unsloth
GLM 4.7 Flash REAP 23B A3B
GLM 4.7 Flash REAP 23B A3B is a 23.0B mixture-of-experts deepseek2 model from unsloth, with about 3.0B parameters active per token. Configured context length is 202,752 tokens. License: mit. Quant sizes are the GGUF file bytes published on Hugging Face when a file exists.
Parameters23.0B / 3.0B active
Context198K
Attentionmla
Licensemit
47 layers · 20 KV heads · head dim 64 · GGUF repo · config
Quant file sizes
| Quant | File size | Source |
|---|---|---|
| BF16 | 42.9 GB | Hugging Face file bytes |
| UD-Q8_K_XL | 25.6 GB | Hugging Face file bytes |
| Q8_0 | 22.8 GB | Hugging Face file bytes |
| UD-Q6_K_XL | 18.8 GB | Hugging Face file bytes |
| Q6_K | 17.7 GB | Hugging Face file bytes |
| UD-Q5_K_XL | 15.6 GB | Hugging Face file bytes |
| Q5_K_M | 15.3 GB | Hugging Face file bytes |
| Q5_K_S | 14.9 GB | Hugging Face file bytes |
| UD-Q4_K_XL | 13.3 GB | Hugging Face file bytes |
| Q4_K_M | 13.1 GB | Hugging Face file bytes |
| Q4_K_S | 12.4 GB | Hugging Face file bytes |
| IQ4_NL | 12.3 GB | Hugging Face file bytes |
| IQ4_XS | 11.7 GB | Hugging Face file bytes |
| Q4_1 | 13.6 GB | Hugging Face file bytes |
| Q4_0 | 12.4 GB | Hugging Face file bytes |
| UD-Q3_K_XL | 10.7 GB | Hugging Face file bytes |
| Q3_K_M | 10.5 GB | Hugging Face file bytes |
| Q3_K_S | 9.59 GB | Hugging Face file bytes |
| UD-IQ3_XXS | 9.35 GB | Hugging Face file bytes |
| UD-Q2_K_XL | 8.39 GB | Hugging Face file bytes |
| Q2_K | 16.4 GB | Hugging Face file bytes |
| UD-IQ2_XXS | 7.67 GB | Hugging Face file bytes |
| UD-IQ1_M | 7.07 GB | Hugging Face file bytes |
| UD-IQ1_S | 6.71 GB | Hugging Face file bytes |
On common machines
| Device | Verdict | Quant | Speed |
|---|---|---|---|
| GeForce RTX 4060 Ti 8GB | Expert offload | Q4_K_M | 69–95 |
| GeForce RTX 3060 12GB | Tight fit | Q3_K_S | 111–151 |
| GeForce RTX 4060 Ti 16GB | Tight fit | UD-Q4_K_XL | 69–94 |
| GeForce RTX 4090 | Tight fit | UD-Q6_K_XL | 180–245 |
| GeForce RTX 5090 | Tight fit | UD-Q8_K_XL | 244–333 |
| Apple M4 Pro | Tight fit | UD-Q5_K_XL | 57–78 |
| Apple M5 | Tight fit | Q3_K_S | 47–64 |
| CPU only · DDR5-5600 | Runs on CPU | Q8_0 | 14–18 |
Smallest devices that fit it fully
- GeForce RTX 3080 10GB — Tight fit, UD-IQ2_XXS, 275–375 tok/s
- Intel Arc B570 — Tight fit, UD-IQ2_XXS, 138–188 tok/s
- GeForce RTX 2080 Ti — Tight fit, UD-Q2_K_XL, 209–285 tok/s
- GeForce GTX 1080 Ti — Tight fit, UD-Q2_K_XL, 164–224 tok/s
- GeForce RTX 5070 — Tight fit, Q3_K_S, 207–282 tok/s
- GeForce RTX 4070 Ti — Tight fit, Q3_K_S, 155–212 tok/s
Run it
Commands for the GeForce RTX 4060 Ti 16GB estimate (UD-Q4_K_XL).
llama.cpp
llama-server -hf unsloth/GLM-4.7-Flash-REAP-23B-A3B-GGUF:UD-Q4_K_XL -c 8192 -ngl 47
No matching name in the Ollama library snapshot.
LM Studio
lms get unsloth/GLM-4.7-Flash-REAP-23B-A3B-GGUF # Then load quant UD-Q4_K_XL