DeepSeek V4 Pro 0813 1.6T
DeepSeek V4 Pro 0813 1.6T needs roughly 1091.4 GB VRAM at Q4_K_M quantization (3584.1 GB at FP16). 0 GPUs we track can run it fully in VRAM at 8k context.
0 GPUs run this natively · 0 with CPU offload
- Q2_K total
- 682.9 GB
- at 8k context
- KV cache, full context
- 13.1 GB
- 61 layers cache, ~10× fewer entries than tokens
- Inputs
- Text
- MIT, released August 2026
- Active parameters
- 49B
- 3.1% of the 1.6T total
- Reasoning modes
- Non-think / High / Max
- selectable per request
- Context ceiling
- 1,048,576 tokens
- 384K max output
DeepSeek V4 Pro 0813 1.6T is a Mixture of Experts (MoE) model with 1600B total parameters but only 49B active per token developed by DeepSeek. The official release DeepSeek's API quietly switched to on August 12, 2026, closing out the April preview with no separate announcement or new weights upload. The 1.6T-total, 49B-active architecture, 61 layers of hybrid CSA/HCA attention, and 1M-token context are all unchanged; only the post-training run behind the checkpoint moved.
To run DeepSeek V4 Pro 0813 1.6T locally: Not a local model for anyone outside a datacenter. The generic Q2_K ladder alone puts the weights at about 610GB, past every single machine this site tracks, including a 512GB Mac Studio M3 Ultra. No vetted quantized build exists yet; the API is the realistic way to use this model. As a MoE model, inference speed depends on active parameters (49B) rather than total size.
The agentic benchmarks jumped hard for an architecture-unchanged release: DeepSWE went from 12.8 to 62.7 and CyberGym from 52.7 to 83.3 on DeepSeek's own comparison against Claude Fable 5. Terminal-Bench 2.1 lands at 87.9. On that same nine-benchmark table Fable 5 still leads by an average of 5.3%, though DeepSeek's blended API price is roughly 1/46th of Fable 5's.
DeepSeek V4 Pro 0813 VRAM at 1M context: 13.1 GB of KV cache, not 131 GB
The 0813 release keeps the exact attention design the April preview shipped with: DeepSeek V4's hybrid Compressed Sparse Attention and Heavily Compressed Attention, the same trick that makes DeepSeek V4 Flash 0731's full context window nearly free. CSA layers fold every four token positions into one cached entry, HCA layers fold 128 into one, and both keep a 128-token uncompressed window so recent tokens stay at full resolution. Nothing about that stack changed between the preview and 0813, since the only thing DeepSeek touched was post-training.
KV cache only, at FP16. Weights and activation overhead sit on top of these figures.
Filling the entire 1,048,576-token window costs about 13.1 GB of KV cache; the same 61-layer stack storing one entry per token would want roughly 131.0 GB instead, the identical 10x saving DeepSeek V4 Flash 0731 gets from the same design. It just doesn't move the needle the way it does for Flash. The recommended Q2_K quantization already totals 609.6 GB in weights alone, so the roughly 118 GB the compression saves is real but small next to a model that needs a multi-GPU datacenter node before context length is even a consideration. Going from an 8k session to the full million-token window moves the Q2_K total from 682.9 GB to 697.4 GB, about a 2% difference.
VRAM at each quantization
Figures below assume 8k context; KV cache grows linearly as context length increases.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 6400.0 GB | 0.10 GB | 7168.1 GB |
| BF16 | 3200.0 GB | 0.10 GB | 3584.1 GB |
| FP16 | 3200.0 GB | 0.10 GB | 3584.1 GB |
| Q8_0 | 1700.8 GB | 0.10 GB | 1905.0 GB |
| Q6_K | 1313.6 GB | 0.10 GB | 1471.3 GB |
| Q5_K_M | 1139.2 GB | 0.10 GB | 1276.0 GB |
| Q4_K_M | 974.4 GB | 0.10 GB | 1091.4 GB |
| Q3_K_M | 769.6 GB | 0.10 GB | 862.1 GB |
| Q2_Krec | 609.6 GB | 0.10 GB | 682.9 GB |
| NVFP4cuda | 800.0 GB | 0.10 GB | 896.1 GB |
KV cache figures assume 8k context at FP16. NVFP4 quantization requires a CUDA-capable GPU. Enable TurboQuant in the calculator to see reduced KV cache estimates.
Benchmarks
DeepSeek V4 Pro 0813 on Terminal-Bench 2.1: second only to a model 75% larger
Terminal-Bench 2.1 is the one benchmark DeepSeek has published specifically for the 0813 checkpoint: 87.9, from its own comparison against Claude Fable 5. Plotted by total parameter count against every other model tracked here that reports the same benchmark, V4 Pro 0813 sits second, behind only Kimi K3, a model 75% larger by total parameters, and just ahead of Qwen3.8 2.4T-A95B, which is over 50% larger still. It barely clears its own 284B Flash sibling: DeepSeek V4 Flash 0731 scores 82.7 with less than a fifth of Pro's total parameters and about a quarter of its active parameters.
GPUs that run DeepSeek V4 Pro 0813 1.6T natively (0)
No single GPU in our list fits this model at Q4 with 8k context. Browse all GPUs or try multi-GPU setups.
Notes
The official release build DeepSeek's API quietly switched to on August 12, 2026, replacing the April preview with no separate announcement, changelog entry, or new weights upload; the version string on the pricing page is the only public confirmation. Architecture is unchanged: 1.6T total parameters, 49B active, 61 layers, hybrid Compressed Sparse Attention and Heavily Compressed Attention, a 1M-token context window and 384K max output, so the VRAM math is identical to the preview. DeepSeek attributes the gains entirely to post-training, and on its own nine-benchmark comparison against Claude Fable 5 the jumps are large on agentic evals not tracked in this site's Benchmarks fields: DeepSWE rose from 12.8 to 62.7 and CyberGym from 52.7 to 83.3. Terminal-Bench 2.1 landed at 87.9, DeepSeek's own run. On the same nine-benchmark table Claude Fable 5 still leads by an average of 5.3%, narrowing to about 2.8% once Humanity's Last Exam (the single largest gap) is set aside, with Claude Fable 5's API pricing roughly 46x DeepSeek's blended rate. Pricing itself did not change: $0.435 per million input tokens, $0.87 per million output, same as the preview. No independently vetted quantized GGUF build exists for this size at the time of writing; a few community conversions claim Q2_K-class builds in the 400-500GB range, but none carry the kind of verification this site requires to publish as a fixed build.
Compare DeepSeek V4 Pro 0813 1.6T with other models
Should you run DeepSeek V4 Pro 0813 locally?
Almost certainly not, whatever your budget. The recommended Q2_K quantization needs about 609.6 GB of weights alone, more than the 512 GB of unified memory in the largest single machine this site tracks, a Mac Studio M3 Ultra. There's no lower rung to fall back to, either: Q2_K is the floor of the standard quantization ladder, and unlike DeepSeek V4 Flash 0731, no independently vetted quantized build exists yet for this checkpoint, only a handful of unverified community conversions claiming Q2_K-class sizes in the 400 to 500 GB range. Actually reaching this model locally means pooling several datacenter GPUs, the same category of hardware Qwen3.8 2.4T-A95B and Nemotron 3 Ultra need. And there's a specific reason not to bother even with that hardware in hand: on DeepSeek's own comparison against Claude Fable 5, Fable 5 leads by an average of 5.3% across nine shared benchmarks, while DeepSeek's blended API rate runs at roughly 1/46th of Fable 5's. If the API is the realistic option either way, buying or renting a multi-GPU node to self-host this specific model is hard to justify outside of auditability, fine-tuning, or an air-gapped requirement.
Frequently asked questions
- What are the VRAM requirements for DeepSeek V4 Pro 0813 1.6T?
- DeepSeek V4 Pro 0813 1.6T requires approximately 1091.4 GB of VRAM at Q4_K_M quantization, 1905.0 GB at Q8, and 3584.1 GB at FP16. These numbers assume 8k context window; VRAM scales linearly with context length due to the KV cache.
- How many parameters does DeepSeek V4 Pro 0813 1.6T have?
- DeepSeek V4 Pro 0813 1.6T has 1600 billion total parameters, but only 49 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 DeepSeek V4 Pro 0813 1.6T run on a 16 GB GPU?
- No. At Q4_K_M, DeepSeek V4 Pro 0813 1.6T needs 1091.4 GB of VRAM — more than 16 GB. You will need a multi-GPU server.
- Can DeepSeek V4 Pro 0813 1.6T run on a 24 GB GPU?
- No. Even at Q4_K_M, DeepSeek V4 Pro 0813 1.6T needs 1091.4 GB. Consider a multi-GPU server with 1092 GB+ of combined VRAM.
- What is the smallest quantization for DeepSeek V4 Pro 0813 1.6T that fits in 24 GB of VRAM?
- DeepSeek V4 Pro 0813 1.6T cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 682.9 GB at Q2_K.
- What GPU do I need to run DeepSeek V4 Pro 0813 1.6T locally?
- You need a multi-GPU server. At Q4_K_M, DeepSeek V4 Pro 0813 1.6T needs 1091.4 GB VRAM, more than any single consumer GPU. That's roughly 14x 80 GB datacenter GPUs (H100, A100, or similar) pooled together.
- How much VRAM does DeepSeek V4 Pro 0813 need?
- About 682.9 GB at a short 8k context and 697.4 GB with the entire 1,048,576-token window filled, both at the recommended Q2_K quantization: roughly 609.6 GB of weights plus KV cache and activation overhead. That's larger than any single machine this site tracks, including a 512 GB Mac Studio M3 Ultra, so this figure only becomes reachable by pooling several datacenter GPUs.
- Can DeepSeek V4 Pro 0813 run on a Mac Studio or a single GPU?
- No. Even the largest unified-memory machine this site tracks, a 512 GB Mac Studio M3 Ultra, falls roughly 100 GB short of the smallest standard quantization, Q2_K, at around 609.6 GB of weights alone. No single consumer or prosumer GPU comes remotely close either. Unlike DeepSeek V4 Flash 0731, there is no independently vetted low-bit GGUF build to fall back on yet, only unverified community conversions.
- What changed between the April preview and DeepSeek V4 Pro 0813?
- Post-training only. The architecture is identical: 1.6T total parameters, 49B active, 61 layers, hybrid CSA/HCA attention, and a 1M-token context window, so the VRAM math didn't move. DeepSeek's own comparison against Claude Fable 5 shows the re-post-train's size on agentic benchmarks: DeepSWE went from 12.8 to 62.7 and CyberGym from 52.7 to 83.3. The switch itself was quiet, a version-string change on the API pricing page on August 12, 2026 with no separate announcement.
- Does DeepSeek V4 Pro 0813 support image input?
- No. It's a text-only model, despite the April preview's own record on this site listing vision and video support. DeepSeek's V4-Pro-DSpark model card lists text generation only, and no public DeepSeek API exposes image input for either V4 Pro or V4 Flash.
- How does DeepSeek V4 Pro 0813 compare to Claude Fable 5?
- Close on capability, far apart on price. Across the nine benchmarks DeepSeek used for its own comparison, Claude Fable 5 leads by an average of 5.3%, narrowing to about 2.8% once Humanity's Last Exam, the single largest gap, is set aside; DeepSeek wins outright on two of the nine. The bigger number is the price: Claude Fable 5's API runs roughly 46 times DeepSeek's blended rate, so V4 Pro 0813 gets you most of the agentic capability for a small fraction of the cost, provided you're fine using DeepSeek's API rather than self-hosting.
- Is DeepSeek V4 Pro 0813 worth running over DeepSeek V4 Flash 0731?
- Not if you're self-hosting. Flash 0731 scores 82.7 on Terminal-Bench 2.1 against Pro 0813's 87.9, a 5-point gap, while needing roughly a sixth of the weights (96.8 GB at UD-Q2_K_XL versus Pro's ~609.6 GB at Q2_K) and running on hardware that actually exists outside a datacenter. Pro's edge only makes sense through the API, where the extra capability costs a bit more per million tokens rather than several hundred extra gigabytes of memory.