ERNIE-4.5-21B-A3B-Thinking
- In Input modalities: Text
- Reasoning
- Tool use
- Open weights
Prices
0 offerings · prices as listedNo priced offering is listed for this model yet.
Details
- Max input
- 96K
- Licence
- Apache License 2.0
- Family
- ernie-thinking
Not stated on Baidu’s pages: output modalities, reasoning efforts, reasoning can be off, max thinking budget, structured output, built-in tools, knowledge cutoff, aliases.
Run locally
Weights: baidu/ERNIE-4.5-21B-A3B-Thinking- Parameters 21.8B
- Active 3B · 64 experts
- Native precision BF16
- Layers 28
- KV cache per token 56 KB
- Config positions 128K (config.json max_position_embeddings)
- API context 128K (maker’s API)
- Needs
- ≈ 18.0 GB (weights 16.7 + KV cache 0.4 + overhead 0.8)
- Usable
- 21.6 GB (90% of its memory)
- Generation speed
- 172–258 tok/s (reads the active 3B per token)
- Longest context
- 74K (longer doesn’t fit)
- First token
- about 0.10–0.15 s for a 2K prompt
| Precision | Weights | Needs at 8K | On your hardware | Longest context |
|---|---|---|---|---|
| BF16 (native)maker’s file | 40.7 GB | ≈ 43.1 | ○Needs system memory 13–19 tok/s | — |
| Q8_0 | 21.6 GB | ≈ 23.1 | ○Needs system memory 91–137 tok/s | — |
| FP8 | 20.3 GB | ≈ 21.8 | ○Needs system memory 123–184 tok/s | — |
| Q6_K | 16.7 GB | ≈ 18.0 | ●Runs well 172–258 tok/s | 74K |
| Q5_K_M | 14.5 GB | ≈ 15.6 | ●Runs well 194–290 tok/s | 117K |
| Q5_0 | 14.0 GB | ≈ 15.1 | ●Runs well 199–299 tok/s | 127K |
| Q4_K_M | 12.3 GB | ≈ 13.4 | ●Runs well 220–330 tok/s | 128K |
| Q4_0 | 11.4 GB | ≈ 12.4 | ●Runs well 234–350 tok/s | 128K |
| IQ4_XS | 10.8 GB | ≈ 11.8 | ●Runs well 244–366 tok/s | 128K |
| Q3_K_Mquality drops | 9.9 GB | ≈ 10.9 | ●Runs well 260–390 tok/s | 128K |
| IQ3_XXSquality drops | 7.8 GB | ≈ 8.7 | ●Runs well 312–467 tok/s | 128K |
| Q2_Kquality drops | 7.6 GB | ≈ 8.6 | ●Runs well 316–474 tok/s | 128K |
| IQ2_XXSquality drops | 5.2 GB | ≈ 6.2 | ●Runs well 406–609 tok/s | 128K |
Click a row to choose a precision. “Maker’s file” sizes are the maker’s own files; others are parameters × bits per weight.
Compared with other hardware
Each at the highest precision it runs. Click one for details.
Smallest that runs it (Q4_K_M, 8K): GPU GeForce RTX 5080; unified memory Apple M5 Pro (16-core GPU) · 24 GB.
Run commands
The maker publishes no files at this precision. Choose its native precision or one of its own builds, or quantise it yourself with llama.cpp.
Memory to fine-tune
Batch 1, 2,048-token sequences, gradient checkpointing; LoRA and QLoRA train rank-16 adapters on the attention projections.
Estimates; facts from the maker’s Hugging Face repositories, as of 2026-11-06. How it’s estimated
Cited scores
From third-party leaderboards, not measured by usThis model isn’t on any leaderboard we cite yet.
Changes
All changes →No changes recorded yet.
Sources
- ERNIE-4.5-21B-A3B-Thinking_文心大模型_大语言模型_推理大模型-飞桨AI Studio星河社区in Chinese (opens in a new tab) aistudio.baidu.com · checked Nov 5, 2026
- 【ERNIE】10分钟部署ERNIE-4.5-21B-A3B-Thinking - 飞桨AI Studio星河社区in Chinese (opens in a new tab) aistudio.baidu.com · checked Oct 31, 2026
- ERNIE-4.5-21B-A3B-Thinking (opens in a new tab) huggingface.co · checked Nov 3, 2026
- baidu/ERNIE-4.5-21B-A3B-Thinking config.json (opens in a new tab) huggingface.co · checked Nov 5, 2026
- baidu/ERNIE-4.5-21B-A3B-Thinking model card (opens in a new tab) huggingface.co · checked Nov 3, 2026
- 模型更新记录 - 百度千帆·大模型服务及Agent开发平台in Chinese (opens in a new tab) cloud.baidu.com · checked Nov 3, 2026
- baidu/ERNIE-4.5-21B-A3B-Thinking README.md (opens in a new tab) huggingface.co · checked Nov 4, 2026
- baidu/ERNIE-4.5-21B-A3B-Thinking files (opens in a new tab) huggingface.co · checked Nov 1, 2026
- LLM_API (opens in a new tab) ai.baidu.com · checked Nov 3, 2026
Some values on this page are still waiting to be checked against official pages: a filled dot is verified, a hollow one is not yet (with or without a source on file). How we verify data
Baidu · China · cloud.baidu.com
Related models
Snapshots, the same family and other models from this provider