The frontier

A field
in motion.

General reasoning, multimodal intelligence, and autonomous agents. The models moving research forward.

22 labs & model teams2026.10.12

Global frontier labs

01OpenAI · GPTGPT-6 Astra / GPT-6.1 Sol / GPT-6 LunaReasoning that carries through to action

Model family

2026-09-29

Progress: GPT-6 Astra, introduced on September 3, combines scientific reasoning, software engineering, and computer use in multi-step workflows. GPT-6.1 Sol followed on September 29, bringing much of that capability to a lower-cost tier, alongside the high-throughput GPT-6 Luna. OpenAI reports that Sol matches Astra on DeepSWE v1.1 at roughly one-fifth the cost, while improving computer use and scientific workflows over GPT-6 Sol.
Significance: Reasoning combined with sustained tool use and cheaper experimentation connects model capability to end-to-end coding and scientific workflows.

GPT-6 Astra: A new generation of intelligence Introducing GPT-6.1 Sol GPT-6.1 Sol system card addendum

02Google DeepMind · GeminiGemini 3.8 Flash / 3.1 Pro / Omni 1.1 FlashMultimodal intelligence meets agentic work

Model family

2026-09-24 · model-card update

Progress: Gemini spans 3.8 Flash, 3.1 Pro, and 3.5 Flash-Lite, with 3.8 Audio and Omni 1.1 Flash extending the family beyond text. Gemini 3.8 Flash accepts text, images, audio, and video in a million-token context window. Its model card reports improved software engineering and agentic knowledge work, including 73.7% on DeepSWE v1.1 versus 65.3% for 3.7 Flash. Omni adds conversational video creation and editing.
Significance: Connecting perception, reasoning, and useful outputs gives agents a richer basis for understanding changing environments and completing complex projects.

Google DeepMind model cards Gemini 3.8 Flash model card Gemini Omni Flash and Omni 1.1 Flash model card

03Anthropic · ClaudeOpus 5.5 / Sonnet 5.5 / Haiku 5.5 / Fable 5.1 / Mythos 5.1Long-running work and scientific workflows

Model family

2026-10-07

Progress: Claude's current range includes Opus 5.5, Sonnet 5.5, and Haiku 5.5, alongside Fable 5.1 and trusted-access Mythos 5.1. Opus targets demanding work, Sonnet everyday throughput, and Haiku high-volume applications. Anthropic reports Opus 5.5 reaches Fable 5.1-level performance on most work while reducing typical running costs by 40% versus Opus 5. Its Fable/Mythos release also describes stronger scientific workflows and GPU optimizations for computational biology.
Significance: Sustained investigation, coding, testing, and coordination extend what models can accomplish across a whole project. Lower running costs bring more of that capability into everyday research and engineering.

Anthropic release announcements Introducing Claude Opus 5.5 Introducing Claude Fable 5.1 and Claude Mythos 5.1

04xAI / GrokGrok 4.7Long-horizon agents and verification

Model family

2026-09-21

Progress: Released on September 21, 2026, Grok 4.7 combines a larger base model with longer reinforcement learning on difficult, multi-hour tasks. Its release report emphasizes stronger self-checking, context management, and native training for the Grok Bot environment, alongside improvements in coding and professional deliverables.
Significance: Preserving a goal, inspecting intermediate results, and working through demanding workflows extends the usefulness of autonomous agents. Retaining Grok 4.6’s standard pricing makes longer problem-solving more practical to run.

Introducing Grok 4.7

05MetaMuse Spark 1.3 / Llama 4Personal agents and the open-model ecosystem

Model family

2026-09-02

Progress: Meta released Muse Spark 1.3 on September 2, 2026, strengthening long-horizon coding, task switching, and tool-driven work across multiple agent environments. Meta’s engineering comparisons report about 20% fewer tool calls and 25% fewer tokens than Spark 1.2. Its earlier Llama 4 Scout and Maverick release, dated April 5, 2025, established downloadable, natively multimodal mixture-of-experts models.
Significance: Muse develops the continuity and coordination needed for useful personal agents. Llama gives developers a foundation to adapt, investigate, and extend, connecting capable individual systems with a broader experimentation ecosystem.

Introducing Muse Spark 1.3 The Llama 4 herd

06Mistral AIMistral Large 4 PreviewMultimodal reasoning and an open-weight roadmap

Model family

2026-10-06

Progress: Mistral Large 4 entered public preview on October 6, 2026. Its model documentation describes a multimodal mixture-of-experts system with 1.05 trillion total parameters, 52 billion active parameters, and a one-million-token context window. The launch details reinforcement learning across scientific problems, verification, and long-horizon tool use; downloadable weights are scheduled for the end of October.
Significance: Visual understanding, executable research workflows, and adaptable deployment form a useful combination for advanced agents. The planned weight release could give more teams the ability to operate, customize, and investigate those systems.

Introducing Mistral Large 4 Mistral Large 4 model documentation

07AmazonNova 2.5 Sonic / Nova 2 LiteVoice agents and deployable intelligence

Model family

2026-10-05

Progress: Amazon released Nova 2.5 Sonic on October 5, 2026, improving reasoning, instruction following, and tool calling in real-time speech-to-speech interaction. Alongside it, Nova 2 Lite offers adjustable reasoning effort, code execution, web grounding, remote MCP tools, and a one-million-token context window, with customization through Nova Forge.
Significance: Real-time voice brings reasoning into situations where typing is impractical. Tool integration, adjustable inference, and customization connect those conversations to repeatable professional workflows and domain-specific agents.

Announcing Amazon Nova 2.5 Sonic Amazon Nova foundation models

08NVIDIANemotron 3 UltraEfficient reasoning and agent orchestration

Model family

2026-06-04

Progress: NVIDIA released Nemotron 3 Ultra on June 4, 2026: an open 550-billion-parameter model with 55 billion active parameters and context support up to one million tokens. Its hybrid Mamba–attention architecture, multi-token prediction, and multi-teacher on-policy distillation target efficient reasoning and long-running agents. NVIDIA also publishes checkpoints, training recipes, and reinforcement-learning resources.
Significance: Combining specialized teachers into a broader reasoner offers a concrete way to consolidate expertise. Efficient inference and open training resources make sustained reasoning and agent coordination easier for other teams to run, reproduce, and improve.

Nemotron 3 Ultra launch and technical overview NVIDIA Nemotron 3 Ultra official model card Nemotron 3 Ultra technical report

09Thinking Machines LabInkling / Inkling-Small · TinkerCustomizable multimodal intelligence

Model family

2026-07-30

Progress: Thinking Machines Lab released the open-weight Inkling model on July 15, 2026, followed by Inkling-Small on July 30. Inkling has 975 billion total and 41 billion active parameters; Small reduces those figures to 276 billion and 12 billion. Both support native image and audio reasoning and adjustable thinking effort. Tinker is the company’s training API, giving researchers control over data, loss functions, and post-training algorithms while managing distributed compute.
Significance: Multimodal foundations provide breadth, while flexible training tools let domain experts turn their own feedback and task knowledge into specialized capabilities. This connects model development with practical experimentation across many different fields.

Inkling: Our Open-Weights Model Introducing Inkling-Small Tinker training API documentation

10MBZUAI · Institute of Foundation ModelsK2 Horizon 375B-A23BReproducible agents across model scales

Model family

2026-09-03

Progress: MBZUAI’s Institute of Foundation Models released K2 Horizon on September 3, 2026. Its flagship is K2 Horizon 375B-A23B, a sparse model with approximately 23 billion active parameters and a 512K context window, within a six-model family spanning 0.9B to 375B. The release includes model weights, intermediate checkpoints, training code, and data or data-construction recipes, with models and code under Apache 2.0.
Significance: Intermediate checkpoints and training resources let researchers study how reasoning and tool use develop, compare stages of learning, and transfer methods between deployment scales. This shared experimental foundation can accelerate progress across independent teams.

Introducing K2 Horizon K2 Horizon 375B-A23B official model card MBZUAI announces K2 Horizon

Chinese model teams

11DeepSeekDeepSeek-V4.1-FlashEfficient long-horizon intelligence

Model family

2026-09-10

Progress: DeepSeek-V4.1-Flash, released September 10, 2026, combines native vision and a million-token context with a 552B-parameter mixture of experts. Its causal encoder–decoder activates 8B parameters when reading inputs and 16B when generating outputs. Cross-layer cache reuse and FP4 caching reduce the global KV-cache footprint to roughly one quarter of V4-Flash's; a separate deployment optimization cuts persistent cache storage to about one eighth. Model checkpoints are publicly available.
Significance: DeepSeek advances the economics of sustained intelligence. Cheaper context processing can make longer investigations, richer tool histories, and repeated verification practical. Its architectural work and released weights give other teams concrete building blocks for scaling reasoning and autonomous workflows.

Introducing DeepSeek-V4.1-Flash DeepSeek-V4.1-Flash technical report

12Alibaba · QwenQwen3.8-2.4T-A95B / Omni-FlashOpen reasoning across text, audio and video

Model family

2026-09-22 · Omni report

Progress: Qwen's August 12, 2026 release opened Qwen3.8-2.4T-A95B, a 2.4T-parameter model with 95B active parameters. Its model card describes adjustable reasoning effort and retained reasoning context for multi-step work. The September 22 Qwen3.8-Omni report extends the story to native audio-video agents: Omni-Flash uses multimodal co-training and a million-token context, alongside open frameworks for media tools and real-time agent orchestration.
Significance: Qwen connects frontier reasoning with a broad, reusable development ecosystem. Opening a flagship-class model widens the field for experimentation, while carrying planning and tool use across media moves assistants toward handling complete, mixed-format projects. The important direction is intelligence that can perceive, coordinate, and deliver across everyday digital environments.

Qwen3.8 official repository Qwen3.8-2.4T-A95B model card Qwen3.8-Omni technical report

13Moonshot AI · KimiKimi K3Open infrastructure for agentic work

Model family

2026-07-27

Progress: Kimi K3 arrived in July 2026, with weights and training infrastructure released on July 27. The 2.8T-parameter mixture-of-experts model activates 104B parameters and supports native vision and a million-token context. Kimi Delta Attention, Attention Residuals, and reinforcement learning across reasoning, coding, and general-agent tasks underpin its long-horizon work. Moonshot also released MoonEP, FlashKDA, and AgentEnv, covering expert communication, attention kernels, and agent-training environments.
Significance: Kimi pushes the open frontier toward sustained, tool-using work. The combination of visual understanding, long context, and agent training is especially relevant to software and research workflows. Sharing both weights and infrastructure gives the wider research community more of the machinery needed to improve these systems.

Kimi K3 technical blog Kimi K3 weights and infrastructure release Kimi K3 technical report

14Z.ai · GLMGLM-5.3 / GLM-5.3-FlashAgentic engineering and efficient inference

Model family

2026-08-26 · Flash update

Progress: GLM-5.3 improves complex coding and long-horizon tasks through post-training on the same base model as GLM-5.2. The August 26, 2026 GLM-5.3-Flash release adds a newly trained, natively multimodal foundation: 320B total parameters, 18B active, hybrid sparse-plus-linear attention, and manifold-constrained hyper-connections. Its MIT-licensed weights make the model deployable beyond Z.ai's service, while visual input helps agents inspect interfaces and the results of their own code.
Significance: GLM highlights two complementary levers: teaching models to sustain difficult work, and making that work cheaper to run. Closing the loop between code, visual feedback, and revision is a meaningful step toward more capable engineering agents and iterative problem-solving systems.

GLM official release notes GLM-5 official repository GLM-5.3-Flash model card

15ByteDance · Doubao / SeedSeed2.1 / SeedRealtime / Seedance 2.5Agents that see, listen and act

Model family

2026-08-05

Progress: Seed2.1, released June 23, strengthens cross-tool execution, repository-level coding and computer use; ByteDance also describes agents assisting model evaluation, data synthesis and training infrastructure. SeedRealtime, announced August 5, jointly processes live audio, video and text, with proactive speech and tool calls. Seedance 2.5, released July 31, extends joint audio-video generation to 30-second clips with multimodal references and targeted editing.
Significance: A tighter perception–action–research loop connects the different parts of an intelligent workflow. Real-time grounding makes agents more responsive to changing environments, while research assistance can accelerate the next round of model development.

Introducing Seed2.1 Introducing SeedRealtime Introducing Seedance 2.5

16MiniMaxM3 / H3 / Speech 2.8Long-context multimodal agency

Model family

2026-07-31

Progress: M3, released June 1, uses MiniMax Sparse Attention for a one-million-token context and natively accepts images and video, alongside coding and desktop operation. MiniMax reports autonomous paper-reproduction and kernel-optimization experiments. H3, released July 31, conditions on text, images, video and audio to generate up to 15 seconds of 2K video with stereo sound. Speech 2.8 adds native tags for expressive vocal details.
Significance: Sustained, feedback-driven work becomes more practical as inference gets cheaper. Keeping observations, code and experiment history together can extend the depth of autonomous research, while multimodal generation widens what those systems can produce.

Introducing MiniMax M3 Introducing MiniMax H3 Introducing MiniMax Speech 2.8

17Baidu · ERNIEERNIE 5.1 / 5.0Elastic models, scalable agent learning

Model family

2026-05-09

Progress: The ERNIE 5.0 technical report describes shared autoregressive training across text, images, video and audio, using modality-agnostic expert routing. ERNIE 5.1's May 9 release builds on that foundation: Baidu reports roughly one-third of 5.0's total parameters and half its active parameters. Its training stack separates agent rollouts, inference, rewards and training, then integrates specialist capabilities through multi-teacher on-policy distillation.
Significance: Capability reuse and scalable learning extend the reach of intelligent agents. A shared multimodal foundation can transfer knowledge between tasks, while elastic subnetworks and asynchronous training make repeated agent improvement less resource-intensive.

Introducing ERNIE 5.1 ERNIE 5.0 technical report

18Tencent · Hunyuan / HyHy3 / HY-World 2.0Efficient agents and persistent 3D worlds

Model family

2026-07-06

Progress: Hy3, officially released July 6, is a hybrid-thinking MoE with 295 billion total and 21 billion active parameters, a 256K context and Apache 2.0 licensing. Tencent emphasizes software work, task planning and tool coordination. Separately, HY-World 2.0's April 16 release describes a multimodal pipeline that generates or reconstructs navigable scenes as editable meshes and Gaussian splats, rather than only rendering transient frames.
Significance: Efficient open models broaden experimentation, while persistent 3D representations provide a useful substrate for spatial learning and planning. Together, these research directions expand the environments in which future intelligent agents can be developed and tested.

Tencent Hy3 release HY-World 2.0 official repository

19iFLYTEK · SparkSpark-X2.5-4B / 1.7BLong-context agents on smaller models

Open-model family

2026-09-01

Progress: The Spark-X2.5-4B and 1.7B series launched globally on September 1. Its architecture alternates one full-attention layer with three sliding-window layers, supporting a native one-million-token context and over 200 languages. The team reports training on Huawei Ascend clusters, then reinforcement learning and multi-teacher on-policy distillation to consolidate specialist abilities. Apache 2.0 weights and multiple inference-backend integrations support local experimentation.
Significance: Smaller agents make repeated tool interactions, local adaptation, and distributed experimentation more accessible. This expands the range of devices and teams that can participate in developing more capable autonomous systems.

Spark-X2.5 open-model repository Spark-X2.5-4B model card

20Xiaomi · MiMoMiMo-V2.6-ProScaled reinforcement learning · Omnimodal agents · Open research

Open-model family

2026-09-22

Progress: Released on September 22, 2026, MiMo-V2.6-Pro brings text, image, video, and audio understanding into a million-token context. Its open Pro-RL checkpoint has 1.02T total parameters and 42B active parameters. Xiaomi scales reinforcement learning across mixed tasks and agent frameworks, pairing larger training runs with comparative grading and verifier cross-checks. The release also opens weights, training environments, and reinforcement-learning resources for further research.
Significance: MiMo makes the learning loop itself a central contribution: richer environments generate experience, while better feedback helps select useful behavior. Combined with multimodal perception, this is a promising route toward agents that can investigate, act, and improve through structured training. Opening the surrounding research tools makes that route easier for others to examine and extend.

MiMo-V2.6: Scaling Reinforcement Learning Toward Self-Improvement MiMo-V2.6-Pro-RL model card MiMo Code v0.1.15 release

21StepFun · StepStep 5 PreviewEfficient agentic work · Long-horizon execution · Visual reasoning

Model preview

2026-09-20

Progress: Step 5 Preview, announced September 20, 2026, is a 600B-parameter sparse mixture-of-experts model with 27B active parameters and a million-token context. It accepts text, images, and video, supports tool calls and adjustable reasoning effort, and targets software engineering and professional knowledge work. StepFun's demonstrations include repeated GPU-kernel optimization and revisions to a post-training data pipeline. The model is available through products and API; open weights are scheduled for October 15.
Significance: StepFun emphasizes sustained work at practical computational cost. Long context and visual feedback can help an agent keep track of a project, test changes, and revise its approach. Its model-improvement experiments also point toward AI becoming a more useful participant in the engineering work that advances future systems.

Step 5 Preview: Advancing the Pareto Frontier StepFun announces flagship Step 5 Preview Step 5 Preview — StepFun Documentation

Superintelligence research

22Safe SuperintelligenceMission and infrastructureA singular focus on safe superintelligence

Mission and infrastructure

2026-07-26

Progress: The official mission statement and a July 2026 update on compute partnerships.
Significance: SSI brings capability development and control into the same research agenda, making safety part of the architecture of future intelligent systems.

The mission to build safe superintelligence Research infrastructure and official updates

Ideas & methods

Global frontier labs

Chinese model teams

Superintelligence research

The research behind the progress.