GLM 4.6V Flash
10.3B parameters · 128k context · MIT · released 2025-12-07
visionagenticcodingreasoningchat
In short
GLM 4.6V Flash is an open model from Z.ai with 10.3B parameters. At Q4_K_M, GLM 4.6V Flash needs about 7 GB of memory at 8k context and 7 GB at 8k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in the US that runs GLM 4.6V Flash fully in memory at 8k context is the NVIDIA Tesla P40, $345 at Amazon, at about 21 tok/s (estimated).
Updated . Speeds are estimates; how they are worked out.
What to run it on in the US
Fully in memory at 8k context, priced at retailers today.
Cheapest
NVIDIA Tesla P4021tok/sest.Q8_0
$345
Best value
NVIDIA GeForce RTX 3080 10GB73tok/sest.Q4_K_M
$469
Fastest
NVIDIA GeForce RTX 5090108tok/sest.Q8_0
$4,921
Mac
Mac mini M4 Pro 48GB17tok/sest.Q8_0
$2,149
All hardware that runs it
Everything we price, best value first. Two cards split the model by layers.
| Hardware | Quant | Fits | Decode | Price | tok/s per 1,000 |
|---|---|---|---|---|---|
| NVIDIA Tesla P40 | Q8_0 | fits | 21 tok/s | $345 | 60.3 |
| Intel Arc B580 | Q8_0 | fits | 22 tok/s | $346 | 63.9 |
| NVIDIA GeForce RTX 3060 12GB | Q8_0 | fits | 22 tok/s | $360 | 60.0 |
| Intel Arc A770 16GB | Q8_0 | fits | 27 tok/s | $400 | 67.8 |
| NVIDIA GeForce RTX 3080 10GB | Q4_K_M | fits | 73 tok/s | $469 | 154.5 |
| AMD Radeon RX 7600 XT | Q8_0 | fits | 15 tok/s | $480 | 31.9 |
| AMD Radeon RX 9060 XT 16GB | Q8_0 | fits | 17 tok/s | $520 | 32.7 |
| AMD Radeon RX 7800 XT | Q8_0 | fits | 33 tok/s | $550 | 60.4 |
| NVIDIA GeForce RTX 5060 Ti 16GB | Q8_0 | fits | 27 tok/s | $618 | 43.5 |
| Intel Arc Pro B60 | Q8_0 | fits | 22 tok/s | $650 | 34.0 |
| NVIDIA GeForce RTX 4070 Super | Q8_0 | fits | 30 tok/s | $699 | 43.2 |
| Mac mini M4 16GB | Q8_0 | fits | 8 tok/s | $749 | 10.0 |
| NVIDIA GeForce RTX 5070 | Q8_0 | fits | 40 tok/s | $750 | 53.7 |
| NVIDIA GeForce RTX 4060 Ti 16GB | Q8_0 | fits | 17 tok/s | $770 | 22.5 |
| AMD Radeon RX 7900 XT | Q8_0 | fits | 43 tok/s | $800 | 53.3 |
| AMD Radeon RX 9070 XT | Q8_0 | fits | 34 tok/s | $815 | 41.8 |
| NVIDIA GeForce RTX 4070 Ti Super | Q8_0 | fits | 40 tok/s | $984 | 41.0 |
| NVIDIA GeForce RTX 5070 Ti | Q8_0 | fits | 54 tok/s | $995 | 54.1 |
| NVIDIA GeForce RTX 4080 Super | Q8_0 | fits | 44 tok/s | $1,199 | 36.9 |
| Mac mini M4 32GB | Q8_0 | fits | 8 tok/s | $1,420 | 5.3 |
| AMD Radeon RX 7900 XTX | Q8_0 | fits | 51 tok/s | $1,533 | 33.3 |
| NVIDIA GeForce RTX 3090 | Q8_0 | fits | 56 tok/s | $1,600 | 35.1 |
| NVIDIA GeForce RTX 5080 | Q8_0 | fits | 58 tok/s | $1,600 | 36.0 |
| AMD Radeon AI PRO R9700 | Q8_0 | fits | 34 tok/s | $1,700 | 20.1 |
| NVIDIA GeForce RTX 3090 Ti | Q8_0 | fits | 61 tok/s | $2,060 | 29.4 |
| NVIDIA GeForce RTX 4090 | Q8_0 | fits | 61 tok/s | $2,073 | 29.2 |
| Mac mini M4 Pro 48GB | Q8_0 | fits | 17 tok/s | $2,149 | 8.0 |
| Mac mini M4 Pro 64GB | Q8_0 | fits | 17 tok/s | $2,900 | 5.9 |
| Mac Studio M4 Max 64GB | Q8_0 | fits | 34 tok/s | $3,400 | 10.1 |
| Framework Desktop Ryzen AI Max+ 395 128GB | Q8_0 | fits | 14 tok/s | $3,449 | 3.9 |
| GMKtec EVO-X2 Ryzen AI Max+ 395 128GB | Q8_0 | fits | 14 tok/s | $3,650 | 3.7 |
| Mac Studio M3 Ultra 96GB | Q8_0 | fits | 52 tok/s | $4,399 | 11.7 |
| NVIDIA GeForce RTX 5090 | Q8_0 | fits | 108 tok/s | $4,921 | 21.8 |
| Mac Studio M4 Max 128GB | Q8_0 | fits | 34 tok/s | $4,999 | 6.9 |
| ASUS Ascent GX10 128GB | Q8_0 | fits | 16 tok/s | $6,100 | 2.7 |
| NVIDIA DGX Spark 128GB | Q8_0 | fits | 16 tok/s | $7,399 | 2.2 |
| NVIDIA RTX PRO 6000 Blackwell | Q8_0 | fits | 108 tok/s | $7,997 | 13.4 |
| Mac Studio M3 Ultra 256GB | Q8_0 | fits | 52 tok/s | $8,499 | 6.1 |
| Mac Studio M3 Ultra 512GB | Q8_0 | fits | 52 tok/s | $23,000 | 2.2 |
| Framework Desktop Ryzen AI Max+ 395 64GB | Q8_0 | fits | 14 tok/s | – | – |
Memory needed
Weights, KV cache and runtime buffers at each context length (KV cache at fp16).
| Quant | File | @8k | @32k | @64k | @128k |
|---|---|---|---|---|---|
| Q4_K_Mgguf | 5.7 GB | 6.8 GB | 7.9 GB | 9.3 GB | 12.2 GB |
| Q8_0gguf | 9.3 GB | 10.4 GB | 11.5 GB | 12.9 GB | 15.8 GB |
Run it
Copy a command for the runtime you use.
- Q4_K_M
- Ollama
ollama run hf.co/ggml-org/GLM-4.6V-Flash-GGUF:Q4_K_M - LM Studio
lms get ggml-org/GLM-4.6V-Flash-GGUF@q4_k_m - llama.cpp
llama-server -hf ggml-org/GLM-4.6V-Flash-GGUF:Q4_K_M
- Ollama
- Q8_0
- Ollama
ollama run hf.co/ggml-org/GLM-4.6V-Flash-GGUF:Q8_0 - LM Studio
lms get ggml-org/GLM-4.6V-Flash-GGUF@q8_0 - llama.cpp
llama-server -hf ggml-org/GLM-4.6V-Flash-GGUF:Q8_0
- Ollama
Notes
This repo hosts GLM-4.6V-Flash (9B), the lightweight variant, not the 106B GLM-4.6V; vLLM serving needs vllm>=0.12.0 and transformers>=5.0.0rc0 (SGLang needs >=0.5.6.post1). The card admits that pure-text QA is weak and that the model can overthink or repeat itself.
Questions
- How much VRAM does GLM 4.6V Flash need?
- At Q4_K_M, GLM 4.6V Flash needs about 7 GB of memory at 8k context and 7 GB at 8k, counting the weights, the KV cache and runtime buffers.
- What is the best GPU for running GLM 4.6V Flash locally?
- The cheapest hardware in the US that runs GLM 4.6V Flash fully in memory at 8k context is the NVIDIA Tesla P40, $345 at Amazon, at about 21 tok/s (estimated). The best value is the NVIDIA GeForce RTX 3080 10GB (about 73 tok/s for $469). The fastest we price is the NVIDIA GeForce RTX 5090, at about 108 tok/s.
- Can I run GLM 4.6V Flash on a 16 GB GPU?
- Yes: at Q8_0 it fits fully in 16 GB at 8k context. On the NVIDIA GeForce RTX 5080 it runs at about 58 tok/s (estimated).
- Can I run GLM 4.6V Flash on a 24 GB GPU?
- Yes: at Q8_0 it fits fully in 24 GB at 8k context. On the NVIDIA GeForce RTX 3090 Ti it runs at about 61 tok/s (estimated).
- How do I run GLM 4.6V Flash locally?
- With Ollama: ollama run hf.co/ggml-org/GLM-4.6V-Flash-GGUF:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
- What licence is GLM 4.6V Flash released under?
- MIT, which allows commercial use.