GLM-5.3-Flash 320B
GLM-5.3-Flash 320B needs roughly 218.5 GB VRAM at Q4_K_M quantization (717.0 GB at FP16). 8 GPUs we track can run it fully in VRAM at 8k context.
8 GPUs run this natively · 2 with CPU offload
- Q2_K total
- 136.8 GB
- at 8k context
- Smallest GPU
- 180 GB
- NVIDIA B200 180GB, at Q2_K
- KV cache, full context
- 23.6 GB
- 11 of 45 layers cache
- Inputs
- Text · Image · Video
- MIT, released August 2026
- Active experts
- 8 routed + 1 shared
- of 288 total
- Active parameters
- 18B
- 5.6% of the 320B total
- Context ceiling
- 1,048,576 tokens
- native, no scaling needed
GLM-5.3-Flash 320B is a Mixture of Experts (MoE) model with 320B total parameters but only 18B active per token developed by Z.ai. Released 25 August 2026 under MIT, the first natively multimodal model in the GLM-5 series (text, image, and video input) and the first GLM release with hybrid attention: 45 layers, eleven repeats of three linear-attention layers followed by one DeepSeek Sparse Attention layer (Multi-head Latent Attention, a 512-dimension compressed cache), plus one closing linear-attention layer. It's a mixture of experts, 288 routed plus 1 shared expert with 8 routed active per token (18B of 320B total active), trained on a 30-trillion-token multimodal corpus and using Manifold-Constrained Hyper-Connections (mHC) for scaling stability. Native context is 1,048,576 tokens.
To run GLM-5.3-Flash 320B locally: a Q2_K GGUF is roughly 121.9GB of weights, about 136.8GB in total at 8k context. That already lands in offload territory on a single 141GB H200 once its 5% VRAM reserve is subtracted, so the smallest single card this site tracks that actually fits it is a 180GB B200; a 256GB Apple M3 Ultra clears it too. The hybrid attention is what keeps a full 1,048,576-token window reachable at all: the real 11-of-45 caching layers cost 23.6GB of KV cache at that length, versus 96.6GB if every layer cached conventionally, and the full-context total comes to 163.0GB, still fitting both the 180GB B200 and the 256GB M3 Ultra. An all-attention version of the same model would need 244.8GB at that context: it still (barely) fits the 256GB Mac, but no longer fits the 180GB B200 at all. As a MoE model, inference speed depends on active parameters (18B) rather than total size.
Z.ai's own release chart puts GLM-5.3-Flash ahead of GLM-5.2 on every benchmark it reports: Terminal-Bench 2.1 84.3 (vs 81.0), DeepSWE v1.1 63.4 (vs 46.2), AutomationBench 48.8 (vs 26.2), Agents' Last Exam 26.3 (vs 20.4), HLE with tools 55.3 (vs 54.7), and GDPVal-AA v2 1773 (vs 1504, scored by Artificial Analysis). The same chart places it close to, and on several benchmarks past, Claude Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash, at roughly a tenth of Opus-class API pricing. These are Z.ai's own reported figures against its own chosen comparison set; no independent lab had re-run them at time of writing.
Eleven sparse-attention layers carry the whole KV cache
GLM-5.3-Flash is a mixture of experts: 288 routed experts exist in the weights, and a router picks 8 plus one shared expert for each token, so 18 of the 320 billion parameters actually compute per step. Its 45 layers are also the first hybrid-attention stack in the GLM-5 line: a block of three linear-attention layers (gated recurrent, labelled "KDA" in the model's own config) followed by one DeepSeek Sparse Attention layer, repeated eleven times, with one more linear-attention layer closing out the stack.
11 of 45 layers keep a KV cache that grows with the full context. The other 34 hold a fixed-size recurrent state, so their memory cost does not move as the context window grows.
The 34 linear-attention layers carry a fixed-size recurrent state that costs the same at 2,000 tokens or 1,048,576. Only the 11 DeepSeek Sparse Attention layers store a per-token cache, and even that cache is compressed: Multi-head Latent Attention keeps one 512-dimension latent per token instead of separate per-head keys and values, the same compression approach DeepSeek's V4 line uses. A lightweight indexer then selects which of those cached entries each query actually attends to, which cuts compute, not what has to stay resident in memory.
What the full 1,048,576-token window actually costs
GLM-5.2 reached 1M context with its own sparse-attention design (IndexShare) but kept every layer caching. GLM-5.3-Flash goes further: only a quarter of its layers cache at all.
KV cache only, at FP16. Weights and activation overhead sit on top of these figures.
Filling the full 1,048,576-token window costs 23.6 GB of KV cache with the real hybrid stack, versus 96.6 GB if every layer cached, a roughly 4x reduction. Add Q2_K's 121.9 GB of weights and the full-context total is 163.0 GB, which fits a 180 GB B200 or a 256 GB Apple M3 Ultra. The all-attention version of the same model would need 244.8 GB at that context: it still (barely) fits the 256 GB Mac, but no longer fits the 180 GB B200 at all.
VRAM at each quantization
Figures below assume 8k context. This model's hybrid attention stack caches only some layers, so KV cache grows much slower than linearly as context increases.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 1280.0 GB | 0.18 GB | 1433.8 GB |
| BF16 | 640.0 GB | 0.18 GB | 717.0 GB |
| FP16 | 640.0 GB | 0.18 GB | 717.0 GB |
| Q8_0 | 340.2 GB | 0.18 GB | 381.2 GB |
| Q6_K | 262.7 GB | 0.18 GB | 294.4 GB |
| Q5_K_M | 227.8 GB | 0.18 GB | 255.4 GB |
| Q4_K_M | 194.9 GB | 0.18 GB | 218.5 GB |
| Q3_K_M | 153.9 GB | 0.18 GB | 172.6 GB |
| Q2_Krec | 121.9 GB | 0.18 GB | 136.8 GB |
| NVFP4cuda | 160.0 GB | 0.18 GB | 179.4 GB |
KV cache is calculated at 8k context (FP16). Note that NVFP4 only runs on CUDA GPUs. Turn on TurboQuant in the calculator above for lower KV cache estimates.
Benchmarks
Where GLM-5.3-Flash actually pulls ahead of GLM-5.2
Terminal-Bench 2.1 is the only benchmark both releases report in this site's own Benchmarks fields (84.3 vs 81.0). Z.ai's own release chart covers several more agentic suites with no field in this site's schema; the two below are the largest point gains on that chart.
Z.ai, GLM-5.3-Flash model card benchmark chart (huggingface.co/zai-org/GLM-5.3-Flash, 25 August 2026). Both scores are Z.ai's own vendor-reported figures; independent reproduction was not yet available at time of writing.
AutomationBench nearly doubles, 26.2 to 48.8, and DeepSWE gains 17.2 points, 46.2 to 63.4, the two largest jumps on Z.ai's own chart. Terminal-Bench 2.1, the one benchmark both releases share in this site's own data, moves by a smaller 3.3 points (81.0 to 84.3), so the newly trained base model and hybrid attention design look like they bought the most ground on long-horizon agentic tasks specifically, not on general coding ability.
GPUs that run GLM-5.3-Flash 320B natively (8)
- NVIDIA B300 288GBNVFP4 · 172.3 t/s
- NVIDIA B200 180GBQ2_K · 225.6 t/s
- AMD Instinct MI300XQ3_K_M · 118.6 t/s
- Apple M5 Ultra (512GB)Q8_0 · 15 t/s
- Apple M5 Ultra (256GB)Q4_K_M · 26.1 t/s
Show 3 more
- Apple M3 Ultra (512GB)Q8_0 · 10.2 t/s
- Apple M3 Ultra (256GB)Q4_K_M · 17.8 t/s
- Apple M2 Ultra (192GB)Q3_K_M · 22 t/s
Plus 2 GPUs that run it with CPU offload (slower)
- NVIDIA H200 141GBQ2_K · 135.4 t/s
- Intel Data Center GPU Max 1550Q2_K · 68.6 t/s
Notes
The first natively multimodal model in the GLM-5 series and the first GLM release with hybrid attention: a newly trained base model (not a post-train of GLM-5.2, unlike the text-only GLM-5.3 flagship released the same week) combining linear-attention and DeepSeek-style sparse-attention layers, plus Manifold-Constrained Hyper-Connections (mHC) for scaling stability, trained on a 30-trillion-token multimodal corpus. It's a mixture of experts: 288 routed experts plus 1 shared, 8 routed active per token (18B of 320B active), with the first 3 of 45 layers using a dense (non-routed) feed-forward network. On Z.ai's own release chart it beats GLM-5.2 on every benchmark shown: Terminal-Bench 2.1 84.3 vs 81.0, DeepSWE v1.1 63.4 vs 46.2, AutomationBench 48.8 vs 26.2, Agents' Last Exam 26.3 vs 20.4, HLE with tools 55.3 vs 54.7, and GDPVal-AA v2 1773 vs 1504 (the last scored by Artificial Analysis). Z.ai's chart also places it near or ahead of Claude Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash on most of those same benchmarks, at roughly a tenth of Opus-class API pricing ($0.113 per million input tokens, $0.394 per million output). These are all Z.ai's own reported figures against its own chosen comparison set; no independent lab had re-run them at time of writing.
So, should you actually run this?
Not on a single consumer GPU: even the most aggressive standard quantization this site tracks, Q2_K, already totals 136.8 GB at a short 8k-token context, before the full 1M window even enters the picture. What the hybrid attention buys is context budget, not a smaller weights floor: a 180 GB B200 or a 256 GB unified-memory Mac now holds the model at its full native 1,048,576-token window, something an all-attention design at the same weight class would make much harder. If you need a self-hosted, Opus-class coding and agentic model with a million-token window and already have that class of hardware, this is a genuinely new option; everyone else gets the same weights through Z.ai's API at roughly a tenth of Opus's price.
Frequently asked questions
- What are the VRAM requirements for GLM-5.3-Flash 320B?
- GLM-5.3-Flash 320B requires approximately 218.5 GB of VRAM at Q4_K_M quantization, 381.2 GB at Q8, and 717.0 GB at FP16. These numbers assume 8k context window; its hybrid attention stack caches far fewer than all layers, so VRAM grows much slower than linearly with context.
- How many parameters does GLM-5.3-Flash 320B have?
- GLM-5.3-Flash 320B has 320 billion total parameters, but only 18 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
- Can GLM-5.3-Flash 320B run on a 16 GB GPU?
- No. At Q4_K_M, GLM-5.3-Flash 320B needs 218.5 GB of VRAM, more than 16 GB. You will need a multi-GPU server.
- Can GLM-5.3-Flash 320B run on a 24 GB GPU?
- No. Even at Q4_K_M, GLM-5.3-Flash 320B needs 218.5 GB. Consider a multi-GPU server with 219 GB+ of combined VRAM.
- What is the smallest quantization for GLM-5.3-Flash 320B that fits in 24 GB of VRAM?
- GLM-5.3-Flash 320B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 136.8 GB at Q2_K.
- What GPU do I need to run GLM-5.3-Flash 320B locally?
- You need a multi-GPU server. At Q4_K_M, GLM-5.3-Flash 320B needs 218.5 GB VRAM, more than any single consumer GPU. That's roughly 3x 80 GB datacenter GPUs (H100, A100, or similar) pooled together.
- Can GLM-5.3-Flash run on a single consumer GPU?
- No. Even the most aggressive standard quantization this site tracks, Q2_K, needs about 121.9 GB of weights alone, before any KV cache or activation overhead. The smallest single card this site tracks that actually fits it at a short 8k-token context is a 180 GB NVIDIA B200; a 256 GB Apple M3 Ultra clears it too. Nothing in the 16-48 GB range consumer cards occupy gets close.
- Is GLM-5.3-Flash the same model as GLM-5.3?
- No, despite the shared version number. GLM-5.3, released earlier the same month, is a text-only post-training refresh of the existing GLM-5.2 base model with no architecture change. GLM-5.3-Flash is a separate, newly trained model: a smaller 320B/18B-active checkpoint with a new hybrid linear/sparse-attention design and native image and video input, neither of which GLM-5.3 has.
- Why does only a quarter of the model's layers keep a KV cache?
- GLM-5.3-Flash's 45 layers alternate three linear-attention layers for every one DeepSeek Sparse Attention layer, eleven times over, plus one closing linear-attention layer. Linear attention carries a fixed-size recurrent state that doesn't grow with context, so only the 11 sparse-attention layers store a per-token cache. That's what keeps the full 1,048,576-token context to 23.6 GB of KV cache instead of the 96.6 GB an all-attention version of the same model would need.