Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › ERNIE-4.5 family

ERNIE-4.5-21B-A3B-Thinking

Hugging Face ↗ Follow changes

baidu Apache 2.0 out 14 months ago

21.8B parameters, 3B active, text in, text out, 128K tokens of context.

Verified 2026-10-25 Rank #607 in LLMs · score 73.9 method v1 Sources 3 Downloads 30d 73.3k · #351 of 619 tracked

Can I run it?13GB at INT4Estimatefits a 16 GB cardmethod
Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?73.3kdownloads, 30 days#351 of 619 we track · below the median+1 likes, downloads −16% this week794 likesHugging Face2026-10-24

Worth reading

Picked by rule from 6 tracked pieces: the lab's own post, the most shared, a field report, the latest.

6 best of 6 tracked

The 1 pick above, plus these 5. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleBaidu: The Open Source Release of the Ernie 4.5 Model Familyernie.baidu.com · 2025-06-30 · 9 HN pts · official
  2. ArticleThe Open Source Release of the Ernie 4.5 Model Familyyiyan.baidu.com · 2025-06-30 · 4 HN pts · official
  3. ArticleBaidu Releases Ernie 4.5 and X1 Models Outperforming GPT-4.5 for 1% the Costanalyticsvidhya.com · 2025-03-17 · 5 HN pts
  4. ArticleChina Releases GPT-4.5 Rival Ernie 4.5 and X1 at 1% the Costvincentschmalbach.com · 2025-03-17 · 1 HN pts
  5. ArticleBaidu to join open-source movement with Ernie 4.5 models publicly availablescmp.com · 2025-07-03 · 1 HN pts

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 24 quantized builds and 12 Spaces.

  • As a service0no API provider listed
  • On own hardware24quantized builds
  • Built on96 fine-tunes · 2 adapters · 1 merges
  • In apps12Hugging Face Spaces

Top Spaces running it: akhaliq/ERNIE-4.5-21B-A3B-Thinking ♥ 21 · jairwaal/image ♥ 16 · Jakob08/moneychatbot ♥ 3 · synthetic-data-universe/synth ♥ 2

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 13 GB at INT4, 26 GB at FP8, 52 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 13 GB of weights (estimate; KV cache and activations extra)INT413 GBFP8: about 26 GB of weights (estimate; KV cache and activations extra)FP826 GBBF16: about 52 GB of weights (estimate; KV cache and activations extra)BF1652 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 13 GB of weights (estimate; KV cache and activations extra)INT413 GBFP8: about 26 GB of weights (estimate; KV cache and activations extra)FP826 GBBF16: about 52 GB of weights (estimate; KV cache and activations extra)BF1652 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2025-09-08), amber triangles are 6 dated articles (each links to it), grey ticks are repo commits.

AprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2025-11-26: 1 repo commit2025-11-10: 1 repo commit2025-10-20: 1 repo commit2025-09-17: 1 repo commit2025-09-11: 3 repo commits2025-09-10: 1 repo commit2025-09-09: 2 repo commits2025-09-08: 4 repo commitsBaidu: The Open Source Release of the Ernie 4.5 Model Family · 2025-06-30The Open Source Release of the Ernie 4.5 Model Family · 2025-06-30Probing Chinese LLM Safety Layers: Reverse-Engineering Kimi and Ernie 4.5 · 2025-11-24Baidu Releases Ernie 4.5 and X1 Models Outperforming GPT-4.5 for 1% the Cost · 2025-03-17China Releases GPT-4.5 Rival Ernie 4.5 and X1 at 1% the Cost · 2025-03-17Baidu to join open-source movement with Ernie 4.5 models publicly available · 2025-07-03Weights 2025-09-08 · 2025-09-08Weights 2025-09-08now
AprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2025-11-26: 1 repo commit2025-11-10: 1 repo commit2025-10-20: 1 repo commit2025-09-17: 1 repo commit2025-09-11: 3 repo commits2025-09-10: 1 repo commit2025-09-09: 2 repo commits2025-09-08: 4 repo commitsBaidu: The Open Source Release of the Ernie 4.5 Model Family · 2025-06-30The Open Source Release of the Ernie 4.5 Model Family · 2025-06-30Probing Chinese LLM Safety Layers: Reverse-Engineering Kimi and Ernie 4.5 · 2025-11-24Baidu Releases Ernie 4.5 and X1 Models Outperforming GPT-4.5 for 1% the Cost · 2025-03-17China Releases GPT-4.5 Rival Ernie 4.5 and X1 at 1% the Cost · 2025-03-17Baidu to join open-source movement with Ernie 4.5 models publicly available · 2025-07-03Weights 2025-09-08 · 2025-09-08Weights 2025-09-08now

What people built on it

33 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

33 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

24 quantized (Hugging Face model tree)6 fine-tunes (Hugging Face model tree)1 merge (Hugging Face model tree)2 adapters (Hugging Face model tree)24 quantized6 fine-tunes1 merge2 adapters
24 quantized (Hugging Face model tree)6 fine-tunes (Hugging Face model tree)1 merge (Hugging Face model tree)2 adapters (Hugging Face model tree)24 quantized6 fine-tunes1 merge2 adapters
GGUF unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF for llama.cpp, Ollama, LM Studio 132.4k downloads MLX nightmedia/ERNIE-4.5-21B-A3B-Thinking-mxfp4-mlx for Apple silicon 699 downloads AWQ cyankiwi/ERNIE-4.5-21B-A3B-Thinking-AWQ-4bit for vLLM, SGLang 37.2k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

What the card says

1 cited sentence

Summary generated, cited

ERNIE-4.5-21B-A3B-Thinking is a text MoE post-trained model with 21.8B total parameters and 3B active parameters.[1][2][3]

  1. parameters human
  2. active parameters human
  3. “21B total parameters and 3B activated parameters”

Generated 2026-10-07 15:14 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

4 use cases

What it is good at 4 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

23 fields

Listed in LLMs · MoE

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesErnie4_5_MoeForCausalLM
Model typeernie4_5_moe
Dtypebfloat16
Params by dtypeBF16: 21,821,012,480 · F32: 4,425,408
Size bandL
MoEyes
Experts64
Context length131,072 (config.json max_position_embeddings)
Languagesen, zh
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Ernie4_5_MoeForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for ERNIE-4.5-21B-A3B-Thinking: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)52 GB · Single high-end GPU (<80 GB)
FP8 (1 byte/param)26 GB · Single high-end GPU (<80 GB)
INT4 (0.5 bytes/param)13 GB · Single GPU (<24 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

33 variants

Lineage

Original model Family: baidu/ERNIE-4.5 33 derivatives on the Hub

quantized: 24finetune: 6merge: 1adapter: 2

Top variants (33 shown, ranked by downloads)

Relation
Quant
Size
33 rows
Variant Relation Quant Params Band License Author Likes Downloads
unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF quantized GGUF — — apache-2.0 unsloth 60 132.4k
bartowski/baidu_ERNIE-4.5-21B-A3B-Thinking-GGUF quantized GGUF — — — bartowski 3 39k
cyankiwi/ERNIE-4.5-21B-A3B-Thinking-AWQ-4bit quantized AWQ 4.53B M apache-2.0 cyankiwi 4 37.2k
gabriellarson/ERNIE-4.5-21B-A3B-Thinking-GGUF quantized GGUF — — apache-2.0 gabriellarson 17 20.8k
mradermacher/ERNIE-4.5-21B-A3B-Thinking-i1-GGUF quantized GGUF — — apache-2.0 mradermacher 1 10.6k
yujiepan/ernie-4.5-moe-tiny-random finetune — 904K S — yujiepan 1 8.3k
enacimie/ERNIE-4.5-21B-A3B-Thinking-Q4_K_M-GGUF quantized GGUF — — apache-2.0 enacimie 0 5.7k
XythicK/ERNIE-4.5-21B-A3B-Thinking-GGUF quantized GGUF — — apache-2.0 XythicK 0 5.1k
mradermacher/ERNIE-4.5-21B-A3B-Thinking-GGUF quantized GGUF — — apache-2.0 mradermacher 5 4.4k
DevQuasar/baidu.ERNIE-4.5-21B-A3B-Thinking-GGUF quantized GGUF — — — DevQuasar 0 3.3k
cyankiwi/ERNIE-4.5-21B-A3B-Thinking-AWQ-8bit quantized AWQ 7.13B M apache-2.0 cyankiwi 4 2k
mmnga/ERNIE-4.5-21B-A3B-Thinking-gguf quantized gguf — — apache-2.0 mmnga 0 2.2k
wekW/ERNIE-4.5-21B-A3B-Thinking-Q8_0-GGUF quantized Q8_0 — — apache-2.0 wekW 1 1.8k
DavidAU/ERNIE-4.5-37B-A3B-Thinking-Brainstorm20x finetune — 37.2B XL apache-2.0 DavidAU 6 503
nightmedia/ERNIE-4.5-21B-A3B-Thinking-mxfp4-mlx quantized mxfp4 21.8B L apache-2.0 nightmedia 1 699
lovedheart/ERNIE-4.5-21B-A3B-Thinking-GGUF-IQ1_S quantized GGUF — — apache-2.0 lovedheart 1 673
DavidAU/ERNIE-21B-A3B-GLM-4.7-Flash-Thinking finetune — 21.8B L apache-2.0 DavidAU 4 437
DavidAU/ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2 finetune — 21.8B L apache-2.0 DavidAU 5 307
noctrex/ERNIE-4.5-21B-A3B-Thinking-MXFP4_MOE-GGUF quantized GGUF — — — noctrex 1 484
DavidAU/ERNIE-21B-A3B-Claude-4.5-High-OPUS-Thinking finetune — 21.8B L apache-2.0 DavidAU 3 357
AmpereComputing/ernie-4.5-a3b-21b-thinking-gguf quantized gguf — — — AmpereComputing 0 506
qurk41/ERNIE-4.5-21B-A3B-Thinking-mlx-4Bit quantized mlx 21.8B L apache-2.0 qurk41 0 391
seawolf2357/Darwin-gpt-ernie-20b merge — 20.9B L apache-2.0 seawolf2357 0 326
NikolayKozloff/ERNIE-4.5-21B-A3B-Thinking-Q3_K_M-GGUF quantized GGUF — — apache-2.0 NikolayKozloff 1 246
NikolayKozloff/ERNIE-4.5-21B-A3B-Thinking-Q3_K_S-GGUF quantized GGUF — — apache-2.0 NikolayKozloff 1 242
NikolayKozloff/ERNIE-4.5-21B-A3B-Thinking-Q2_K-GGUF quantized Q2_K — — apache-2.0 NikolayKozloff 1 221
MrMeOrYou/ERNIE-4.5-21B-A3B-Thinking-Q4_K_M-GGUF quantized GGUF — — apache-2.0 MrMeOrYou 0 243
guodayang/ERNIE-4.5-21B-A3B-Thinking-Q8_0-GGUF quantized Q8_0 — — apache-2.0 guodayang 0 188
guodayang/ERNIE-4.5-21B-A3B-Thinking-Q4_K_M-GGUF quantized GGUF — — apache-2.0 guodayang 0 183
kkioikk/ERNIE-4.5-21B-A3B-Thinking-Q3_K_S-GGUF quantized GGUF — — apache-2.0 kkioikk 0 172
tiny-random/ernie-4.5-moe finetune — 904K S — tiny-random 0 137
Nikskic/elementor-ernie45-custom adapter — — — apache-2.0 Nikskic 0 15
Siddhu217256/food_nutrition_coach adapter — — — afl-3.0 Siddhu217256 0 0

Repository changelog

14 commits

Repository changelog 14 commits

DateCommitSHA
2025-11-26 Update README.md 4341bb4
2025-11-10 Update README.md 00e0afa
2025-10-20 Update README.md (#9) e647ff1
2025-09-17 Update README.md 960d37e
2025-09-11 Update README.md 6baa154
2025-09-11 Update tokenizer_config.json f8154a4
2025-09-11 Update config.json 7f14894
2025-09-10 Upload folder using huggingface_hub 004009b
2025-09-09 update benchmark graph 78d7a20
2025-09-09 update benchmark graph 2141482
2025-09-08 Update README.md 4956ee9
2025-09-08 Upload benchmark.png with huggingface_hub 6450fa8
2025-09-08 Upload folder using huggingface_hub eb280a0
2025-09-08 initial commit d9008e6

Articles and coverage

6 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official2

Baidu: The Open Source Release of the Ernie 4.5 Model Family
ernie.baidu.com2025-06-30officialabout the ERNIE 4.5 familyHN 9 pts · 2 comments1,126 words
Baidu releases the ERNIE 4.5 model family, a new set of large-scale multimodal models available under Apache 2.0. The family includes variants ranging from a 0.3B dense model to a 424B total parameter MoE model, achieving 47% Model FLOPs Utilization in pre-training. Post-training utilizes Unified Preference Optimization (UPO), while inference supports 4-bit/2-bit lossless quantization.written, reviewed, cited
  • The MoE architecture enhances multimodal understanding without compromising performance on text-related tasks.
  • ERNIE-4.5-VL-424B-A47B thinking mode narrows or surpasses the gap to OpenAI-o1 on challenging benchmarks.
  • The FastDeploy toolkit supports multi-machine PD Disaggregation with multi-level load balancing for deployment.
The Open Source Release of the Ernie 4.5 Model Family
yiyan.baidu.com2025-06-30officialabout the ERNIE 4.5 familyHN 4 pts · 0 comments43 words

Analysis4

Probing Chinese LLM Safety Layers: Reverse-Engineering Kimi and Ernie 4.5
zenodo.org2025-11-24about the ERNIE 4.5 familyHN 3 pts · 1 comments497 words
A controlled study compares censorship behavior in Kimi.com and Ernie 4.5 Turbo Chinese LLMs. Researchers found that emotional framing, specifically soft and empathic tones, can temporarily soften censorship responses in Kimi.com, revealing neutral references to normally censored historical events like Tiananmen 1989. An update on 24 november 2025 added a supplementary note analyzing topic-gated persona behavior in Ernie 4.5 Turbo alongside newly documented behaviors such as sovereignty recognition cascades in the V2 paper.written, reviewed, cited
Baidu Releases Ernie 4.5 and X1 Models Outperforming GPT-4.5 for 1% the Cost
analyticsvidhya.com2025-03-17about the ERNIE 4.5 familyHN 5 pts · 0 comments1,681 words
Baidu released ERNIE 4.5 and ERNIE X1, multimodal and deep-reasoning models that outperform GPT-4.5 in benchmarks while costing 1% of the price. ERNIE 4.5 API access costs $0.55 for input and $2.20 for output per million tokens, whereas ERNIE X1 is priced at $0.28 and $1.10 respectively but remains unavailable via API. Both models are accessible through ERNIE Bot, though registration is currently limited to Chinese nationals.written, reviewed, cited
  • ERNIE 4.5 is the first AI chatbot interface to incorporate audio analysis for source identification and significance description.
  • The ERNIE X1 model will soon be made available on the Baidu AI Cloud Qianfan platform for API access.
  • ERNIE 4.5 allows users to upload multiple files of various types including docs, PDFs, PPTs, and Excel sheets simultaneously.
China Releases GPT-4.5 Rival Ernie 4.5 and X1 at 1% the Cost
vincentschmalbach.com2025-03-17about the ERNIE 4.5 familyHN 1 pts · 1 comments5,483 words
Baidu launched ERNIE 4.5, a multimodal foundation model, and ERNIE X1, a specialized reasoning model with autonomous tool use. ERNIE 4.5 pricing is ¥0.004 per 1,000 input tokens and ¥0.016 per 1,000 output tokens. ERNIE X1 delivers performance on par with DeepSeek R1 at only half the price and costs just under 2% of OpenAI's corresponding model.written, reviewed, cited
  • ERNIE X1 is the first model of its kind that can use tools autonomously to solve problems.
  • Baidu's CEO noted they ran out of GPUs serving the GPT-4.5 preview due to its computational cost.
Baidu to join open-source movement with Ernie 4.5 models publicly available
scmp.com2025-07-03about the ERNIE 4.5 familyHN 1 pts · 0 comments

Similar models

3 models

Similar models

  • gpt-oss-20b · 20.9B
    same size band (L), also MoE, same task (text-generation), more providers, permissive license
  • gpt-oss-safeguard-20b · 21.5B
    same size band (L), also MoE, same task (text-generation), more providers, permissive license
  • LFM2.5-24B-A2B · 24.0B
    same size band (L), also MoE, same task (text-generation)

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:42 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:14 UTC · Last verified 2026-10-25

Follow ERNIE-4.5-21B-A3B-Thinking

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.