Model Atlas

Explore

Models, agents, hardware and technology, in context.

20072026327 events

LLM154Agent50Hardware54Technology69

September 202617 events

Gemini 4 Argon

GoogleLLM

A new multimodal model for complex coding, knowledge work, and cyber defense, initially available to invited teams.

Source

GPT-6.1 Sol

OpenAILLM

Improves coding, computer use, and professional work, with task delegation available within a request.

Source

Claude Sonnet 5.5

AnthropicLLM

The second Claude 5.5 model supports coding, document processing, and office tasks.

Source
3 events

Claude Opus 5.5

AnthropicLLM

Opus targets complex knowledge work, coding, and judgment-intensive tasks.

Source

Grok 4.7

xAILLM

Grok improves coding, self-verification, and context management.

Source

Step 5 Preview

StepFunLLM

A flagship preview targets long-running engineering, knowledge work, and research.

Source

Claude / Cowork

AnthropicAgent

Cowork and chat merge into one Claude experience for documents and ongoing tasks.

Source

GPT-6 Astra

OpenAILLM

GPT combines reasoning, computer use, and multi-step professional work.

Source
3 events

Qwen3.8-Max-0902

AlibabaLLM

Improves engineering projects, multi-tool orchestration, chart reasoning, and document parsing.

Source

August 202610 events

2 events

GLM-5.3-Flash

Z.aiLLM

Native vision and hybrid linear/sparse attention support coding, browser, and GUI tasks.

Source

Qwen3.8-Flash

AlibabaLLM

An efficient multimodal model supports coding and high-volume agent use.

Source

GLM-5.3

Z.aiLLM

GLM strengthens complex coding and extended task execution.

Source
3 events

DeepSeek Harness

DeepSeekAgent

Plugins compose models, tools, sessions, sandboxes and execution loops, with a recorded task trajectory.

Source
2 events

Grok 4.6

xAILLM

Grok improves sustained work across agents, interactive apps, and vision.

Source

Qwen3.8-Max

AlibabaLLM

A native vision-language flagship supports a one-million-token context.

Source

July 20269 events

Atlas 950 SuperPoD

HuaweiHardware

A 1,024-card SuperPoD is publicly demonstrated with unified addressing and high-speed links.

Source
2 events

Grok 4.5

xAILLM

Grok advances coding, agent tasks, and professional knowledge work.

Source

Kimi K3

Moonshot AILLM

Kimi combines native vision, long context, and complex coding.

Source
2 events

GPT-5.6

OpenAILLM

Sol, Terra, and Luna address different knowledge-work and coding needs.

Source

Hunyuan Hy3

TencentLLM

Tencent improves long-running agents and product integration.

Source

June 20265 events

Claude Sonnet 5

AnthropicLLM

Sonnet expands planning, browser use, terminal work, and sustained execution.

Source

GLM-5.2

Z.aiLLM

GLM extends long-running task and engineering capabilities.

Source

Kimi K2.7 Code

Moonshot AILLM

Kimi improves long-running coding, tool calls, and efficient thinking.

Source

MiniMax-M3

MiniMaxLLM

Sparse attention adds native image and video understanding.

Source

May 20265 events

Qwen3.7-Max

AlibabaLLM

Qwen improves agentic coding, office work, and long-running execution.

Source

April 20269 events

2 events

GPT-5.5

OpenAILLM

GPT improves tasks spanning multiple tools and applications.

Source

Muse Spark

MetaLLM

Meta introduces a new family combining multimodality, reasoning, and tools.

Source

GLM-5.1

Z.aiLLM

GLM improves long-running engineering, planning, and execution.

Source

Gemma 4

GoogleLLM

Gemma advances open-model reasoning and agent capabilities.

Source

March 20268 events

2 events

MiniMax-M2.7

MiniMaxLLM

Software engineering and professional work gain stronger agent collaboration.

Source

Mistral Small 4

MistralLLM

One model combines reasoning, vision, and coding with adjustable thinking.

Source
2 events

GPT-5.4

OpenAILLM

Reasoning, coding, and native computer use target professional work.

Source

FlashAttention-4

FlashAttentionTechnology

Algorithm and kernel-pipeline co-design adapts attention to Blackwell’s compute and bandwidth characteristics.

Source

WorkBuddy

TencentAgent

A desktop workspace agent handles files, content and coding tasks with tools and extensible skills.

Source

February 202612 events

Hermes Agent

Nous ResearchAgent

A persistent agent combines memory, reusable skills and tools to carry out tasks locally or on a server.

Source

Gemini 3.1 Pro

GoogleLLM

Gemini strengthens reasoning for scientific, engineering, and complex tasks.

Source

Qwen3.5

AlibabaLLM

Native multimodality and agent capabilities come together in a large expert model.

Source

Seed 2.0

ByteDanceLLM

ByteDance introduces multimodal models for real-world agent tasks.

Source
3 events

GLM-5

Z.aiLLM

GLM moves from individual coding tasks toward broader software engineering.

Source

MiniMax-M2.5

MiniMaxLLM

Reinforcement learning targets practical coding, search, office work, and tools.

Source
2 events

GPT-5.3-Codex

OpenAILLM

Coding and general reasoning combine in longer engineering workflows.

Source
2 events

Codex app

OpenAIAgent

A desktop app organizes parallel coding agents and code review.

Source

Step 3.5 Flash

StepFunLLM

Sparse mixture-of-experts modeling balances speed with agent capabilities.

Source

January 20265 events

OpenClaw

OpenClawAgent

The personal-agent project adopts the OpenClaw name across messaging and local tools.

Source

Kimi K2.5

Moonshot AILLM

Kimi combines visual understanding, coding, and agent collaboration.

Source

Claude Cowork

AnthropicAgent

Claude Code-style execution expands to files, documents, and office work.

Source

December 202511 events

MiniMax-M2.1

MiniMaxLLM

MiniMax improves programming across languages and practical office tasks.

Source

GLM-4.7

Z.aiLLM

GLM improves coding and long-context task execution.

Source

MiMo-V2-Flash

XiaomiLLM

Xiaomi opens an efficient model for reasoning, coding, and agent tasks.

Source

GPT-5.2

OpenAILLM

GPT advances professional knowledge work, coding, and long-context tasks.

Source
4 events

Kiro autonomous agent

AmazonAgent

Handles software tasks asynchronously in isolated environments, retaining context from project history and review.

Source

Mistral 3

MistralLLM

Mistral expands its open-weight multimodal model family.

Source

DeepSeek-V3.2

DeepSeekLLM

DeepSeek advances efficient attention, reasoning, and agentic tool use.

Source

November 20257 events

Pi coding agent

Mario ZechnerAgent

A minimal coding harness emphasizes tools, context control, and session management.

Source
2 events

Google Antigravity

GoogleAgent

Agents work across the editor, terminal and browser, presenting artifacts that document progress and verification.

Source

Gemini 3 Pro

GoogleLLM

Google introduces a new generation of multimodal reasoning models.

Source

Kiro CLI / IDE GA

AmazonAgent

Kiro reaches general availability and introduces a CLI for specification-driven development in the terminal.

Source

GPT-5.1

OpenAILLM

Instant and Thinking improve instructions, style, and adaptive reasoning.

Source

October 20258 events

OpenCode 1.0

AnomalyAgent

An open-source coding agent connects to different models to read files, edit code and use project tools.

Source

MiniMax-M2

MiniMaxLLM

MiniMax opens a model designed around coding and agent workflows.

Source

Kimi CLI

Moonshot AIAgent

A publicly released terminal coding agent connecting shell commands, project files and MCP tools.

Source

CrewAI 1.0

CrewAIAgent

Organizes agents around roles, tasks and flows as its open-source core reaches general availability.

Source
2 events

Agent Skills

AnthropicAgent

Instructions, scripts, and resources package reusable, on-demand agent expertise.

Source

Manus 1.5

ManusAgent

Handles research, analysis, presentations and website development, including full-stack web applications.

Source

NVIDIA DGX Spark

NVIDIAHardware

GB10 and unified memory move into a desktop system for local model development.

Source

September 20254 events

GLM-4.6

Z.aiLLM

GLM improves code, reasoning, and tool calls.

Source
2 events

Claude Agent SDK

AnthropicAgent

Claude Code SDK becomes Claude Agent SDK, expanding its focus to general-purpose agents.

Source

Qwen3-Next

AlibabaLLM

Hybrid attention and highly sparse activation target efficient long-context processing.

Source

August 20255 events

2 events

Qoder

AlibabaAgent

Plans and executes development tasks with repository context; Quest mode supports longer delegated work.

Source

DeepSeek-V3.1

DeepSeekLLM

Thinking and non-thinking modes are unified with stronger agent capabilities.

Source

GPT-5

OpenAILLM

A new GPT generation advances reasoning, coding, and everyday work.

Source
2 events

July 20259 events

Deep Agents

LangChainAgent

A reusable harness combines planning, a filesystem, and subagents for longer tasks.

Source

GLM-4.5

Z.aiLLM

GLM combines reasoning, coding, and agentic workflows.

Source
2 events

Qwen Code

AlibabaAgent

An open-source command-line coding agent released alongside Qwen3-Coder for repository work and task execution.

Source

Qwen3-Coder

AlibabaLLM

An open mixture-of-experts model targets agentic coding and long tasks.

Source
2 events

ChatGPT agent

OpenAIAgent

Research and action come together through a virtual computer, browser, and terminal.

Source

TRAE SOLO

ByteDanceAgent

Connects requirements, coding, browser validation and deployment within one development task.

Source

Kiro

AmazonAgent

An IDE organizes requirements, design and task lists before agents edit code, with event-triggered automation.

Source

Kimi K2

Moonshot AILLM

Kimi introduces open weights focused on coding and agentic tool use.

Source

Grok 4

xAILLM

xAI advances reasoning and tool-assisted problem solving.

Source

June 20255 events

Gemini CLI

GoogleAgent

Google releases an open-source terminal agent for coding, troubleshooting, and tools.

Source

Warp 2.0

WarpAgent

Expands the terminal into an agent development environment for managing tasks, code changes and commands.

Source

MiniMax-M1

MiniMaxLLM

Hybrid attention explores efficient reasoning and long-context work.

Source
2 events

ROCm 7

AMDTechnology

AMD previews ROCm 7 with updated software support for Instinct accelerators.

Source

May 20257 events

DeepSeek-R1-0528

DeepSeekLLM

A major R1 update improves reasoning, coding, structured output, and tool calls.

Source

Jules

GoogleAgent

An asynchronous agent reads repositories, edits code and runs tests in a cloud VM, then presents changes for review.

Source
2 events

Codex Cloud

OpenAIAgent

Cloud sandboxes support parallel engineering tasks and reviewable code changes.

Source

Strands Agents

AmazonAgent

Define an agent with a model, instructions and tools, letting the model choose the next steps.

Source

Mistral Medium 3

MistralLLM

An enterprise model balances multimodal capability, coding, and deployment cost.

Source

April 20259 events

Qwen3

AlibabaLLM

Thinking and non-thinking modes share one model family.

Source
2 events

Codex CLI

OpenAIAgent

An open-source terminal coding agent works directly with local code and tools.

Source

GPT-4.1

OpenAILLM

An API model family improves coding, instructions, and long-context processing.

Source
3 events

Agent2Agent · A2A

GoogleAgent

An open protocol supports agent discovery, task delegation and status updates across vendors.

Source

Llama 4

MetaLLM

Scout and Maverick introduce multimodal mixture-of-experts models.

Source

March 20257 events

Gemini 2.5 Pro

GoogleLLM

Thinking becomes central to complex reasoning, coding, and multimodal tasks.

Source

Hunyuan T1

TencentLLM

Tencent introduces a deep-reasoning model trained with reinforcement learning.

Source
2 events

NVIDIA Dynamo

NVIDIATechnology

An open-source framework coordinates distributed model serving across GPUs.

Source

Gemma 3

GoogleLLM

The open-weight family expands image understanding and long-context processing.

Source

OpenAI Agents SDK

OpenAIAgent

Agent orchestration adds handoffs, guardrails, and tracing around model and tool calls.

Source

Apple M3 Ultra

AppleHardware

Larger unified-memory configurations increase capacity for running large models locally.

Source

February 20255 events

GPT-4.5

OpenAILLM

A research preview explores further gains from scaling pretraining.

Source
2 events

Claude Code

AnthropicAgent

A terminal agent reads code, edits files, runs tests, and uses command-line tools.

Source

Grok 3 Beta

xAILLM

xAI describes reasoning capabilities and the DeepSearch experience.

Source

January 20256 events

Goose

BlockAgent

A local desktop and command-line agent carries out tasks through connected tools and extensions.

Source

vLLM V1

vLLMTechnology

A rebuilt inference core simplifies scheduling and improves prefix caching and execution.

Source

Operator

OpenAIAgent

A research preview carries out tasks through a browser interface.

Source

DeepSeek-R1

DeepSeekLLM

DeepSeek releases a reasoning model and distilled variants with downloadable weights.

Source

GeForce RTX 5090

NVIDIAHardware

The Blackwell consumer flagship expands to 32 GB of GDDR7 for local AI and graphics workloads.

Source

December 20249 events

smolagents

Hugging FaceAgent

Models express actions as code and use tools and execution environments to complete multi-step tasks.

Source

DeepSeek-V3

DeepSeekLLM

Open weights expand large-scale mixture-of-experts modeling.

Source

Phi-4

MicrosoftLLM

Synthetic data and training improvements strengthen small-model reasoning.

Source

Llama 3.3

MetaLLM

A revised 70B instruction model improves multilingual dialogue and deployment efficiency.

Source

OpenAI o1

OpenAILLM

The full o1 model launches with ChatGPT Pro and adds image understanding.

Source
2 events

Amazon Nova

AmazonLLM

Amazon introduces text and multimodal models across several cost tiers.

Source

November 20243 events

Cursor Agent

AnysphereAgent

Composer gains an early agent that retrieves context and uses the terminal.

Source

October 20242 events

September 20244 events

Llama 3.2

MetaLLM

Vision models and small on-device text models join the Llama family.

Source

Qwen2.5

AlibabaLLM

General, coding, and math models improve instruction following and structured output.

Source

Replit Agent

ReplitAgent

Natural-language app building combines setup, coding, and execution.

Source

July 20245 events

2 events

Llama 3.1

MetaLLM

A 405B model extends the family with longer context, multilingual support, and tool use.

Source

GPT-4o mini

OpenAILLM

A lower-cost text and vision model expands lightweight applications.

Source

FlashAttention-3

FlashAttentionTechnology

Attention kernels are redesigned around Hopper GPUs’ asynchronous execution and low-precision features.

Source

June 20243 events

Gemma 2

GoogleLLM

A redesigned architecture introduces efficient 9B and 27B open-weight models.

Source

Claude 3.5 Sonnet

AnthropicLLM

Sonnet improves coding and visual understanding alongside the introduction of Artifacts.

Source

Qwen2

AlibabaLLM

A range of model sizes improves multilingual, coding, math, and long-context performance.

Source

May 20246 events

GPT-4o

OpenAILLM

Text, vision, and audio move toward a unified, more natural interface.

Source
3 events

SWE-agent

SWE-agent teamAgent

A purpose-built computer interface lets agents navigate code, edit files, and run tests.

Source

DeepSeek-V2

DeepSeekLLM

Mixture-of-experts and latent attention improve training and inference efficiency.

Source

E2B Code Interpreter SDK

E2BTechnology

Packages isolated code execution into an SDK so agents can run generated programs and inspect their results.

Source

April 20244 events

Phi-3

MicrosoftLLM

Microsoft introduces small language models suited to local deployment.

Source

Llama 3

MetaLLM

8B and 70B weights improve general-purpose open-model capabilities.

Source

Intel Gaudi 3

IntelHardware

Gaudi 3 combines matrix compute, HBM, and Ethernet links for large-model training and inference.

Source

March 20244 events

Cerebras WSE-3 / CS-3

CerebrasHardware

The third wafer-scale engine integrates a large compute array and on-chip memory on one wafer.

Source

Devin

CognitionAgent

A sandboxed agent plans engineering tasks using an editor, terminal, and browser.

Source

Claude 3

AnthropicLLM

Opus, Sonnet, and Haiku form a new vision-capable model family.

Source

February 20243 events

Gemma

GoogleLLM

Google introduces lightweight open-weight models for local development and research.

Source

Gemini 1.5

GoogleLLM

Gemini 1.5 Pro demonstrates million-token context processing.

Source

January 20241 event

LangGraph

LangChainAgent

Stateful graphs and cycles organize controllable agent execution.

Source

December 20237 events

SGLang

SGLangTechnology

A structured generation language and runtime reuse prefix computation and cached state.

Source

Mixtral 8×7B

MistralLLM

A sparse mixture-of-experts model expands efficient open-weight inference.

Source
3 events

Google TPU v5p

GoogleHardware

A fifth-generation TPU focused on large-model training is announced with AI Hypercomputer.

Source

Gemini 1.0

GoogleLLM

Google introduces the multimodal Gemini family: Ultra, Pro, and Nano.

Source

MLX

AppleTechnology

An array framework brings automatic differentiation and unified-memory execution to Apple silicon.

Source

Mamba

State SpacesTechnology

Selective state-space models offer an alternative approach to sequence modeling.

Source

November 20233 events

NVIDIA H200

NVIDIAHardware

Hopper gains HBM3e to increase memory capacity and bandwidth for large-model inference.

Source

GPT-4 Turbo

OpenAILLM

A developer preview expands context to 128K and advances vision and tool use.

Source

Grok

xAILLM

xAI announces Grok with early access and information from X.

Source

October 20231 event

TensorRT-LLM

NVIDIATechnology

An inference library for large language models on NVIDIA GPUs becomes publicly available.

Source

September 20232 events

AutoGen

MicrosoftAgent

Orchestrate workflows through conversations between models, tools, and people.

Source

August 20233 events

Google TPU v5e

GoogleHardware

The cost-focused fifth-generation TPU supports both training and inference.

Source

Ascend 910B

HuaweiHardware

Powers the Spark all-in-one system introduced by iFLYTEK and Huawei for enterprise LLM deployment.

Source

Qwen-7B

AlibabaLLM

Qwen releases 7B base and chat model weights.

Source

July 20233 events

Llama 2

MetaLLM

Pretrained and chat-tuned weights become available for research and commercial use.

Source

FlashAttention-2

FlashAttentionTechnology

Revised work partitioning improves GPU parallelism for attention computation.

Source

Claude 2

AnthropicLLM

Improved coding, reasoning, and long-document processing arrive with a public chat beta.

Source

June 20232 events

AWQ

MITTechnology

Activation statistics protect salient weights for low-bit model deployment.

Source

May 20234 events

DPO

StanfordTechnology

Direct optimization on preference data simplifies language-model alignment.

Source

Aider

AiderAgent

Terminal-based code editing uses a repository map to give the model context across files.

Source

QLoRA

U. WashingtonTechnology

Low-bit quantization and low-rank adaptation reduce memory requirements for model fine-tuning.

Source

April 20233 events

LLaVA

LLaVA TeamLLM

Visual instruction tuning connects an image encoder and a language model for multimodal conversation.

Source

Auto-GPT 0.2

Significant GravitasAgent

An early autonomous agent uses repeated model calls and tools, adding Selenium-based web browsing in this release.

Source

March 20236 events

NVIDIA L4

NVIDIAHardware

A low-power Ada GPU succeeds T4 for generative AI and video inference.

Source

PyTorch 2.0

PyTorchTechnology

torch.compile adds compilation-based acceleration to the established PyTorch workflow.

Source
2 events

Claude

AnthropicLLM

Anthropic introduces Claude and Claude Instant.

Source

GPT-4

OpenAILLM

Stronger language capabilities are joined by demonstrated image understanding.

Source

Stanford Alpaca

StanfordLLM

Synthetic instruction data adapts LLaMA into a small instruction-following model.

Source

llama.cpp

ggmlTechnology

A C/C++ implementation runs LLaMA locally with quantized inference.

Source

February 20231 event

LLaMA

MetaLLM

Meta introduces a family of foundation models for research access.

Source

November 20222 events

2 events

ChatGPT

OpenAILLM

A GPT-3.5-based research preview supports multi-turn conversation.

Source

Speculative Decoding

GoogleTechnology

A small model drafts tokens and a target model verifies them, accelerating generation while preserving the target distribution.

Source

October 20224 events

GPTQ

ISTATechnology

Approximate second-order information supports post-training weight quantization for lower-memory inference.

Source
2 events

Flow Matching

MetaTechnology

Vector fields along conditional probability paths train continuous-flow generative models.

Source

ReAct

GoogleTechnology

Reasoning steps and tool actions are interleaved so language models can adapt their plans to environment feedback.

Source

September 20221 event

GeForce RTX 4090

NVIDIAHardware

Ada and 24 GB of memory support local inference, image generation, and model development.

Source

July 20221 event

BLOOM

BigScienceLLM

BigScience releases a 176B multilingual model with weights and training artifacts.

Source

May 20222 events

FlashAttention

FlashAttentionTechnology

Reduced data movement between GPU memory and on-chip storage accelerates exact attention.

Source

Intel Habana Gaudi 2

IntelHardware

The second Gaudi generation moves to 7 nm while retaining Ethernet-based training scale-out.

Source

April 20221 event

PaLM

GoogleLLM

Google explores reasoning and language capabilities at larger scale.

Source

March 20224 events

NVIDIA H100

NVIDIAHardware

Hopper introduces the Transformer Engine and FP8 for large-model training and inference.

Source

Self-Consistency

GoogleTechnology

Multiple sampled reasoning paths are aggregated to improve answer consistency.

Source

January 20222 events

InstructGPT

OpenAILLM

Reinforcement learning from human feedback improves instruction following and alignment with user intent.

Source

December 20211 event

Latent Diffusion

CompVisTechnology

Diffusion in a compressed latent space reduces the cost of high-resolution image generation.

Source

November 20211 event

AMD Instinct MI250X

AMDHardware

CDNA 2 uses a multi-die design to expand memory capacity for training and scientific computing.

Source

August 20211 event

Codex

OpenAILLM

Natural-language instructions become executable code.

Source

July 20211 event

Triton 1.0

OpenAITechnology

A Python-like language and compiler simplify the development of optimized GPU kernels.

Source

June 20211 event

LoRA

MicrosoftTechnology

Frozen base weights and trainable low-rank matrices reduce the cost of model adaptation.

Source

May 20211 event

April 20212 events

2 events

January 20211 event

CLIP

OpenAITechnology

Image–text contrastive pretraining enables visual concepts to be recognized from natural-language descriptions.

Source

November 20202 events

Apple M1

AppleHardware

Unified memory and the Neural Engine come to the Mac as a platform for local machine learning.

Source

October 20201 event

September 20201 event

GeForce RTX 3090

NVIDIAHardware

24 GB of memory and Ampere Tensor Cores expand capacity for local model experiments.

Source

August 20201 event

CANN 3.0

HuaweiTechnology

An updated Ascend software stack covers operator development, model compilation, and execution.

Source

July 20201 event

June 20201 event

DDPM

UC BerkeleyTechnology

Iterative denoising provides a training approach for diffusion-based image generation.

Source

May 20203 events

GPT-3

OpenAILLM

A 175B-parameter model demonstrates broad few-shot learning.

Source

RAG

MetaTechnology

Document retrieval supplies external knowledge to a text-generation model.

Source

NVIDIA A100

NVIDIAHardware

Ampere unifies training and inference and introduces Multi-Instance GPU.

Source

February 20201 event

DeepSpeed · ZeRO

MicrosoftTechnology

Partitioning optimizer state and other training data reduces distributed-training memory use.

Source

January 20201 event

Neural Scaling Laws

OpenAITechnology

A systematic study relates language-model performance to model size, data, and compute.

Source

December 20191 event

October 20192 events

T5

GoogleLLM

A text-to-text formulation unifies language tasks within one framework.

Source

Hugging Face Transformers

Hugging FaceTechnology

A unified API makes Transformer architectures and pretrained models accessible for research and applications.

Source

September 20191 event

Megatron-LM

NVIDIATechnology

Intra-layer model parallelism distributes large Transformer training across GPUs.

Source

August 20192 events

Ascend 910

HuaweiHardware

Ascend 910 launches as a dedicated processor for neural-network training.

Source

Cerebras WSE

CerebrasHardware

The first wafer-scale engine integrates compute, memory, and interconnect on one large chip.

Source

July 20191 event

Cloud Hypervisor v0.1.0

Cloud HypervisorTechnology

A Rust-based virtual machine monitor provides a compact VM layer for modern cloud workloads.

Source

June 20191 event

Habana Gaudi

IntelHardware

An Ethernet-connected architecture targets scalable neural-network training.

Source

February 20191 event

GPT-2

OpenAILLM

Scaling improves coherent text generation and zero-shot task performance.

Source

December 20182 events

JAX

GoogleTechnology

NumPy-style programming combines with automatic differentiation and accelerator compilation.

Source

ONNX Runtime

MicrosoftTechnology

A cross-platform runtime connects model formats to execution backends.

Source

November 20182 events

Firecracker

AWSTechnology

AWS releases lightweight microVM technology to isolate short-lived, multi-tenant workloads with separate kernels.

Source

October 20182 events

BERT

GoogleLLM

Bidirectional context advances language understanding and reading comprehension.

Source

Ascend 310 / 910

HuaweiHardware

Huawei announces the Ascend chip family and Da Vinci architecture for inference and training.

Source

September 20181 event

NVIDIA Tesla T4

NVIDIAHardware

Turing Tensor Cores bring low-precision compute to data-center inference.

Source

August 20181 event

SentencePiece

GoogleTechnology

Subword models train directly on raw text without language-specific word segmentation.

Source

June 20181 event

GPT-1

OpenAILLM

Unsupervised language pretraining and task-specific fine-tuning support transfer across tasks.

Source

May 20183 events

Kata Containers 1.0

Kata ContainersTechnology

A container runtime uses lightweight VMs to give containers or pods their own guest kernel.

Source

Google TPU v3

GoogleHardware

Third-generation TPUs expand training scale with liquid-cooled pods.

Source

gVisor

GoogleTechnology

Google introduces a user-space kernel sandbox that adds isolation for untrusted code in containers.

Source

September 20171 event

ONNX

MicrosoftTechnology

Microsoft and Facebook introduce a common model representation for framework interoperability.

Source

July 20171 event

PPO

OpenAITechnology

Constrained policy updates offer a practical approach to stable reinforcement learning.

Source

June 20172 events

2 events

Transformer

GoogleLLMTechnology

Google researchers introduce the Transformer, an attention-based sequence model, in Attention Is All You Need.

Source

May 20172 events

Cloud TPU · v2

GoogleHardware

Second-generation TPUs add training support and a route to external access through Cloud TPU.

Source

Tesla V100

NVIDIAHardware

Volta introduces Tensor Cores to accelerate matrix operations for deep learning.

Source

January 20172 events

Sparsely-Gated MoE

GoogleTechnology

Sparse routing selects a small subset of experts to increase capacity while limiting computation.

Source

PyTorch

PyTorchTechnology

Dynamic computation graphs and a Python workflow support deep-learning research.

Source

July 20161 event

Layer Normalization

U. TorontoTechnology

Normalization statistics are computed within each example rather than across a batch.

Source

May 20161 event

Google TPU

GoogleHardware

Google describes its custom tensor processor for neural-network inference.

Source

April 20161 event

Tesla P100

NVIDIAHardware

Pascal combines HBM2 and NVLink for deep learning and multi-GPU computing.

Source

December 20151 event

ResNet

MicrosoftTechnology

Residual connections make very deep networks easier to optimize and train.

Source

November 20151 event

TensorFlow

GoogleTechnology

Google releases its machine-learning framework as open-source software.

Source

August 20151 event

March 20151 event

December 20141 event

Adam

U. AmsterdamTechnology

First- and second-moment gradient estimates adapt optimization step sizes.

Source

November 20141 event

Tesla K80

NVIDIAHardware

A dual-GPU accelerator expands memory and throughput for machine learning and scientific computing.

Source

September 20143 events

cuDNN

NVIDIATechnology

A GPU deep-learning primitive library provides optimized operations for frameworks.

Source

Additive Attention

U. MontréalTechnology

A translation model learns to attend to different input positions as it generates each word.

Source

June 20141 event

GAN

U. MontréalTechnology

Adversarial training between a generator and a discriminator learns a data distribution.

Source

December 20131 event

March 20131 event

Docker

DockerTechnology

Image-based application packaging makes reproducible environments easier to build, share, and run.

Source

January 20131 event

word2vec

GoogleTechnology

Efficient word-vector learning represents relationships in a continuous space.

Source

December 20121 event

AlexNet

U. TorontoTechnology

A deep convolutional network trained on GPUs advances ImageNet image classification.

Source

February 20071 event

CUDA SDK

NVIDIATechnology

A public beta of the CUDA toolkit and SDK brings C-based programming to GPU computing.

Source
BenchmarksArtificial AnalysisArenaSWE-benchMLPerf

Search names, organizations, or keywords across all timelines.