LLM Artificial Analysis Index >08.2025<

🏢 OpenAI (USA)

GPT‑4o: OpenAI’s flagship multimodal model with ~1.8 trillion parameters and ~128K token context window. High performance, fast inference, and multimedia support (explodingtopics.com, codingscape.com).
GPT‑4.1: A follow-up to GPT‑4 (released April 2025) with a 1 million token context window, optimized for reasoning, vision, and coding workloads (turn0search1, turn0search3).
GPT‑5 (in development): Expected release imminently (as of August 6, 2025), promising enhanced reasoning and “test‑time compute” capabilities (turn0news20, turn0search16).
GPT‑OSS‑120B & GPT‑OSS‑20B: Open‑weight models released August 5, 2025. The 20B model runs on consumer laptops (~16 GB VRAM); the 120B model requires a single GPU. Both perform comparably to OpenAI’s proprietary o3‑mini and o4‑mini models (turn0news25, turn0news24, turn0news23).

Gemini 2.5 Pro: Released June 2025 with ~1 million token context window, full multimodal support (text, image, video, audio), top-tier benchmark performance (turn0search1, turn0search3, turn0search8).
Gemini 2.5 Flash / Flash‑Lite: Cost- and latency-optimized variants with same massive context window and multimodal capabilities (turn0search3, turn0search8).
Gemma 3 / Gemma 3n (open): Google’s open models (1B to 27B parameters), multilingual and multimodal, designed for on-device deployment (turn0search35, turn0search14).

Claude 3.7 Sonnet and Claude Opus 4 (released May 2025): Proprietary, multimodal, 200K token context window, focused on safety, chat-based conversational and reasoning tasks (turn0search4, turn0search3, turn0search16).

DeepSeek‑V3 and DeepSeek‑R1 (R1‑0528): Open‑source (open‑weight), ~671B parameters with ~37B active per token, 128K token context window. Excellent on coding, math, reasoning benchmarks; V3 released Dec 2024, R1 in May 2025 (turn0search36, turn0search3).

Falcon 180B: Open‑weight 180B parameter model, Apache‑2.0 licensed, strong multilingual and coding benchmarks, widely used in open deployments (turn0search37, turn0search3).

Qwen 3‑235B: Released April–July 2025, open‑weight (Apache‑2.0), both dense and MoE variants (~22B active parameters), strong performance in Chinese/English and reasoning tasks (turn0search38, turn0search3).
Other Qwen models include Qwen 2.5 (dense and multimodal), up to ~72B parameters (turn0search36, [turn0search38]).

Mistral Large 2: Open‑weight ~123B parameter model with 128K token context window. Known for coding prowess and efficiency. Released late 2024, research license available (turn0search37).
Pixtral Large: Multimodal variant (124B + image encoder), research license (turn0search37).

Nemotron‑4 (NVIDIA): ~340B parameters, open‑weight under NVIDIA license; enterprise focus, less widely adopted than Falcon or LLaMA (turn0search14).
Gemma 3 family (Google open): Highly popular small‑to‑medium models, multilingual multimodal performance, deployed widely via Hugging Face (turn0search35).
MiniMax‑Text‑01 (Minimax): ~456B open‑source model released in early 2025 (turn0search36).

Company / Model	Proprietary or Open‑Weight/Source	Approx. Params	Context Window	Highlights
OpenAI – GPT‑4o	Proprietary	~1.8T	~128K	Multimodal general-purpose flagship
OpenAI – GPT‑4.1	Proprietary	Not disclosed	~1M tokens	Reasoning and vision across domains
OpenAI – GPT‑OSS‑120B/20B	Open‑Weight (Apache 2.0)	120B / 20B	~200K	First open‑weight models from OpenAI since 2019
Google – Gemini 2.5 Pro	Proprietary	Not disclosed	~1M tokens	Multimodal, extreme context, reasoning leader
Google – Gemma 3 / 3n	Open‑Source	1–27B	~128K	Lightweight, multilingual, on-device capable
Anthropic – Claude Opus/3.7	Proprietary	Not disclosed	~200K	Safety-first, complex reasoning and chat-oriented
DeepSeek – V3 / R1	Open‑Weight / Open Source	671B (37B active)	~128K	Strong benchmarks, efficient sparse architecture
Alibaba – Qwen 3‑235B	Open‑Weight (Apache 2.0)	235B / ~22B active	~262K	Multilingual, large context, Chinese/English
TII – Falcon 180B	Open‑Weight (Apache 2.0)	180B	~2K	Efficient, widely adopted open model
Mistral – Large 2 / Pixtral	Open‑Weight / Research License	123–124B	~128K	Coding optimization, multimodal via Pixtral
NVIDIA – Nemotron‑4	Open‑Weight / NVIDIA license	~340B	~8K	Enterprise-grade open model
Minimax – MiniMax‑Text‑01	Open‑Source	456B	~4M tokens	Large open-text model with unique context length

Most powerful proprietary: OpenAI GPT‑4o, GPT‑4.1, Gemini 2.5 Pro, Claude Opus 4.
Most significant open source / open‑weight: DeepSeek V3/R1, Qwen 3‑235B, Mistral Large 2, Falcon 180B, Nemotron‑4, and Gemma 3.
Highest context windows: GPT‑4.1 and Gemini 2.5 Pro (~1M tokens).
Accessibility: GPT‑OSS models and open‑weight releases by DeepSeek, Qwen, Mistral, and Mistral’s Pixtral offer full-download and fine-tuning capability.

Model	Company	~Params (Active)	License	Why It Stands Out
DeepSeek‑V3 / R1	DeepSeek‑AI	671B (37B active)	MIT (fully open)	Top benchmark scores; ultra-efficient; widely replicable (GitHub, Financial Times)
LLaMA 4	Meta AI	~405B (dense)	Meta License	Broad community use, flexible deployment (Instaclustr, Wikipedia)
Falcon 180B	TII (UAE)	180B (dense)	Apache‑2.0	Strong multilingual capabilities, open access (Bestarion, Reddit)
Qwen 3‑235B	Alibaba Cloud	235B (dense) / MoE	Apache‑2.0	Competitive in multi-language tasks; commercial‑grade quality (Exploding Topics, Bestarion)
DBRX	Databricks / Mosaic ML	~unknown large (MoE)	Open Model License	Excellent reasoning/math performance benchmarks (Wikipedia, Exploding Topics)
Nemotron‑4	NVIDIA	~340B	NVIDIA license	Open release by major tech company; still niche adoption (Wikipedia, Bestarion)

BIGGEST

Rank	Model	Params (approx.)	License
1	PanGu‑Σ	~1.085 T	Proprietary
2	GPT‑4o	~1.8 T (est.)	Proprietary (API)
3	DeepSeek‑V3 / R1	671 B (37 B active)	Open source
4	Qwen 3	235 B	Open source
5	Llama 4	up to 405 B	Open source
6	GLM‑4.5	~355 B	Proprietary
7	Grok 3	~314 B	Proprietary
8	Mistral Large 2	~123 B	Open / research
9	gpt‑oss‑120B	117 B	Open (Apache 2.0)
…	Others…	Various larger / smaller	Mixed