DeepSeek V4.1 Flash needs 550.14 GiB at 8,192 tokens
DeepSeek-V4.1-Flash FP8 safetensors (deepseek-ai/DeepSeek-V4.1-Flash) is 510,286,023,000 bytes (475.24 GiB, 510.29 GB). At 8,192 tokens, 1 sequence, FP16 KV, vllm, typical overhead, the calculator estimates 476.49 GiB–550.14 GiB. Rough estimate. Coverage is partial. The other catalog quant, INT4 theoretical, has no weight file. On NVIDIA DGX Spark (GB10, 128 GB unified), the net budget is 112.00 GiB. FP8 safetensors at 4,096 tokens estimates 476.18 GiB–549.77 GiB. The verdict is Does not fit. On GeForce RTX 5090 32GB (desktop), the net budget is 30.00 GiB. FP8 safetensors at 4,096 tokens estimates 476.18 GiB–549.77 GiB. The verdict is Does not fit. This catalog has no calibration trace and no planned benchmark for this model. The ranges are calculator estimates.
- DeepSeek-V4.1-Flash on NVIDIA DGX Spark (GB10, 128 GB unified)
- DeepSeek-V4.1-Flash on GeForce RTX 5090 32GB (desktop)
- DeepSeek-V4.1-Flash on 2× DGX Spark (EXL3 2.9 bpw)
- DeepSeek-V4.1-Flash on 3× DGX Spark (SGLang MXFP4 / FP8)
- DeepSeek-V4.1-Flash on 4× DGX Spark (1M Long-Context Setup)
- rhys101 SGLang TP4/EP4 NVMe Engram
- MiaAI-Lab SGLang TP4/EP1
VRAM by quant at 8,192 tokens
1 sequence, FP16 KV, typical overhead. Each total is the calculator for that catalog quant. This catalog has no calibration trace and no planned benchmark for this model. The ranges are calculator estimates.
| Quantization | Selected files | Engine | 8,192 tokens |
|---|---|---|---|
| FP8 safetensors | 510,286,023,000 (475.24 GiB) | vllm | 476.49 GiB–550.14 GiB |
| INT4 theoretical Theoretical rows are not downloadable artifacts. | No weight file | llama.cpp | 356.15 GiB–459.09 GiB |
Known working setups
3 known setupsCommunity and maintainer-tested deployments on real hardware with verified context, quantization, and runtime configurations.
- ✓ generation
- ✓ 262,144 context
- ✓ tool calling
- ✓ vision
- ✓ speculative decoding
2-node DGX Spark setup serving DeepSeek-V4.1-Flash with vLLM tensor-parallel size 2 over ConnectX-7 RoCE, EXL3 2.9 bpw mul1 weights, and native NVMe-resident Engram tables.
- ✓ generation
- ✓ ~208K Tested (256K cfg)
- ✓ tool calling
- ✓ vision
- ✓ speculative decoding
3-node DGX Spark cluster running native MXFP4 experts and FP8 dense weights with SGLang TP=3 over ConnectX-7 RoCE triangle, with NVMe-resident Engram tables.
- ✓ generation
- ✓ 1,048,576 context
- ✓ tool calling
- ✓ vision
- ✓ speculative decoding
4-node DGX Spark cluster running native MXFP4 MoE and FP8 dense weights with SGLang TP=4, verified 1M context needle test, and 8M pinned KV token pool.
Runs on
Publisher and provenance
- PublisherDeepSeek
- Hugging Face
deepseek-ai/DeepSeek-V4.1-Flash - Revision
dba1be0a40aa45a94ad051997016db3960a90277 - Total parameters763,205,315,794
- Active parametersUnknown / not independently sourced
- Layers40
- Query heads64
- KV heads1 (mla)
- Head dimension512
- Recorded context capability1,048,576
- Completenesspartial · partial
Active parameters describe tokens' routed compute, not the full resident model weight footprint. Full-resident weight estimates use total parameters.
Memory layers in this estimate
Artifact bytes, raw KV, engine extras, and total demand are separate. This is not a runtime trace.
- Weight artifact
- 475.24 GiB (510,286,023,000 bytes)
- Raw KV cache
- 0.63 GiB (671,088,640 bytes)
- Engine extras / workspace
- 0.63 GiB–74.28 GiB
- Total estimated demand
- 476.49 GiB–550.14 GiB
Published artifacts
| Quantization | Selected files | Bytes | Engine | Evidence |
|---|---|---|---|---|
| FP8 safetensors fp8 | 510,286,023,000 (475.24 GiB / 510.29 GB) | vllm, sglang, unspecified | file-metadata | |
| INT4 theoretical | No published file | 381,602,657,897 (theoretical) | — | theoretical |
Memory by quantization and context
Concurrent sequences = 1. KV defaults to FP16. The range includes documented runtime extras for the listed engine. These figures use the same core as the interactive calculator.
| Quantization | 2,048 tokens | 8,192 tokens | 32,768 tokens |
|---|---|---|---|
| FP8 safetensors | 476.02 GiB–549.59 GiB | 476.49 GiB–550.14 GiB | 478.37 GiB–552.36 GiB |
| INT4 theoretical Theoretical rows are not downloadable artifacts. | 355.68 GiB–458.51 GiB | 356.15 GiB–459.09 GiB | 358.02 GiB–461.40 GiB |
Raw KV contribution
Related pages
Sources
- DeepSeek-V4.1-Flash model card
- config.json — revision dba1be0a40aa45a94ad051997016db3960a90277
- safetensors index — metadata.total_size=510286023000
Assumptions that affect this result
- Requested context is within the recorded native window of 1048576 tokens.
- Full-resident weight estimates use total parameters, including inactive MoE experts.
- Activated parameter count is unknown / not independently sourced. Active parameters describe routed compute, not the full resident model weight footprint.
- Weight bytes come from the selected weight files only, not from every file in the repository.
- Download file size is not peak GPU memory. Runtime layout, allocator padding, and KV are extra.
- Weight bytes come from 48 published safetensors shards totaling 510286023000 bytes.
- This checkpoint has a vision component. Text-only inference does not include vision encoder or projector bytes.
- Multi-token prediction / draft weights exist and are not included unless enabled.
Calculator
Loading catalog snapshot…
Search or pick a catalog checkpoint to see memory demand.