LLM Size and Parameter Comparison

As of July 2026, OpenAI has never officially disclosed how many parameters GPT-5 has, and no GPT-5.x version ships a public size; the most-cited independent estimate puts GPT-5 at roughly 100 billion active parameters, inferred from price and speed by Epoch AI, with no credible figure for its total size ^[1]. The largest language model whose size is both officially disclosed and downloadable is DeepSeek V4-Pro, a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active per token, released under an MIT license in April 2026 ^[2]. Bigger numbers circulate, Baidu's ERNIE 5.0 is described as roughly 2.4 trillion parameters and a leaked estimate put GPT-4 near 1.8 trillion, but those are proprietary rough claims or unconfirmed leaks, and Meta's announced 2-trillion-parameter Llama 4 Behemoth was never released ^[3][4][5].

Verdict: if you want the biggest model you can actually download and audit, it is DeepSeek V4-Pro at 1.6T total parameters (49B active). If you are asking about the frontier proprietary chatbots (ChatGPT, Claude, Gemini), the honest answer is that none of their vendors publish parameter counts, so every number you see for them is an estimate. This page separates the two cleanly: open-weight rows are exact figures taken from official model cards and HuggingFace configs, and proprietary rows are labeled as estimates or undisclosed with the source of each estimate. Last verified: July 2026.

How many parameters does GPT-5 have?

OpenAI has not disclosed a parameter count for GPT-5, GPT-5.1, GPT-5.2, or the current flagship GPT-5.5 (released April 2026) ^[1]. This is deliberate: OpenAI stopped publishing model sizes after GPT-3 (175B), and even the widely repeated GPT-4 figure of about 1.8 trillion parameters comes from a 2023 SemiAnalysis leak that OpenAI never confirmed ^[4]. The best-supported public estimate for GPT-5 is from Epoch AI, which reasons from API pricing, throughput, and industry trends to put GPT-5 at roughly 100 billion active parameters, in the same class as Grok-2 (about 115B active), and explicitly notes there is no reliable estimate of its total size ^[1]. Treat any specific GPT-5 total (the "2 trillion" and "5 trillion" numbers on content-farm sites) as fabricated. Anthropic and Google are the same story: no Claude or Gemini model has ever had its parameter count disclosed, and no credible public estimate exists for either ^[1].

What is the largest LLM by parameters?

Among models with an officially disclosed size, DeepSeek V4-Pro is the largest at 1.6T total parameters, with 49B active per token across 384 routed experts plus one shared expert, six routed per token ^[2]. Its config.json on HuggingFace (61 layers, hidden size 7168, 384 experts) reproduces that total, so the figure is verifiable rather than a claim ^[2]. Behind it, Kimi K2.6 from Moonshot AI is a 1 trillion parameter MoE (32B active) ^[6], and GLM-5.2 from Zhipu is 744B (40B active) ^[7].

Several larger-sounding numbers do not belong in the same category:

• ERNIE 5.0 (Baidu): described by Baidu as about 2.4 trillion parameters, but it is proprietary and API-only, the active count is undisclosed (Baidu says under 3 percent activate per query), and there is no technical report, so the total is a rounded marketing figure, not a verified spec ^[3].
• GPT-4 (OpenAI): about 1.8 trillion total, from a leak attributed to SemiAnalysis and George Hotz, never confirmed by OpenAI ^[4].
• Llama 4 Behemoth (Meta): Meta disclosed a "nearly two trillion" parameter spec (288B active, 16 experts) in April 2025 while it was still training, then paused it; as of July 2026 it has never been released as open weights ^[5][8].
• Qwen3-Max (Alibaba): announced as "over 1 trillion" parameters, but that is a rough lower bound, the exact total and active counts were never disclosed, and it is proprietary/API-only ^[9].

So the leader depends on what you count. Largest disclosed and downloadable: DeepSeek V4-Pro (1.6T). Largest number a vendor has ever attached to a shipping product: ERNIE 5.0 (about 2.4T, but proprietary and unverifiable). Largest ever announced as open weights: Llama 4 Behemoth (about 2T), which never shipped.

Open-weight LLMs by disclosed parameter count

Every figure below is officially disclosed on the model card, technical report, or the HuggingFace config.json, and the models are downloadable. Total and active parameters are stated by the developer. Sorted by total parameters, largest first. "MoE (N exp)" means a mixture-of-experts model with N total experts; "active" is parameters used per token.

Scroll sideways for more →

Notes on this table: DeepSeek's headline for the V3 line (V3, V3.1, V3.2) is 671B total / 37B active; HuggingFace shows 685B on-disk because the weights include a 14B multi-token-prediction module ^[11]. Baidu also ships an ERNIE-4.5-300B-A47B text-only model (300B total, 47B active) alongside the 424B vision-language model listed above ^[13]. Qwen3-Coder-480B and Qwen3-235B are from the 2025 Qwen3 generation; Qwen3.5-397B (February 2026) is the largest open-weight Qwen flagship as of July 2026 ^[9]. Licenses vary widely: DeepSeek V4, Kimi K2.6, and GLM use MIT or Modified MIT; Mistral Large 3, Mixtral, and Gemma 4 use Apache-2.0; Llama 4 uses Meta's community license; Grok-2 is source-available (not open source), and Mistral Large 2 is research/non-commercial only.

Proprietary and unreleased models: estimates and undisclosed sizes

None of the numbers in this table are officially disclosed parameter counts. They are vendor rough claims, third-party estimates, leaks, or blanks where nothing credible exists. Do not cite any of these as a confirmed figure.

Scroll sideways for more →

Total vs active parameters: why MoE changes the question

For a dense model like Gemma 4 31B or Mistral Large 2, every parameter is used on every token, so "size" is one number ^[10][18]. Almost every frontier open-weight model in 2026 is instead a mixture-of-experts model, where a router activates only a few of many expert subnetworks per token. DeepSeek V4-Pro has 1.6T total parameters but activates only 49B per token; Kimi K2.6 has 1T total but 32B active ^[2][6]. The total parameter count sets memory and storage requirements (you must load all 1.6T weights), while the active count drives compute cost and speed per token. This is why a 1T MoE like Kimi K2.6 can be cheaper to serve than a 123B dense model despite having eight times the parameters, and it is why comparing raw totals across dense and MoE models can mislead. When someone asks which model is "biggest," specify whether you mean total capacity or the active compute per token.

Which labs disclose parameter counts, and which do not?

The open-weight labs (DeepSeek, Alibaba Qwen, Zhipu, Moonshot, MiniMax, Mistral, and Baidu for its ERNIE 4.5 open models) disclose exact total and active counts, because releasing the weights makes the numbers checkable anyway ^{[2][6][7][9][10][13]}. xAI is partial: it stated Grok-1's 314B/86B exactly and open-sourced it under Apache-2.0, but for Grok-2 it published weights without stating a figure, so the roughly 270B total is counted from the released tensors, not disclosed ^[15][16]. The frontier proprietary labs disclose nothing: OpenAI (GPT-5.x), Anthropic (Claude), and Google (Gemini 3) have never published a parameter count for their current models, and only OpenAI's older GPT-4 has even a leaked estimate ^[1][4]. Baidu and Alibaba occupy a middle ground for their closed flagships, offering rounded headline claims (ERNIE 5.0 "2.4T", Qwen3-Max "over 1T") without exact or active figures ^[3][9]. The practical rule: if a model's weights are downloadable, its size is a fact; if it is API-only, its size is at best an estimate.

References

1. Epoch AI, "Notes on GPT-5 training compute" (2025), estimating GPT-5 at roughly 100B active parameters via inference economics and confirming OpenAI has not disclosed a figure. https://epochai.substack.com/p/notes-on-gpt-5-training-compute ↩
2. DeepSeek V4 model card and API release notes, and the DeepSeek-V4-Pro config.json (61 layers, 384 routed + 1 shared experts, 6 active), 1.6T total / 49B active; V4-Flash 284B / 13B. https://api-docs.deepseek.com/news/news260424 and https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro ↩
3. Baidu, "ERNIE 5.0: A 2.4 Trillion-Parameter Unified Multimodal Foundation Model" (2026). https://ernie.baidu.com/blog/ ↩
4. SemiAnalysis / The Decoder, GPT-4 architecture leak (about 1.8T total, 16 experts, 2 active), never confirmed by OpenAI (2023). https://the-decoder.com/gpt-4-architecture-datasets-costs-and-more-leaked/ ↩
5. Meta AI, "The Llama 4 herd" (April 2025), disclosing Llama 4 Behemoth at nearly 2T total / 288B active while still training. https://ai.meta.com/blog/llama-4-multimodal-intelligence/ ↩
6. Moonshot AI, Kimi K2.6 model card (1T total / 32B active, 384 experts), and the Kimi K2 technical report (arXiv 2507.20534). https://huggingface.co/moonshotai/Kimi-K2.6 ↩
7. Zhipu / Z.ai, GLM-5.2 model card and config.json (744B total / 40B active) and GLM-4.5/4.6 (355B / 32B); GLM-4.5 technical report arXiv 2508.06471. https://huggingface.co/zai-org/GLM-5.2 and https://docs.z.ai/release-notes/new-released ↩
8. Computerworld (WSJ-sourced), "Meta hits pause on Llama 4 Behemoth" (May 2025), and Wikipedia "Llama (language model)" confirming Behemoth was never released. https://www.computerworld.com/article/3987990/meta-hits-pause-on-llama-4-behemoth-ai-model-amid-capability-concerns.html ↩
9. Alibaba Qwen model cards: Qwen3-235B-A22B-Instruct-2507 (235B/22B), Qwen3-Coder-480B-A35B (480B/35B), Qwen3.5-397B-A17B (397B/17B); Qwen3-Max announcement ("over 1 trillion parameters"). https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct and https://huggingface.co/Qwen/Qwen3.5-397B-A17B ↩
10. Mistral AI, "Mistral 3" (December 2025), Mistral Large 3 at 675B total / 41B active (Apache-2.0); "Mistral Large 2" (July 2024) at 123B dense (research license). https://mistral.ai/news/mistral-3/ and https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512 ↩
11. DeepSeek-V3 technical report (arXiv 2412.19437): 671B total / 37B active, 256 routed + 1 shared experts, 8 active per token. https://arxiv.org/abs/2412.19437 ↩
12. MiniMax model cards and technical reports: MiniMax-M1 (456B / 45.9B active, arXiv 2506.13585) and MiniMax-M3 (about 428B / 23B active, native multimodal). https://huggingface.co/MiniMaxAI/MiniMax-M3 and https://huggingface.co/MiniMaxAI/MiniMax-M1-80k ↩
13. Baidu, ERNIE 4.5 blog and model cards: ERNIE-4.5-VL-424B-A47B (424B / 47B) and ERNIE-4.5-300B-A47B text model (300B / 47B), Apache-2.0. https://ernie.baidu.com/blog/posts/ernie4.5/ and https://huggingface.co/baidu/ERNIE-4.5-300B-A47B-PT ↩
14. Meta AI, Llama 4 model cards: Scout (109B total / 17B active, 16 experts) and Maverick (400B / 17B, 128 experts). https://ai.meta.com/blog/llama-4-multimodal-intelligence/ and https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct ↩
15. xAI, Grok-1 open-source announcement and repository (314B total / 86B active, 8 experts, Apache-2.0); current Grok 4.x models are proprietary. https://github.com/xai-org/grok-1 ↩
16. xAI, Grok-2 open-weights release on HuggingFace (August 2025), about 270B total / 115B active counted from the released tensors, Grok 2 Community License (source-available). https://huggingface.co/xai-org/grok-2 ↩
17. Mistral AI, "Cheaper, Better, Faster, Stronger" Mixtral 8x22B (April 2024): 141B total / 39B active, 8 experts, top-2 routing, Apache-2.0. https://mistral.ai/news/mixtral-8x22b/ ↩
18. Google, Gemma 4 documentation and Open Source Blog: largest dense model 31B (30.7B), plus a 26B-A4B MoE (25.2B / 3.8B), Apache-2.0; Gemma 3 27B is dense. https://ai.google.dev/gemma/docs/core and https://opensource.googleblog.com/2026/03/gemma-4-expanding-the-gemmaverse-with-apache-20.html ↩