DeepSeek-V4.1-Flash on DGX Spark (GB10): 128 GB unified, Does not fit, Rough estimate
DeepSeek-V4.1-Flash VRAM lists each catalog quant at 8,192 tokens.
GB10 is the superchip named on this record, NVIDIA DGX Spark (GB10, 128 GB unified). 128 GB LPDDR5x unified memory on the GB10 Superchip.
Does not fit. DeepSeek-V4.1-Flash FP8 safetensors on one DGX Spark (GB10) (128 GB LPDDR5x coherent unified system memory) estimates 476.18 GiB–549.77 GiB at 4,096 tokens, 1 sequence, FP16 KV, vllm, against 112.00 GiB net unified memory. The weight file is 510,286,023,000 bytes. Rough estimate. V4.1-Flash official packed weights on Spark. Coverage is partial: MLA KV is a latent bound, vision/Engram/MTP bytes are not separable, so this is a weights-lower-bound plus incomplete cache — not a full fit claim.
DeepSeek V4.1 Flash on DGX Spark (128 GB unified): Does not fit. DeepSeek-V4.1-Flash on DGX Spark (GB10). Estimate 476.18 GiB–549.77 GiB against 112.00 GiB net (FP8 safetensors, 4,096 tokens).
Conditional memory judgement for one configuration. Change context, quant, or engine and the conclusion can change. Configuration used on this page: FP8 safetensors, 4,096 cached tokens, 1 sequence, engine vllm. Status: Insufficient memory. Software: Not verified..
The generic calculator judges the canonical checkpoint. 3 community-tested deployment recipes are also listed for this setup.
Assumptions that affect this result
- Unified memory is a shared CPU/GPU pool. Host RAM is not added on top of this capacity.
- 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.
- Weight bytes come from the selected weight files only, not from every file in the repository.
- Full-resident weight estimates use total parameters, including inactive MoE experts.
- This checkpoint has a vision component. Text-only inference does not include vision encoder or projector bytes.
- Engram (or similar extra memory module) is present and not included in this text-backbone figure.
- Multi-token prediction / draft weights exist and are not included unless enabled.
- MLA KV is not estimated with the GQA formula.
- Context budget is the number of cached tokens per sequence (prompt plus reserved generation).
- vLLM often preallocates KV from remaining GPU memory. This estimate reports needed KV plus CUDA/graph extras, not the allocator leftover grab.
- Prefill peak and steady decode occupancy are not the same. The high scenario is closer to prefill/workspace pressure.
Model and recipes
Known working setups
3 known setupsCommunity and maintainer-tested deployment recipes for this model on this hardware platform.
- ✓ 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.
Fixed page configuration
V4.1-Flash official packed weights on Spark. Coverage is partial: MLA KV is a latent bound, vision/Engram/MTP bytes are not separable, so this is a weights-lower-bound plus incomplete cache — not a full fit claim.
- ModelDeepSeek-V4.1-Flash
- HardwareNVIDIA DGX Spark (GB10, 128 GB unified)
- QuantizationFP8 safetensors
- Context length (cached tokens)4,096
- Engine
vllm - Concurrent sequences1
- KV cache typefp16
Technical identifiers and share data
- Model
deepseek-v4.1-flash - Hardware
nvidia-dgx-spark-128gb-unified - Quantization
fp8 - Share fragment for this identity
v2.eyJzY2hlbWFWZXJzaW9uIjoyLCJjYWxjdWxhdG9yVmVyc2lvbiI6IjAuMy4wIiwiZGF0YXNldFZlcnNpb24iOiIyMDI2LTEwLTAyLjEiLCJjb25maWciOnsibW9kZWxJZCI6ImRlZXBzZWVrLXY0LjEtZmxhc2giLCJoYXJkd2FyZUlkIjoibnZpZGlhLWRneC1zcGFyay0xMjhnYi11bmlmaWVkIiwiZGV2aWNlQ291bnQiOjEsInF1YW50aXphdGlvbklkIjoiZnA4IiwiY29udGV4dFRva2VucyI6NDA5NiwiY29uY3VycmVudFNlcXVlbmNlcyI6MSwiZW5naW5lIjoidmxsbSIsImt2VHlwZSI6ImZwMTYifX0
Memory breakdown
- Weight artifact
- 475.24 GiB (510,286,023,000 bytes)
- Raw KV cache
- 0.31 GiB (335,544,320 bytes)
- Engine extras / workspace
- 0.63 GiB–74.22 GiB
- Total estimated demand
- 476.18 GiB–549.77 GiB
- Net calculator budget
- 112.00 GiB
- OS / display reserve
- 16.00 GiB
- Remaining vs net budget
- -437.77 GiB–-364.18 GiB
549.77 GiB high estimate vs 112.00 GiB net
Enough memory does not mean this engine, quant, and OS will run. Support does not mean it will be fast.
Other published files on this device
Same model, device and context (4,096 tokens, 1 sequence, FP16 KV). Each row is one published weight-file set with its usual engine, sorted by memory.
| Quantization | Engine | Estimated memory | Fit |
|---|---|---|---|
| FP8 safetensors This page | vllm | 476.18 GiB–549.77 GiB | Does not fit · Rough estimate |
Related pages
Sources
- DeepSeek-V4.1-Flash model card
- config.json — revision dba1be0a40aa45a94ad051997016db3960a90277
- safetensors index — metadata.total_size=510286023000
- NVIDIA DGX Spark product page — 128 GB LPDDR5x coherent unified system memory. This is not dedicated VRAM. Default reserve deducts OS and other processes. Multiple Sparks are not one pool.
- NVIDIA DGX Spark hardware docs — 128 GB LPDDR5x unified memory on the GB10 Superchip.
Calculator
This combo page keeps the route model and device. Customize to change them on the main calculator; the share then keeps that identity.
Loading catalog snapshot…
Search or pick a catalog checkpoint to see memory demand.