Best AI Models - Page 25
View:
-
1
GPT-Live
OpenAI
GPT-Live is a new generation of voice models for natural human-AI interaction, now powering ChatGPT Voice. It is built to make talking with AI feel much more like having a real conversation through a full-duplex architecture, meaning it can listen and speak at the same time. During conversations, GPT-Live can show it is paying attention with short acknowledgments like “mhmm” or “yeah,” engage in quick back-and-forth, or stay quiet when the user needs a moment to think. Instead of processing separate turns one after another, GPT-Live continuously processes input while generating output, allowing it to decide many times per second whether to speak, keep listening, pause, interrupt, or invoke a tool. For questions that require web search, deeper reasoning, or more complex work, GPT-Live can delegate to a frontier model behind the scenes and bring the result back into the conversation when it is ready, while still maintaining the flow of the voice interaction. -
2
GPT-Live-1
OpenAI
GPT-Live-1 is one of the two new GPT-Live voice models rolling out to ChatGPT users globally, built to make talking with AI feel much more like having a real conversation. It is powered by a full-duplex architecture, so it can listen and speak at the same time instead of waiting for one rigid turn to end before the next begins. During conversations, GPT-Live-1 can show it is paying attention with short acknowledgments, engage in quick back-and-forth, pause when the user needs a moment to think, or stay quiet when asked to listen. It continuously processes input while generating output, allowing the model to decide many times per second whether to speak, keep listening, pause, interrupt, or invoke a tool. GPT-Live-1 also separates natural interaction from deeper work: when a question requires web search, reasoning, or more agentic capabilities, it can delegate the task to a frontier model behind the scenes and bring the result back when it is ready. -
3
GPT-Live-1 mini
OpenAI
GPT-Live-1 mini is one of the two GPT-Live voice models rolling out to ChatGPT users globally, designed to bring more natural, intelligent, and responsive voice interaction to everyday conversations. Built with the same full-duplex approach as GPT-Live, it can listen and speak at the same time instead of waiting for rigid turn-by-turn exchanges. The model continuously processes input while generating output, allowing it to decide many times per second whether to speak, keep listening, pause, interrupt, or invoke a tool. This makes conversations feel faster, smoother, and more natural, with active listening, quick back-and-forth, better timing, and fewer awkward interruptions when the user pauses to think. GPT-Live-1 mini also benefits from the new ChatGPT Voice experience, where users can interrupt with a question, ask ChatGPT to slow down, or tell it to stay quiet and listen. -
4
Muse Image
Meta
Muse Image is Meta’s image generation model from Meta Superintelligence Labs, built into Meta AI for creating, editing, and sharing high-quality visuals. The model can turn simple conversational prompts into detailed images, blend multiple photos together, remove unwanted objects, generate legible text inside visuals, and create styled outputs such as portraits, posters, stickers, room redesigns, infographics, and fantasy scenes. Muse Image uses advanced reasoning through Muse Spark to plan layouts, understand context, look up real-time web information, and combine visual references more intelligently. Users can start with suggested presets, mention Instagram accounts to personalize creations, and sketch or annotate edits directly on top of an image. The model powers creative experiences across Meta AI, Instagram Stories, WhatsApp chats, and soon Facebook, Messenger, and advertiser tools through Meta Advantage+ creative. -
5
Seedream 5.0 Pro
ByteDance
Seedream 5.0 Pro is a multimodal image creation model built for advanced reasoning, efficient content creation, and professional production. In real production environments, visual appeal is only the starting point; what matters is whether the model can efficiently meet complex creative demands, close the gap between the creator’s intent and the final visual output, and deliver true usability. Compared to previous versions, Seedream 5.0 Pro improves image-text alignment, structural coherence, text rendering, and visual aesthetics, while introducing core breakthroughs in complex information visualization, interactive precision editing, realistic imagery, portrait textures, and native multilingual generation. It can accurately transform data, concepts, and dense text into professional layouts for high-density content production, including infographics, educational images, technical drawings, UI designs, posters, and specialized professional visuals. -
6
Bonsai 27B
PrismML
Bonsai 27B is the new multimodal flagship of the Bonsai family and the first 27B-class model to run on a phone. Based on Qwen3.6 27B, it brings a new capability tier to local devices; multi-step reasoning, structured tool calls, vision tasks, and computer-use agentic loops that stay coherent across many steps. Bonsai 27B comes in two variants. Ternary Bonsai 27B uses ternary weights with FP16 group-wise scaling, giving 1.71 effective bits per weight and a 5.9 GB footprint for the quality-oriented laptop-class version. 1-bit Bonsai 27B uses binary weights with the same group-wise scaling, giving 1.125 effective bits per weight and a 3.9 GB footprint that fits within the memory budget of an iPhone 17 Pro. Both variants run end-to-end across the language network, embeddings, attention, MLPs, and LM head with no higher-precision escape hatches. They are multimodal, with a compact 4-bit vision tower, so on-device workflows can understand screenshots, documents, and camera input. -
7
Seed2.1 Turbo
ByteDance
Seed2.1 Turbo is a next-generation AI productivity model designed to execute complex real-world tasks with strong general-agent, coding, and multimodal capabilities. It goes beyond one-off answers by carrying multi-step workflows toward defined goals and producing practical, usable outcomes across tools, environments, and interaction modes. For professional work and everyday consultation, it can support project planning, document and file processing, information analysis, solution design, content planning, tool use, and results consolidation. It also handles teaching, office, and research scenarios such as generating lesson-plan slides, analyzing complex spreadsheets, and producing industry reports. In software engineering, Seed2.1 Turbo supports end-to-end delivery across requirement analysis, feature implementation, bug fixing, environment setup, terminal usage, and result validation, while understanding codebase architecture, dependencies, and business logic to coordinate changes. -
8
Seeduplex
ByteDance
Seeduplex is a native full-duplex speech large language model built on a new “listen while speaking” framework for more natural, fluid, and precisely paced voice interaction. Unlike traditional half-duplex systems that alternate between listening and replying, it continuously receives and understands user-side audio, allowing it to listen and speak simultaneously while tracking the broader acoustic environment. Its high-precision interference suppression distinguishes genuine user interaction from background noise, broadcasts, navigation prompts, side conversations, and overlapping voices, reducing false responses and false interruptions in complex settings. Seeduplex also combines speech and semantic features for adaptive endpoint detection, helping it recognize when a user is thinking, hesitating, correcting themselves, or has actually finished speaking. It can wait patiently through reflective pauses, respond quickly once an utterance ends, and stop smoothly when interrupted. -
9
Laguna XS 2.1
Poolside
Laguna XS 2.1 is an upgraded open weight agentic coding model designed for long-horizon work on a local machine. It uses a 33-billion-parameter Mixture-of-Experts architecture with 3 billion activated parameters per token, retaining the same efficient architecture as Laguna XS.2 while improving multilingual software engineering and terminal-style task performance. The model is built to support coding agents that inspect repositories, reason through complex changes, use tools, execute commands, and continue working across extended tasks. It is served with a 256K context window, giving agents room to work with large codebases, lengthy histories, and multi-step workflows. Laguna XS 2.1 is supported by vLLM, SGLang, NVIDIA TensorRT-LLM, Hugging Face Transformers, and Ollama, with native llama.cpp support planned. It is available in BF16, FP8, INT4, and NVFP4 checkpoints, allowing developers to choose between maximum fidelity and configurations suited to tighter VRAM or compute budgets. -
10
Antares
Cisco
Antares is a family of open-weight security small language models purpose-built to localize known vulnerabilities inside large codebases. Antares-350M and Antares-1B are compact enough to run locally or on premises, helping teams keep proprietary source code inside their environment while reducing inference cost and runtime. Starting from a vulnerability description, advisory, or CWE category, the model follows an iterative investigation process similar to a human analyst, it searches for relevant code patterns, reads candidate files, incorporates new evidence, changes direction when a path is unproductive, and narrows the search to the files most likely to contain the weakness. Antares returns a ranked list of potentially vulnerable source files together with the terminal exploration trace that produced the result, making findings easier to review and prioritize. -
11
MAI-Cyber-1-Flash
Microsoft
MAI-Cyber-1-Flash is Microsoft AI’s compact, code-heavy security model for finding vulnerabilities in complex codebases. Derived from the MAI-Thinking-1 lineage and built from scratch on high-quality data, it is deeply integrated into MDASH, Microsoft’s multi-agent vulnerability identification and remediation harness. MDASH uses more than 100 expert-tuned agents and multiple leading models to find, validate, and remediate software vulnerabilities, while MAI-Cyber-1-Flash efficiently handles up to 90% of tasks. Exceptionally difficult cases can be routed to larger models such as GPT-5.4, creating a well-tuned multi-model system that selects the right model for each task. Together, MDASH and MAI-Cyber-1-Flash achieved 96% on CyberGym, outperforming Mythos, Gemini, and GPT-based alternatives in reasoning over large codebases to identify vulnerabilities. -
12
Grok Voice Think Fast 2.0
SpaceXSI
Grok Voice Think Fast 2.0 is xAI’s flagship voice model for building real-time assistants, phone agents, and interactive voice systems that stream audio and text bidirectionally over WebSocket. Developers can configure system instructions, high or no reasoning effort, built-in or custom voices, automatic server-side voice activity detection, silence duration, idle re-engagement, playback speed, and session resumption after temporary disconnects. It accepts PCM, G.711 μ-law, G.711 A-law, and Opus audio through JSON or raw binary frames, with configurable PCM sample rates from telephone quality to 48 kHz. It supports more than 20 languages with native-quality accents, automatic language detection, natural responses in the speaker’s language, and seamless code-switching. Language hints and up to 100 key terms improve transcription of regional speech, names, products, codes, addresses, and specialized terminology, while pronunciation replacements correct spoken output. -
13
Lyria 3.5
Google
Lyria 3.5 is Google’s newest AI music generation model, designed to help people create richer, high-fidelity tracks with greater musical and technical control. Available in Google Flow Music, it advances musicality with more natural, complex melodic structures and a stronger understanding of rhythm, arrangement, tempo, dynamics, and more. Enhanced lyric generation improves prompt adherence and structural awareness, while upgraded vocals add more realistic expression, emotional nuance, and clearer pronunciation. Users can begin with a simple idea or provide detailed instructions for genre, instrumentation, mood, key, tempo, vocal style, language, and production character, then refine the sound they want. Lyria 3.5 supports variable song lengths, allowing creators to request a quick 60-second clip, a half-length track, or a cohesive song lasting up to three minutes. It can generate music across global languages and genres, from pop, funk, and R&B to reggaeton, jazz, and electronic music. -
14
MiniMax Music 3.0
MiniMax
MiniMax Music 3.0 is a music-generation API for creating songs from a description, lyrics, or reference audio. Developers use the prompt parameter to define style, mood, instrumentation, vocal character, and production direction, while the lyrics parameter supplies vocal content. Its upgraded semantic model improves creative-intent understanding and reduces drift in AI-generated music. Higher sound quality produces clearer mixes and supports specific instruments and playing techniques such as slides and legato. A new vocal engine delivers more natural synthesis with control over melody, pronunciation, breathing, and layered harmonies. Teams can first call the Lyrics Generation API to write full lyrics with sections such as Verse, Chorus, and Bridge, then send them to the Music Generation API, or skip that step and generate a song directly with lyrics optimization. Music 3.0 also supports instrumental-only creation. -
15
Gemini Robotics 2
Google DeepMind
Gemini Robotics 2 is Google DeepMind’s intelligence layer for adaptable robots, bringing whole-body control, advanced dexterity, embodied reasoning, and multi-robot collaboration to physical AI. It includes three models. Gemini Robotics 2 is a vision-language-action model that converts visual and language input into motor control, enabling humanoids and bi-arm robots to act from feet to fingertips. It can coordinate walking, crouching, reaching, balancing, and object manipulation, while controlling five-fingered hands or standard grippers for delicate and precise tasks. Gemini Robotics ER 2 serves as the high-level brain, communicating with people, understanding its surroundings, planning multi-step tasks that last several minutes, coordinating actions with the VLA, tracking progress, self-correcting failures, and allowing different robots to work together. -
16
NVIDIA Parakeet
NVIDIA
NVIDIA Parakeet-RNNT-1.1B is a multilingual automatic speech recognition model built for quality transcription across voice applications. With 1.1 billion parameters and training on more than 90,000 hours of speech, it supports 25 languages and regional variants, including English, Spanish, French, German, Italian, Arabic, Japanese, Korean, Portuguese, Russian, Hindi, Dutch, Danish, Norwegian, Czech, Polish, Swedish, Thai, Turkish, and Hebrew. The model automatically detects the spoken language and uses a universal tokenizer created by training language-specific tokenizers and merging them into a shared vocabulary, enabling efficient cross-lingual learning and deployment. Parakeet-RNNT produces case-sensitive transcripts with upper and lowercase text, punctuation, spaces, and apostrophes, making the output suitable for production voice applications and downstream language understanding. -
17
NVIDIA Alpamayo 2 Super
NVIDIA
NVIDIA Alpamayo 2 Super is a frontier open model for robotaxis and autonomous vehicles, built to reason through rare, complex driving situations and produce decisions developers can inspect, validate, and trust. Based on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning, it combines commercial openness with multitask capabilities for autonomous driving. The model reasons across full-surround camera coverage, fusing front, side, and rear views to understand lane changes, merges, unprotected turns, and complex intersections. For each scenario, it can generate a planned vehicle trajectory, a chain-of-causation trace explaining the decision, a meta-action such as yielding or stopping, reasoning auto-labels for training and validation, and visual question-answering responses grounded in specific image regions. These linked outputs make it easier to connect what the model observed with the action it selected. -
18
Shieldstral
Mistral AI
Shieldstral is a 3B open-weights, policy-adaptive multimodal safety classifier designed to evaluate text, images, and text-plus-image content using policies defined at inference time. Instead of relying on a fixed taxonomy of harm categories, it frames moderation as a binary question-answering task: users provide an instruction describing the evaluation context and strictness, a yes-or-no safety question, and the content to judge. The model reads the “yes” and “no” logits and converts them into a continuous, calibrated safety score, allowing applications to threshold or rank results by confidence rather than depend on a single discrete label. This formulation unifies prompt classification, response moderation, refusal detection, toxicity detection, and multimodal safety in one interface, while letting teams adapt policies without retraining the model. Shieldstral can evaluate prompts, responses, prompt-response pairs, images, and images with accompanying text. -
19
GPT‑5.6‑Cyber
OpenAI
GPT-5.6-Cyber is OpenAI’s most advanced purpose-trained cybersecurity model for approved defenders conducting authorized vulnerability research, exploit validation, and security testing. Built on GPT-5.6 Sol, it is trained to improve performance on specialized cybersecurity tasks such as finding zero-day vulnerabilities, developing exploit chains, testing authentication bypasses, privilege escalation, and advanced security research. The model is designed to reduce unnecessary refusals on legitimate higher-risk, dual-use cybersecurity work while helping trusted defenders conduct real-world security activities. GPT-5.6-Cyber improves performance on exploit development workflows and can generate proof-of-concept exploits alongside technical findings, assess the severity and impact of novel vulnerabilities, and support vulnerability discovery and report writing. It is particularly suited to sustained reasoning across large and unfamiliar codebases. -
20
Nemotron 3.5 Lightning
NVIDIA
NVIDIA Nemotron 3.5 Lightning is an open 30B-parameter mixture-of-experts model with 3B active parameters, designed for high-volume, low-latency execution in long-running and always-on AI agents. Built for the execution layer of agentic systems, it handles frequent tasks such as tool calls, output validation, routine commands, and subagent delegation while larger reasoning models focus on planning and orchestration. Its MoE architecture activates only a fraction of parameters for each token, combining the capacity of a larger model with lower compute requirements. The model is trained for popular agent harnesses and supports speculative decoding through multi-token prediction, DFlash, and DSpark to improve inference speed across different serving scenarios. It is available with BF16 and NVFP4 checkpoints and can run from local systems such as DGX Spark and GeForce RTX hardware to data center environments. -
21
Ling 3.0 Tiny
Ant Group
Ling 3.0 Tiny is an open-weights reasoning model with 7.9B total parameters, 1.3B active parameters, and a 262K-token context window. Built with a mixture-of-experts architecture, it extends the open-weights Pareto frontier for intelligence versus active parameters and is small enough to run locally in many settings. The model scores 25 on the Artificial Analysis Intelligence Index, comparable to gpt-oss-120b (high, 24) while using 15x fewer total parameters and 4x fewer active parameters. This parameter efficiency comes with relatively high token usage, with 213M output tokens required to run the Intelligence Index. Ling 3.0 Tiny also shows substantial improvements in hallucination behavior over Ling-mini-2.0, improving its AA-Omniscience score by 59 points while maintaining similar accuracy. Rather than guessing when uncertain, it attempted only 37% of questions in the evaluation, resulting in a 30% hallucination rate compared with 96% for the previous generation. -
22
Higgs Audio / Avatar
Boson AI
Higgs Audio / Avatar is a family of foundation audio and avatar models designed to generate natural speech, understand tone, emotion, and intent, and give voice interactions a visual presence. The models support text-to-speech, speech-to-text, avatar generation, and automatic voice casting that selects an appropriate voice based on context, sentiment, and content. Built for real-world production, Higgs combines expressive generation, robust speech understanding, and flexible deployment for workloads where quality, latency, and reliability matter. High-accuracy multilingual speech recognition supports major languages, while voice cloning reproduces a speaker’s tone from short reference samples to maintain consistent brand voices across interactions. Sentiment detection reads emotional signals in speech to enable smarter routing, stronger analytics, and more context-aware agent behavior. -
23
GPT-5.6 Sol Ultrafast
OpenAI
GPT-5.6 Sol Ultrafast is a new OpenAI API service tier that runs GPT-5.6 Sol up to 14× faster than Standard processing, bringing frontier intelligence to products and workflows where every second matters. Powered by Cerebras, it can generate up to 750 output tokens per second, allowing advanced reasoning to operate at real-time speeds without requiring a smaller or more specialized model. It is designed for time-sensitive business workflows where faster responses can change what AI can realistically do. Applications include incident response, where models can analyze logs, code changes, traces, and engineer reports while an outage is unfolding; financial research and security, where changing market signals and suspicious transactions can be assessed quickly; and customer support and voice, where complex issues can be resolved without interrupting a live conversation. In commerce, it can answer product questions, check inventory, and personalize recommendations. -
24
Qwen3.8-2.4T-A95B
Alibaba
Qwen3.8-2.4T-A95B is the largest open model in the Qwen3.8 family, bringing Qwen-Max-class capabilities to an open release. Built on the architectural foundation of Qwen3.5, it delivers substantial improvements across coding, professional work, research, and long-horizon agentic tasks, with a focus on carrying complex, multi-step work through to completion more reliably. The causal language model uses a mixture-of-experts architecture with 2.4 trillion total parameters and 95 billion activated parameters, including 512 experts with 10 routed and one shared expert active at a time. It supports a native context length of 262,144 tokens that can be extended to approximately 1.01 million tokens. Agent execution is strengthened through better autonomous planning and improved handling of environment feedback, while broader compatibility with popular agent harnesses and development tools simplifies integration into existing stacks. -
25
Pika Soundtrack
Pika
Pika Soundtrack is a video-to-audio model that turns silent video into a native soundtrack complete with motion-aware sound effects, music, ambience, and voiceover that follow what happens on screen. Users can leave the prompt blank to generate a full soundscape automatically or provide direction specifying what the model should emphasize, include, or leave out. Rather than simply generating sound with video attached, the model is designed to understand what is happening in a scene, place each sound at the right moment, and keep every audio layer coherent throughout the full video. This synchronization approach allows sound effects, ambience, music, and speech to feel as though they belong naturally within the same scene. Pika reports that, in its full-duration benchmark, Soundtrack achieved the strongest semantic alignment and lowest audiovisual desynchronization among the models tested, including LTX-2.3 Foley V2A, HunyuanVideo-Foley, and MMAudio v2. -
26
Pika Music
Pika
Pika Music is a generative music model designed to meet creators wherever an idea starts, turning text prompts, lyrics, voices, reference tracks, or combinations of these inputs into complete songs up to six minutes long. A lyric can become a stripped-back ballad, a dance-pop record, or a cinematic rock performance, giving creators flexibility to explore different musical directions from the same starting material. Vocal references can shape the character and delivery of a performance, while music references can serve as the starting point for something new. The model’s key strength is composability: rather than forcing creators into separate workflows for lyrics, voices, references, and musical direction, it can bring all of these signals together within a single generation process. It supports both text-and-lyrics-to-music workflows and voice-conditioned music, allowing users to experiment with how words, vocal identity, style, and arrangement interact. -
27
Pika SFX
Pika
Pika SFX is a sound-effects generation model that turns natural-language direction into focused, ready-to-use audio for video, games, editing workflows, and creative tools. Users simply describe the sound they want, whether it is a glass shattering, a metal door slamming in an empty warehouse, a cork popping and fizzing into a glass, or something more stylized. The model can handle both a single crisp event and longer sequences, with control over material, space, perspective, timing, texture, and mood. It can generate natural Foley, cartoon-style effects, designed fantasy sounds, environmental ambience, and other prompt-driven audio while following the brief closely. Unless requested, it avoids adding unrelated speech, music, or background noise, helping creators generate clean effects that can be dropped directly into a project. -
28
Pika Speech
Pika
Pika Speech is an expressive text-to-speech model built for the inflection, rhythm, and timbre that make narration, characters, and spoken moments feel human. Rather than simply reading text aloud, it is designed to set the tone and give creators control over how a line is delivered. Users can choose from preset voices or create a voice clone from only a few seconds of reference audio, then direct the performance with a caption describing the desired delivery, such as bright and brisk, low and reflective, crisp and formal, or a custom style. The model generates 48 kHz audio and supports requests up to five minutes long, making it suitable for narration, character dialogue, product experiences, storytelling, and other spoken-content workflows. It is also designed for fast iteration: in Pika’s local testing, a real-time factor of 0.02 means that one minute of speech takes about one second to generate. -
29
MAI-Image-2.6
Microsoft
MAI-Image-2.6 is Microsoft AI’s latest image generation model, designed to push image quality forward across text-to-image generation and editing. It delivers broad improvements over MAI-Image-2.5 across every measured Arena category, with particularly strong gains in text rendering. The model produces stronger portraits and 3D imagery, along with more polished commercial and photorealistic outputs for product, branding, and cinematic use cases. It also expands creative control with support for working across multiple references, richer grounding, and greater control over reasoning, format, and resolution. In independent Arena evaluations, MAI-Image-2.6 ranked No. 2 on the text-to-image leaderboard and reached No. 3 for image editing, demonstrating improvements across both generation and editing workflows. Its image editing performance showed especially large advances in text rendering and product, branding, and commercial design. -
30
Gemini Omni 1.1 Flash
Google
Gemini Omni 1.1 Flash is a production-ready generative video model designed to give developers more control over AI video creation and editing. It can extend an existing scene in 10-second increments up to 40 seconds while analyzing as much as 10 seconds of prior context, improving visual consistency and narrative continuity across longer sequences. Developers can specify both the first and last frame of a shot, and the model generates continuous motion between them for smooth transitions, camera orbits, zooms, and seamless looping clips. A 360p preview mode supports faster prototyping and storyboard iteration, while final videos can be generated at 1080p or upscaled to 4K for polished professional production. Omni 1.1 also accepts up to three seconds of reference video as multimodal input, helping preserve visual context, character consistency, motion, and scene direction.
