Mercury 3
Inception
A diffusion-based language model for reasoning, tool use, and latency-sensitive agentic workflows, served through Inception's API.
The frontier and open-weights models developers are building on — with the specs that actually matter: context window, modalities, licensing and price per token.
Inception
A diffusion-based language model for reasoning, tool use, and latency-sensitive agentic workflows, served through Inception's API.
Meta
An open-weight, natively multimodal model for coding, long-horizon agentic work, and complex software engineering.
NVIDIA
An open-weight multimodal mixture-of-experts model for specialized agents, visual reasoning, coding, and tool use.
Google DeepMind
Gemini 4 Argon is Google's frontier reasoning model for complex software engineering, mathematical and scientific problem solving, and long-running agentic workflows.
Meta
A frontier model for personal and professional agents, combining Active Memory and Active Focus for sustained coding and complex tasks. Available through the Meta Model API and Muse Code.
Mistral AI
Mistral Large 4 is a natively multimodal model with 1 trillion total parameters and 49 billion active parameters, built for coding, agentic workflows and multimodal understanding. Available in public API preview; weights are planned for release by the end of October 2026.
Laminar
flow-1 is our new model for understanding agent traces, trained with reinforcement learning to identify hard-to-spot failures and explain why they happened. On our benchmark, it matches GPT-6-sol in detection quality while being 23x cheaper.
Cloudflare
Clef is a 27B multimodal model that turns a state and a schema of typed questions into decisions. It reads the state as text, JSON, images, or video, and returns a probability for every allowed option of every question in a single forward pass. There is no free-form text generation and no output parsing.
OpenAI
Near-Astra performance for complex coding, computer use, and professional work at a lower cost.
Anthropic
Claude Sonnet 5.5 is a clear upgrade over Claude Sonnet 5, runs 30%+ faster, and costs up to 30% less for most work.
Built for deep creative direction and character design. Create entirely new voices from scratch using natural language prompts to bring characters to life across gaming, immersive audiobooks, podcasts, and interactive media. Direct every performance line by line with granular control over acting cues, pacing, dialect shifts, and backchanneling.
Built for high-volume, cost-efficient scale. Optimized for high-volume dubbing, audio content creation, and expressive voice agents with fine-grained control over tone, pacing, and expressive nuance.
Anthropic
Claude Opus 5.5 leads in agentic coding and knowledge work, and costs 40% less to run than Opus 5 on typical workloads.
Black Forest Labs
FLUX 3 Action is an open weights 7B world action model. It takes camera frames, the robot's state and a text instruction, and returns the next chunk of actions, denoised together with the next video frames.
OpenAI
OpenAI's most efficient model for focused, high-volume tasks
OpenAI
OpenAI model for complex coding and agentic workflows
PrismML
A ternary 27B multimodal reasoning model based on Qwen3.8, supporting coding, mathematics, image understanding, and tool use.
Cactus Compute
A 121M-parameter open model for tool calling, structured extraction and text embedding on tiny devices. One set of weights ships as an intelligence ladder: every depth from 2 to 20 layers is a deployable model, an 8-29 MB CQ2-bit binary.
Built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding.
Built for high-complexity tasks, with increased intelligence and multi-step reasoning.
DeepSeek
We introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. The model natively processes images and text, and generates text autoregressively.
OpenAI
GPT-Image-2.5 Flare brings the same improvements in quality, editing, and speed.
OpenAI
GPT-Image-2.5 Sunburst offers an extra level of precision for detailed creative work with longer generation times.
OpenAI
Our most capable model, built for the hardest end-to-end work.
Microsoft AI
Turn noisy audio into precise, domain-specific transcripts, with leading FLEURS and Artificial Analysis accuracy scores.
Our most intelligent workhorse model, delivering significant improvements from 3.7 Flash across software engineering, agentic tasks, and critical, multi-step reasoning in specialized domains.
Our most capable cybersecurity model with frontier-level performance in vulnerability detection and automated patching, available to trusted defenders through our new Fairwind Program.
Meta
A multimodal reasoning model for coding and long-running agentic workflows, improving instruction following and efficiency over Muse Spark 1.2.
Osaurus
An 8B mixture-of-experts that activates only ~1B parameters per token, runs the full Osaurus tool surface in 6.3 GB, and stays resident on the 8–16 GB Macs most people own.
World Labs
Atlas is an omni model that we pretrained from scratch to natively operate on text, images, video, and 3D. It is a multimodal autoregressive diffusion transformer: all inputs are combined into a shared spatial context.
Anthropic
Claude Fable 5.1 sets a new standard for coding, knowledge work, and long-running problem-solving tasks.
Anthropic
Mythos 5.1 is available only through our trusted access programs; its safeguards are specifically designed to support work in cybersecurity and the life sciences.
Omni now delivers studio-quality video production, including the ability to extend a scene, first and last frame interpolation, crisp 4K upscaling, faster prototyping, and more.
Our most precise speech-to-text model yet, designed for intelligent voice interactions.
Our most precise speech-to-text model yet, designed for intelligent voice interactions.
Zhipu AI
GLM-5.3-Flash is the first native multimodal model in the GLM-5 series, delivering stronger intelligence than GLM-5.2 while maintaining an exceptionally cost-efficient architecture.
Fastino Labs
A new boundary-prediction architecture replaces span enumeration, adding joint entity-relation extraction, constrained classification, and unlimited span length.
Fastino Labs
A new boundary-prediction architecture replaces span enumeration, adding joint entity-relation extraction, constrained classification, and unlimited span length.
Fastino Labs
A new boundary-prediction architecture replaces span enumeration, adding joint entity-relation extraction, constrained classification, and unlimited span length.
Qwen
An open-weights multimodal MoE model that doubles as an early preview of the Qwen4 architecture, the same role Qwen3-Next played for Qwen3.5. It pairs Gated DeltaNet with Qwen Sparse Attention, widens the residual stream into four gated branches, and adds 51B N-gram embedding parameters that cost almost nothing per token. Natively 262K context, extensible to 1M with YaRN.
Zhipu AI
GLM-5.3 is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench.
Qwen
Built on the architectural foundation of Qwen3.5, Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks. Qwen3.8-27B brings these advances to a compact, deployment-friendly dense model: a native vision-language model that understands images and videos, with flexible thinking control, designed to carry complex, multi-step tasks through to completion with greater reliability.
Our most intelligent workhorse model yet for coding and agents.
xAI
xAI's frontier Grok model, tuned for long-running agents, coding, knowledge work, and visual projects.
Liquid AI
LFM2.5-VL-3B is the multimodal variant of LFM2.5, pairing the LFM2.5-2.6B backbone with a SigLIP2 NaFlex 400M vision encoder. It answers directly rather than reasoning, and adds screen understanding, grounding and function calling for on-device use.
Superwhisper
A 0.6B-parameter text normalizer for speech-to-text output. It takes a raw ASR transcript and rewrites it as clean written text: fillers removed, false starts and self-corrections resolved to the value the speaker landed on, punctuation and capitalization applied, and spoken numbers, dates, times, currency and email addresses rendered in written form.
LTX
LTX-2.5 generates multi-shot scenes in one pass, edits real footage, and exports cinema-grade EXR. Open weights you can fine-tune and run on your hardware.
Cactus Compute
An open 45M-parameter model for tool calling, device use, and structured extraction. Needle 2 runs as a 14 MB binary in 28 MB of session RAM.
NVIDIA
A customizable open 30B MoE model with 3B active parameters, providing optimal high-volume execution for autonomous agents.
inclusionAI
We are introducing Ling-3.0-tiny, a lightweight hybrid reasoning MoE model with 7.9B total parameters and only 1.3B activated parameters per token. It is designed to deliver strong reasoning and agentic capabilities at low inference cost, making advanced model capabilities more accessible for local and resource-constrained deployment.
Meta
Muse Glimmer is a 30-billion-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware. The model integrates multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery into a single model that runs locally without requiring cloud infrastructure or network access.
webAI
A LoRA adapter over SmolLM2-1.7B-Instruct for formal logic — FOL translation, entailment, semantic parsing and Lean assistance — quantized to a 1.06 GB download that runs on a phone.
webAI
A 3B formal-logic reasoning model built from SmolLM3-3B by LoRA fine-tuning, checkpoint fusion, WiSE-FT interpolation and entropy-weighted GRPO, which beats gpt-oss-120b on four of five lanes of webAI's formal-reasoning suite.
Black Forest Labs
FLUX 3 is our new multimodal foundation model. It jointly learns from images, videos, and audio within a unified architecture, because what it needs to learn is not any one of these elements in isolation. Instead, a model must learn a representation of the world: how objects hold together, how things move, and how events sound.
Liquid AI
LFM2.5-2.6B is part of LFM2.5, a family of hybrid models designed for on-device deployment. It builds on the LFM2 architecture with a 128K context window and agentic post-training.
Mistral AI
A 3B open-weights, policy-adaptive multimodal safety classifier that matches models up to 7x its size on text safety and sets a new state of the art on multimodal moderation.
MiniMax
MiniMax H3 is a general-purpose, omni-modal generative system. It supports unified understanding of multimodal contexts composed of text, images, video, and audio, and can generate video with native stereo audio at resolutions up to 2K and durations of up to 15 seconds.
Thinking Machines Lab
A quarter the size of Inkling at comparable performance: a multimodal Mixture-of-Experts reasoner (276B total, 12B active) with controllable reasoning effort, fine-tunable on Tinker.
Anthropic
Opus 5 is a step change improvement for the Opus tier powering long-running agents while delivering improvements in coding and professional work.
Our fastest, most cost-effective 3.5-class model, delivering 350 output tokens per second.
Our workhorse model that delivers better coding, knowledge work, and multimodal performance.
Poolside
The most capable agentic coding model in its weight class by a wide margin.
Moonshot AI
Kimi K3 is a 2.8T-parameter model built on Kimi Delta Attention and Attention Residuals, with native vision capabilities and a 1-million-token context window — the world's first open 3T-class model.
Thinking Machines Lab
Thinking Machines Lab's first open-weights model: a multimodal Mixture-of-Experts reasoner (975B total, 41B active) with controllable reasoning effort, fine-tunable on Tinker.
PrismML
The first 27B-class model to run on a phone, based on Qwen3.6 27B: a multimodal flagship shipping in ternary (5.9 GB) and 1-bit (3.9 GB) forms with speculative decoding.
OpenAI
GPT-5.6 model optimized for cost-sensitive workloads.
OpenAI
Frontier model for complex professional work.
OpenAI
GPT-5.6 model that balances intelligence and cost.
xAI
xAI's Grok 4.5 flagship for chat, coding, and agentic tool use, with lower hallucination risk.
Cognition
The most capable model Cognition has trained so far. It reaches frontier-level intelligence at a much lower cost, advancing the cost-performance Pareto curve.
Anthropic
Claude Sonnet 5 is built to be the most agentic Sonnet model yet. It can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models.
Zhipu AI
We're introducing GLM-5.2, our latest flagship model for long-horizon tasks.
DeepSeek
DeepSeek-V4-Flash-0731 is the official release of DeepSeek-V4-Flash, superseding the preview version, with substantially enhanced agentic capabilities.
DeepSeek
DeepSeek-V4-Pro is a 1.6T-parameter Mixture-of-Experts model with 49B activated, supporting a one-million-token context. The 0813 revision is its GA release.
OpenAI
A new class of intelligence for coding and professional work.
Moonshot AI
Kimi K2.6 is an open-source, native multimodal agentic model that advances practical capabilities in long-horizon coding, coding-driven design, proactive autonomous execution, and swarm-based task orchestration.
Reflection AI
Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active, built for coding, reasoning, and agentic workloads.
Cloudflare
A System One extraction model for extracting, computing, estimating, and forecasting structured values from unstructured inputs. Available in preview on Cloudflare Workers AI.
Cloudflare
A smaller decision model based on Qwen3.5-9B that scores schema-constrained options. The October update reduces hosted pricing and context size without changing the released weights.
Cloudflare
A multimodal decision model built on Qwen3-Omni-30B-A3B-Instruct. It scores schema-constrained options from text, images, audio, and video without generating free-form output.
LightOn AI
A local code-search agent that searches repositories with ColGrep, reads relevant code, and returns file paths and line ranges to a coding agent. Post-trained from MiniCPM5-2B.
Liquid AI
d1-2B is a decision model built on LFM3-VL-2B. It reads text and images and answers multiple questions in a single forward pass, each with a calibrated probability.
Liquid AI
d1-Omni-2B is a decision model built on LFM3-Omni-2B. It reads text and images, listens to audio, and answers multiple questions in a single forward pass, each with a calibrated probability.
RadixArk
RadixArk’s first model, built on Qwen3.8-27B and trained with reinforcement learning using Miles on long-horizon agent tasks.
Fine-tuned for finding and fixing cybersecurity vulnerabilities at a lower price per token than larger models.
Google DeepMind
Bringing frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows.
Moonshot AI
Kimi K3.1 Pro is a post-trained version of Kimi K3 with improvements in coding, 3D modeling, financial research, and agentic tasks. Available through Kimi's products and API, with full model weights scheduled for release by November 4, 2026.
Moonshot AI
Kimi K3.1 Pro Highspeed is the faster variant of Kimi K3.1 Pro, which Moonshot AI reports delivers 2–3 times the speed for more responsive interactions. Available through Kimi's products and API.
Liquid AI
LFM2.5-24B-A2B is a reasoning model with 24 billion total parameters and 2 billion active parameters, built for fast, private on-device agents. It combines gated short convolutions and attention with a mixture-of-experts architecture.
Liquid AI
LFM3-1.7B is a text model for local agents and specialized applications.
Liquid AI
LFM3-Audio-2B is a compact audio model that runs locally on laptops and edge devices. It takes text or audio as input and responds with text, speech, or both.
Liquid AI
LFM3-Omni-2B brings vision, audio, and language together in a single on-device model that can see, listen, and talk back.
Liquid AI
LFM3-VL-2B is a vision-language model for video understanding, visual reasoning, and spatial and embodied reasoning on-device.
Meta
A frontier model for personal and professional agents, combining Active Memory and Active Focus for sustained coding and complex tasks. Available to Muse Code subscribers.
Ornith
Ornith 1.5 122B is a mixture-of-experts reasoning model for coding and agentic workflows, with support for tool calling. Its 122 billion parameter weights are available under the MIT license.
Thesys
An open-weight diffusion model fine-tuned from DiffusionGemma to generate user interfaces in OpenUI Lang on consumer GPUs.
Recursive
A mixture-of-experts model built for fast agentic coding on local hardware, with 40 billion total parameters and 2.5 billion active per token. Built from scratch with Recursive’s Eureka Machine automated research system.
Cognition
A coding model post-trained from Kimi K3 with reinforcement learning across reasoning effort levels. Available in Devin Desktop and CLI, with rollout to Devin Web and Fusion.
PrismML
A ternary 2B model based on Qwen3.5, built for local agents with reasoning, visual understanding, and tool use.
PrismML
A ternary 4B model based on Qwen3.5, built for local agents with reasoning, visual understanding, and tool use.
PrismML
A ternary 9B model based on Qwen3.5, built for local agents with reasoning, visual understanding, and tool use.
Paper Instruments
An agentic model for professional knowledge work, post-trained from GLM-5.3 Flash with reinforcement learning on non-verifiable tasks. The announcement presents research results without a public model download or API release.
An occasional email when notable AI dev tools and models land in the directory. No spam, unsubscribe anytime.