preview visual thinking model that revisits images and video while reasoning (zoom, frame extraction, code tools). $0.60/$3.00 per million input/output tokens.
1 release latest Oct 16, 2026
-
MOSS-Image (Base and SFT)
5B text-to-image DiT with JSON layout control, strong text rendering and prompts up to 4K tokens; Apache-2.0 weights.
2 releases latest Oct 16, 2026
-
Beam 1.0 (open weights)
Apache 2.0 weights for the 501B-total / 23B-active MoE in BF16, FP8 and NVFP4, plus a speculative-decoding drafter; public-beta API.
-
Beam
501B-total / 23B-active MoE for coding, reasoning and agents; launched as an early-access preview. Apache 2.0 weights released Oct 16.
6 releases latest Oct 15, 2026
-
FLUX 3 Image Klein
5.6B open-weights text-to-image and editing model with up to 10 references and native 4K; base and step-distilled checkpoints, FLUX Kommunity License.
-
FLUX 3 Image
Image side of FLUX 3: text-to-image, pixel-exact local edits, up to 10 references, bounding-box layout, up to 4K.
-
FLUX 3 Action
Open-weight 7B world-action (robotics) model.
-
FLUX 3
Unified video, image and audio model.
-
FLUX.2
-
FLUX.1 Kontext
1 release latest Oct 15, 2026
-
Krea 3
Image editing model: most edits under 8 seconds at native 2K, up to 6 references, $0.06 per 2K image; also on fal, Replicate, Together AI and OpenRouter.
1 release latest Oct 15, 2026
-
Comet 60B
CompactifAI-compressed Nemotron 3.5 Super VL: about 60B total / 12B active, text, image and video input; BF16, FP8 and NVFP4.
1 release latest Oct 15, 2026
-
Matrix-Game 4.0
Open interactive world model: a game engine plus a causal video model doing streaming neural rendering; three playable games, Apache-2.0.
1 release latest Oct 14, 2026
-
UltraBERT (150M and 380M) and UltraBERT-Embed-150M
Apache 2.0 English encoders with native 32K context for classification and retrieval, plus a 150M embedding model scoring 67.24 on MTEB(eng, v2).
1 release latest Oct 14, 2026
-
Gnani Timbre v3.0
Expressive text-to-speech for Indian languages with audio tags; launches with 13 voices, 13 tags and 10 languages.
1 release latest Oct 14, 2026
-
Mirelo SFX 2.0
Video-to-sound model that keeps dialogue and generates synced sound effects, foley and ambience for videos up to 10 minutes.
1 release latest Oct 14, 2026
-
Ultramarine 1 Flash
Agentic knowledge-work model post-trained with non-verifiable RL on GLM-5.3 Flash; API live; open weights promised.
15 releases latest Oct 14, 2026
-
Grok Voice 3.0
Speech-to-speech model; #1 on Artificial Analysis Speech-to-Speech Index (86.9). Voice Agent API at $0.08/min; grok-voice-latest now points to it.
-
Grok 4.7
-
Grok Voice Transcribe 2.0
Speech-to-text model.
-
Grok 4.6
-
Imagine Image 2.0
-
Grok Voice Think Fast 2.0
Speech-to-speech voice model.
-
Grok 4.5
SpaceXAI's own announcement post is dated July 16.
-
Grok Imagine Video 1.5
-
Grok Imagine (API)
-
Grok 4.1 Fast
-
Grok 4.1
-
Grok 4 Fast
-
Grok Code Fast 1
-
Grok 4
-
Grok 3
1 release latest Oct 14, 2026
-
kettei (fat, chub, core and dot)
Decision model family; fat/chub multimodal via OneHot API; core (3.7B) and dot (900M) open-source text-only weights.
2 releases latest Oct 12, 2026
-
ColGrep-Agent-2B
Local 2B code-search agent that coding agents delegate repo search to; fine-tuned from MiniCPM5-2B; Apache 2.0.
-
LightOnOCR-3 (0.8B, 1B and 4B)
End-to-end document OCR family with grounded layout, image descriptions and chart data extraction; Apache 2.0.
1 release latest Oct 12, 2026
-
Helix World
Real-time audio-visual world model with HKUST: video plus camera-aligned stereo at 24 fps on one H800. Built on LTX-2.3; Apache-2.0 preview weights.
9 releases latest Oct 12, 2026
-
GLM-Legal
Legal-domain agentic model built on GLM-5.3-Flash; 19.2% all-pass on Harvey LAB. Beta with application-gated API access; coming to AutoClaw.
-
GLM-5.3-Flash
First natively multimodal GLM-5 model: 320B total, 18B active, hybrid linear and sparse attention; MIT license.
-
GLM-5.3
Open weights followed on August 28.
-
GLM-5.2
-
GLM-5.1
-
GLM-5
-
GLM-4.7
-
GLM-4.6
-
GLM-4.5
1 release latest Oct 11, 2026
-
Humaneness Voice Converter
Open Chatterbox-based speech-to-speech voice conversion; preserves timing/emotion from a 5–10s target clip; CC BY 4.0 + MIT upstream.
1 release latest Oct 10, 2026
-
THX-01
322M open Apache-2.0 multilingual decision model (mmBERT); one forward pass ~10 ms; HF doofz/THX-01.
1 release latest Oct 10, 2026
-
Decision 3.0 (d3, d3-flash, d3-mini, d3-nano, d3-lite and d3-edge)
Open Apache-2.0 multimodal decision family (0.6B–27B) reading text, images and video; System One probabilities, no text generation.
1 release latest Oct 8, 2026
-
Mellum2.1
12B MoE (2.5B active) Apache-2.0 coding-agent model with RL in real repos; HF JetBrains/Mellum2.1-12B-A2.5B-Thinking.
1 release latest Oct 8, 2026
-
Odyssey-3
Foundation world model (Odyssey-3 Pro + Flash); public research preview for real-time interactive environments; SOTA Physics-IQ Verified 66.1 (Pro).
1 release latest Oct 8, 2026
-
Darwin-27B-ZTC
Apache-2.0 zero-token 27B decision/judge model; one forward pass per typed question; #1 on typed-decisions zero-shot (0.743).
1 release latest Oct 7, 2026
-
Vidu Q4 Preview
Next-gen Vidu video model with native audio, reference voices, up to 15 reference images and 4K 10-bit output.
4 releases latest Oct 6, 2026
-
Falcon-Emirati, Falcon-ASR and Falcon-OCR-Arabic
7B Emirati-dialect LLM on Falcon-H1-Arabic (Falcon Chat), plus a 1.6B multilingual ASR model and a 270M Arabic OCR model.
-
Falcon Perception and Falcon OCR
-
Falcon H1R 7B
-
Falcon-H1
Hybrid Transformer-Mamba family.
5 releases latest Oct 5, 2026
-
Amazon Nova 2.5 Sonic
Speech-to-speech model for real-time voice agents with improved reasoning, instruction following and tool calling; GA on Amazon Bedrock.
-
Amazon Nova 2 Sonic
-
Amazon Nova 2 Lite and Nova 2 Pro (preview)
-
Amazon Nova Premier
-
Amazon Nova Sonic
3 releases latest Oct 5, 2026
-
Rho-1
19B omni-reasoning research preview trained from scratch; understands and generates text, images and video, and emits robot actions in one model.
-
Reka Flash 3.1
-
Reka Flash 3
1 release latest Oct 3, 2026
-
Kolibri-1
English-German MoE, Apache 2.0.
2 releases latest Sep 22, 2026
-
Solar Mini 4
35B MoE (3B active) cost-efficient agent model with Korean/English/Japanese, reasoning mode, and 512K context.
-
Solar Pro 4
Upstage flagship for agentic use, strongest at documents, coding and business tasks. 512K context, 128K output, reasoning mode; params undisclosed. Replaces Solar Pro 3.
3 releases latest Sep 22, 2026
-
MiMo-V2.6-Pro and MiMo-V2.6-Flash
Native omni-modal MoE; Pro is 1.02T total / 42B active. MIT-licensed weights, plus a Pro-UltraSpeed API variant.
-
MiMo-V2.5-Pro
-
MiMo-V2-Pro
Previewed on OpenRouter as "Hunter Alpha".
4 releases latest Sep 20, 2026
-
Step 5 Preview
600B MoE (27B active) flagship for agentic coding and knowledge work with vision and 1M context; open weights planned Oct 15.
-
Step 3.7 Flash
-
Step 3.5 Flash
-
Step 3
1 release latest Sep 17, 2026
-
Xing4.0-29B-A4B
Successor to TeleChat: 29B-total/4B-active MoE agentic coding model trained fully on Ascend NPUs; 256K context (512K extensible). Apache-2.0 weights on Hugging Face.
2 releases latest Sep 11, 2026
-
Fugu Max and Fugu Ultra v2
-
Sakana Fugu and Fugu Ultra
Multi-model orchestration system exposed as one API.
2 releases latest Jun 8, 2026
-
Apple Foundation Models 3 (AFM 3 Core and Cloud)
Announced at WWDC26.
-
Apple Foundation Models 2025 (on-device and server)
3 releases latest May 20, 2026
-
Stable Audio 3.0
-
Stable Audio 2.5
-
Stable Video 4D 2.0
3 releases latest Jan 8, 2026
-
Jamba2 (3B and Mini)
-
Jamba Reasoning 3B
-
Jamba 1.6
0 releases
No notable model releases tracked since January 2025.