Radar / Categories / AI models

New AI models, ranked by traction

New frontier, open-weights and on-device AI models. Every launch from the last few weeks, ranked by JEV traction and GitHub stars. Updated four times a day.

Open weights · 55

Soup: fine-tune an 8B LLM on a 4 GB laptop GPU from one YAML fileAn open-source CLI that fine-tunes LLMs from a single YAML file, with layer streaming that trains an 8B model on a 4…models · JEV 0.73Tencent Hunyuan Hy3: open 295B MoE with 21B active under Apache 2.0Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model with only 21B parameters active per token, under…models · JEV 0.73NVIDIA Nemotron 3.5 Retriever collection: Embed 8B/1B, Rerank 8B, Parse 631MNVIDIA's Oct 21 collection drop for agentic retrieval: Nemotron-3.5-Embed in 8B BF16, 1B BF16 and 1B NVFP4…models · JEV 0.72Kimi K3.1 Pro: Moonshot's 2.8T open-weight flagship officially launchesMoonshot AI officially launched Kimi K3.1 Pro on Oct 23, a mixture-of-experts model with 2.8 trillion total / 104…models · JEV 0.72NVIDIA Cosmos 3.5 world models: now openly availableNVIDIA's Cosmos 3.5 family of open world foundation models for robots, autonomous vehicles and vision agents is now…models · JEV 0.71Kimi K3.1 Pro open weights: Moonshot's 2.8T model lands on Hugging FaceMoonshot published the open weights for Kimi K3.1 Pro on November 4 as promised at launch: a natively multimodal…models · JEV 0.69Granite Speech 5.0: IBM's open Apache-2.0 speech models for transcription, translation, and speaker trackingIBM released granite-speech-5.0-2b and updated granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus…models · JEV 0.68Mistral Large 4: open weights for a 1T multimodal flagship built for coding and cybersecurityMistral released downloadable weights for Large 4 (Le Chonk), a ~1T-parameter MoE with 49B active parameters, 1M-token…models · JEV 0.67MiniMax M3: open weights land for the 428B multimodal model with 1M contextMiniMax published the open weights of M3, a 428B-total / 23B-active natively multimodal MoE with a 1M-token context…models · JEV 0.67ERNIE 5.5 Flash Base: Baidu's 280B open checkpoint for further trainingBaidu released the weights of ERNIE-5.5-Flash-Base, a 280B-total / 14B-active sparse MoE base model pre-trained on…models · JEV 0.66ERNIE 5.5 Flash Preview: Baidu's 280B post-trained model, weights open by 25 NovBaidu's first post-trained version of ERNIE 5.5 Flash — a 280B-total / 14B-active sparse MoE with 1M-token context…models · JEV 0.65NVIDIA Nemotron 3.5 Transcribe: open-weight speech model that transcribes, translates and attributes speakers in one passOpen-weight 30B MoE speech model (about 4B parameters active per token) covering 40 languages, released with Granary…models · JEV 0.65Next-Concept Modeling: parallel coarse-to-fine text generationStanford Hazy Research's two-stage framework compresses text into a hierarchy of discrete codes and generates each…models · JEV 0.63Xiaomi-OCR-0: a 0.8B open document model that parses pages into Markdown, tables, and JSONXiaomi's SeerRay team unveiled Xiaomi-OCR-0 on Oct 30: a unified 0.8B vision-language model built on Qwen3.5-0.8B-Base…models · JEV 0.63Tencent Hy4: 770B flagship that sees and operates a computer for the first timeTencent's Hunyuan team released Hy4, a 770B mixture-of-experts model with first-time vision and computer-use…models · JEV 0.63Qwen3-8B GGUF: Alibaba's official ready-to-run local builds with dual reasoning modesQwen's team published official GGUF quantizations of Qwen3-8B (Q4_K_M through Q8_0) for llama.cpp, Ollama, and LM…models · JEV 0.61DeepSeek V3 in ~200 lines of pure JAXThe full DeepSeek-V3 architecture — MLA, DeepSeekMoE, aux-loss-free balancing, multi-token prediction — in 190 lines…models · JEV 0.6Nemotron-Labs-Sigma: NVIDIA's open diffusion language model that writes text in blocksNVIDIA's 8B continuous diffusion language model denoises text in parallel token blocks instead of generating one token…models · JEV 0.6Inkling 2 Small weights: Thinking Machines' 594B MoE goes open under Apache 2.0Thinking Machines Lab published the full weights of Inkling 2 Small on Hugging Face on 12 Nov: a 594B-parameter MoE…models · JEV 0.6Qwen3.8-27B-DSpine-b16: Tencent's parallel draft model for faster Qwen3.8 servingTencent's 2B-parameter draft model for Qwen3.8-27B that guesses 16 tokens at once so the big model verifies them in a…models · JEV 0.58ZONOS3-Mini: Zyphra's open voice-cloning TTS that runs in a browserAn 800M-parameter sparse mixture-of-experts text-to-speech model with only ~100M active parameters per step, doing…models · JEV 0.57MiniCPM-OCR: open document-parsing models with a built-in correctness checkModelBest and OpenBMB's open-source OCR series: MiniCPM-OCR 1.0 (1.3B) for fast document parsing and MiniCPM-OCR Jev…models · JEV 0.55DeepSeek V4.1 Flash: 552B MoE that outruns V4 Pro at flash pricesDeepSeek's Causal Encoder-Decoder flagship — 552B backbone with only 8B active on input and 16B on output, native…models · JEV 0.55Sage S2.1 Flash: Shanghai AI Lab's 320B science model that reads molecules and spectraShanghai AI Lab quietly shipped a 320B/18B-active MoE science model under MIT license that takes molecules, protein…models · JEV 0.55

Frontier · 39

NVIDIA RFD4-Proteina: open protein-design BioNeMo modelOpen protein-design model from NVIDIA and the Baker Lab at UW Medicine's Institute for Protein Design, announced at…models · JEV 0.7AlphaGenome 2: DeepMind's DNA model now works cell by cellGoogle DeepMind released AlphaGenome 2, a model that predicts how DNA variants change gene activity in a specific…models · JEV 0.67Qwen4-Max: Alibaba's new flagship model goes liveAlibaba's 2.4T-parameter flagship is now generally available via API, with a 1M-token context, native vision, and…models · JEV 0.67Ling 3.2 Flash: Ant Group's 560B native multimodal MoE, now free on OpenRouterAnt Group's inclusionAI launched Ling-3.2-flash, a 560B-parameter MoE (25B active per token) that reasons natively…models · JEV 0.63ElevenLabs Scribe v3: speech-to-text with 27% fewer errors in 102 languagesTwo new speech-to-text models — Scribe v3 for recorded audio and Scribe v3 Live for real-time streams — cutting word…models · JEV 0.62LongCat-2.5: Meituan's 1.6T multimodal model for long-horizon coding agentsMeituan released LongCat-2.5 on 2 November, a native multimodal model with 1.6 trillion total parameters and about 48…models · JEV 0.62Claude Fable 5.2: Anthropic's fastest frontier model, 20% cheaperAnthropic's most capable generally available model gets up to 30% faster outputs, clearer communication, and a price…models · JEV 0.62Inkling 2 Small: Thinking Machines' 594B MoE, second-gen model with Apache 2.0 weights comingThinking Machines Lab released Inkling 2 Small, a 594B-total / 17B-active MoE with 1M-token context, scoring 31 on the…models · JEV 0.61ElevenLabs Eleven Realtime: speech-to-speech models for voice agentsElevenLabs launched Eleven Realtime v1 Mini and v1 Max — single speech-to-speech models that fold recognition…models · JEV 0.59AssemblyAI-TTS-1: low-latency text-to-speech built for live voice agentsAssemblyAI's first text-to-speech model, generally available 29 Oct: 4,000+ voices, instant cloning from short…models · JEV 0.57Crouton Alpha: OpenRouter's free, anonymous 1M-context reasoning modelOpenRouter listed stealth/crouton-alpha on November 3: a free multimodal reasoning model from an undisclosed provider…models · JEV 0.55K-EXAONE 2.5: LG opens free API access to its 750B sovereign modelLG AI Research opened free API access to K-EXAONE 2.5, a 750B-parameter Korean and English model with a 512K-token…models · JEV 0.55Gemini 4 Hydrogen: Google's fast daily-driver model for coding and agentic workGoogle's execution-focused Gemini 4 model: 1M-token context, four thinking levels, $1.50/$7.50 per 1M tokens, 2.5x…models · JEV 0.53iFLYTEK Spark X2.6: a 293B MoE trained entirely on domestic Chinese computingiFLYTEK launched Spark X2.6 at its Global 1024 Developer Festival on Oct 22, 2026: a 293B-parameter MoE (30B active)…models · JEV 0.52GPT-6.1 Terra: OpenAI's fast, half-price model for visual reasoning and computer useOpenAI released GPT-6.1 Terra on Oct 26, a mid-tier model for visual reasoning, precise computer use and everyday…models · JEV 0.52Voxtral 3: Mistral's voice agent stack — speech models plus SDK in 50+ languagesMistral's Voxtral 3 family pairs three speech-to-text models and a text-to-speech model with a LiveKit-based voice…models · JEV 0.51Kimi K3.1 Pro Highspeed on Amazon BedrockMoonshot's Kimi K3.1 Pro Highspeed, a post-trained variant that generates output 2-3x faster with improved reasoning…models · JEV 0.5Hy Realtime: Tencent's voice model that talks, calls tools, and singsTencent Hunyuan released Hy Realtime on 13 Nov 2026: a real-time speech model that understands spoken instructions…models · JEV 0.5Dessa: Block's full-duplex speech model trained only on human conversationBlock's in-house AI research lab announced Dessa on Oct 23, a full-duplex speech-to-speech model trained from scratch…models · JEV 0.5Apollo World: Baidu's AI world model that simulates 100,000 km of robotaxi driving a dayBaidu's Apollo World generates real-time multimodal driving simulations, including camera and lidar streams, to train…models · JEV 0.49Cognition SWE-2 FlashA smaller, faster coding model for Devin built on Thinking Machines' Inkling-Small, running 3x faster at one-fifth the…models · JEV 0.49Qwen4-Max Preview: Alibaba's next flagship, open for 14 days onlyAlibaba opened a limited preview of Qwen4-Max, the first model of its Qwen4 family, with a 1M-token context window and…models · JEV 0.47Chariot-1-TTS-Multilingual: Indian TTS that knows how text should be spokenIndian frontier lab Chariot launched Chariot-1-TTS-Multilingual, a text-to-speech model covering 11 Indic languages…models · JEV 0.47Mercury 3: Inception's 1M-context diffusion LLM at 1,068 tokens per secondInception released Mercury 3 on Oct 19, a diffusion language model that drafts many tokens in parallel instead of one…models · JEV 0.47

On device · 24

VoiceStudio: open-source, fully-local ElevenLabs alternative for voice cloning, dubbing, and audiobooksdebpalash/VoiceStudio is a desktop app that runs voice cloning, voice design, video dubbing, dictation, transcription…models · JEV 0.75Edge0-145B: open 145B MoE that runs on a single consumer GPUEdge0 open-sourced 145B and 120B latent-MoE models derived from Qwen3.8-Flash-Next, plus an inference framework that…models · JEV 0.74Strata: run a 125B model on a gaming PCFree, MIT-licensed local inference engine that runs Qwen3.8-Flash-Next (125B MoE, about 6B active per token) on…models · JEV 0.73Google AI Edge Gallery: offline speech recognition and synthesis on-deviceGoogle added fully offline speech recognition and text-to-speech to AI Edge Gallery, with ML Kit APIs on Android and…models · JEV 0.73smolmaxxing: run 16B LLMs on a 1 GB TV box or a 2013 iPadAn MIT-licensed llama.cpp patch set that runs language models up to a 16B MoE on devices with about 1 GB of usable RAM…models · JEV 0.65project-maya: run 700B-class open models on your own GPUsproject-maya is an open-source runner for the newest giant open-weight models — GLM-5.3 (744B), GLM-5.3-Flash (321B)…models · JEV 0.59Opera WebAI: local AI models built into the browser, no cloud neededOpera began shipping local AI models in its browsers on 27 Oct 2026: WebAI, an opt-in early preview in Opera One, lets…models · JEV 0.57llama-modchip: 284B DeepSeek-V4-Flash on 64GB of CPU-only RAMA llama.cpp fork that runs the 284B DeepSeek-V4-Flash at 1.71-bit quantization in under 59 GiB of RAM with no GPU…models · JEV 0.56BotPsy-0.8B: Tether's sub-billion on-device tool-calling modelTether AI Research released BotPsy, a 0.8B-parameter model built on Qwen3.5-0.8B and post-trained for function calling…models · JEV 0.56MACaQ: Huawei's internal quantizer for local inference, now open sourceHuawei-CSL open-sourced MACaQ, an MXFP4 W4A4 post-training quantization method that treats every linear layer…models · JEV 0.54bootgpt: a language model that lives in a 512-byte boot sectorbootgpt packs a tiny character-level neural network into a PC's master boot record: 510 bytes of hand-written 16-bit…models · JEV 0.49Mirage: a 22MB vision model that describes, detects, reads and segments in one passCactus Compute's 51.3M-parameter vision model fits in ~22MB at 2-bit, doing scene description, detection, OCR…models · JEV 0.49Mage-VL-Cowork: Microsoft's 4B computer-use model that clicks or codesA 4-billion-parameter open-weight vision-language model that operates a desktop two ways inside one policy: clicking…models · JEV 0.46Maple: sub-2-bit MoE pruning that runs Gemma 4 26B on a 6GB GPUManifold-Activated Pruning of Latent Experts compresses Gemma 4 26B-A4B from 9.92 GB to a 4.75 GB GGUF that the author…models · JEV 0.45Earlybird: a 29M-param speech-to-text model that outputs words before they endA 17 MB open-weights on-device speech-to-text model that predicts each word correctly a median 90 ms before the…models · JEV 0.45kimi-k3-in-c: a 2.78T model running on one CPU in 8 GB of RAMkimi-k3-in-c is a from-scratch C99 inference engine that runs Moonshot's 2.78-trillion-parameter Kimi K3 checkpoint on…models · JEV 0.41NVIDIA Alpamayo 2 Edge: a 2B reasoning model that drives from inside the carNVIDIA released the open weights for Alpamayo 2 Edge, a 2-billion-parameter vision-language-action model that plans…models · JEV 0.37RealtimeVAD: end-of-speech detection about 110ms earlier for voice assistantsA CPU-only, MIT-licensed drop-in for Silero VAD that adds a 10ms energy check with room-noise tracking, detecting end…models · JEV 0.36ByteSized: a 3KB LLM inference engine in pure CA 3,156-byte C engine that runs SmolLM2-135M token-for-token identical to Hugging Face, with no dependencies, aimed at…models · JEV 0.34NVIDIA Nemotron 3.5 Bolt E4B 4GB: an open small model for local voice agents, game characters and robotsNVIDIA released Nemotron 3.5 Bolt E4B 4GB on Oct 23, 2026: an open small language model post-trained from Google's…models · JEV 0.34Claris Harbor: turn your Macs into a private AI inference serverApple's Claris subsidiary is privately previewing Harbor, a Mac app that runs open-weight models locally on Apple…models · JEV 0.33Opposable: on-device AI that operates your Android phoneAndroid app that reads the screen through the accessibility service and taps and types in any app; when no cloud model…models · JEV 0.33Hugging Face AnySpeechRecognizer and AnyDecisionModel: open speech and decision models for Apple platformsTwo Swift packages from Hugging Face: AnySpeechRecognizer gives Apple apps one interface for on-device speech…models · JEV 0.32RigSpark: know which AI will run on your own computerRigSpark reads a machine's memory, GPU and disk and shows which open models and sizes will actually run on it, which…models · JEV 0.24

Decision · 14

Gemini Decisions API: fast, calibrated multimodal decisionsGoogle DeepMind launched the Gemini Decisions API powered by Gemini Decision 1, scoring up to 1,000 yes/no, choice, or…models · JEV 0.6Perplexity Decider v1.3: open decision models that return probabilities, now with LocatePerplexity released pplx-decider-v1.3 and pplx-decider-flash-v1.3 on 12 Nov 2026: open-weight decision models (27B and…models · JEV 0.58Liquid AI d1.2 and d1.2-ultra: decision models that now hear audioLiquid AI's decision models add audio input and five new typed tasks - reading values from documents, locating objects…models · JEV 0.54Decision Models in Hugging Face: a shared interface for models that decide instead of generatingHugging Face added decision models to Transformers: models that take a state plus typed questions (Choice, Score…models · JEV 0.51Zhipu GLM-5.3-FlashX-Legal reaches public APIsZhipu's legal-specialized GLM-5.3-FlashX model reached public API listings on OpenRouter and Vercel AI Gateway on Nov 2.models · JEV 0.45mammoth-1 and elephant-1: Baseten's open-weight decision models top the Jev Decision IndexBaseten's internal research lab Base Labs released two open-weight decision models — mammoth-1 (27B) and elephant-1…models · JEV 0.42SAP RPT-2.0: tabular foundation model adds anomaly detection, plus an ISS edge deploymentSAP announced SAP-RPT-2.0 at TechEd on 27 Oct 2026, extending its tabular foundation model beyond classification and…models · JEV 0.42Sarvam Prism: open-weight decision model that answers in 44 millisecondsSarvam AI launched Prism, a System 1 decision model built on Sarvam 105B that returns a probability score for each…models · JEV 0.41Strands Decider: AWS's open-weight decision model that picks instead of generatingOpen-weight decision models from AWS Strands Labs that choose from a fixed set of options and return calibrated…models · JEV 0.39Kanana-2-v-decide: Kakao's open decision model answers in 0.04 secondsKakao released Kanana-2-v-decide, a 3.67B multimodal decision model that reads text and images and returns typed…models · JEV 0.38Drex Max v1 open weights: Nace AI's 27B decision model, now runnable locallyOpen weights for Nace AI's 27-billion-parameter decision model: it returns a probability for every option in a single…models · JEV 0.37ESET NeoCortex: self-hosted cybersecurity foundation model for incident triageSpecialized world-model-architecture AI that powers ESET LiveCortex inside the ESET PROTECT platform, correlating…models · JEV 0.3Mouse-1: decision model for agent harnessesDecision model for agent harnesses launched Oct 21 on Product Hunt by maker Pete McGrath. Optimized for the hundreds…models · JEV 0.29Listen Twins: one-to-one AI models of real customers that predict audience reactionsListen Labs launched Listen Twins on 27 Oct 2026: one-to-one AI models of real people in a company's target audience…models · JEV 0.28

Composite · 5

Nemotron for Science: NVIDIA's open model, chemistry dataset and recipes for science agentsNVIDIA packaged its open Nemotron 3.5 Super VL vision-language model with the LitmusBench v0.2 chemistry dataset…models · JEV 0.61Cadence Maestro: domain-specialized AI models for chip design, built on Nvidia NemotronCadence Design Systems launched Maestro on Oct 26: a portfolio of domain-specialized AI models — post-trained language…models · JEV 0.55Sparks of Weight-Space ProgrammingMIT CSAIL researchers show GPT-6 Astra can read raw neural-network weights, identify what a network was trained to do…models · JEV 0.5Qwenputer: a functioning, deterministic RISC-V CPU compiled into Qwen3.5's weightsAn experimental model with a 32-bit RISC-V CPU compiled into Qwen3.5-4B's weights, using Gated DeltaNet layers as RAM…models · JEV 0.33nesformer: a working NES CPU compiled into Qwen3.5 weightsA compiler, not training, turns a stock Qwen3.5-4B checkpoint into a working NES 6502 CPU: the model predicts the…models · JEV 0.32

Ros · 1

NVIDIA Isaac Lab 3.0 GA + Isaac OS open runtime + Agile ONE SN humanoidMajor Isaac platform expansion at GTC Berlin Oct 21: Isaac Lab 3.0 is generally available as the open-source robot…models · JEV 0.73

Other · 1

WeVisOne-VL: Tencent's open 2B/4B vision model for detection, segmentation and mattingWeChat Vision (Tencent) releases Apache-2.0 WeVisOne-VL 2B and 4B weights, a VLM that handles detection, referring…models

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.