DeepSeek V4.1 Flash 1017 552B
DeepSeek V4.1 Flash 1017 552B needs roughly 571.6 GB VRAM at Q4_K_M quantization (1236.5 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
- Q4_K_M total
- 571.6 GB
- at 8k context
- KV cache, full context
- 2.3 GB
- 40 layers cache, ~37× fewer entries than tokens
- Inputs
- Text · Image
- MIT, released October 2026
- MoE routing
- 6 of 384 experts
- 16B of 552B active on decode
- Prefill vs decode
- 8B / 16B active
- same design as the September build
- Update from Sept 10
- Post-training only
- same config, same 510.3 GB of weights
DeepSeek V4.1 Flash 1017 552B is a Mixture of Experts (MoE) model with 552B total parameters but only 16B active per token developed by DeepSeek. Released October 17, 2026 as what DeepSeek's Hugging Face card calls a minor update to DeepSeek V4.1 Flash; DeepSeek's API changelog says V4.1 Flash "has been upgraded" to it, so the deepseek-flash model name now serves this build. The shape did not change. Its configuration file is identical to the September 10 release's, so the model size and VRAM requirements are the same: a 552B-parameter mixture of experts (each token is routed through only 6 of 384 specialist sub-networks, which is why just 16B parameters do work on any one token), 40 layers split into a 20-layer encoder and a 20-layer decoder, 8B parameters active while it reads your prompt (prefill) against 16B while it writes the answer (decode), a 1M-token context window, and native image input. What changed is post-training, aimed at agents: DeepSeek says the build reduces task execution time and output length, and names its own agent harness plus Tencent's CodeBuddy and WorkBuddy as the tools it was tuned for (it recommends harness version 0.2.1-rc.1 or later). Recommended sampling is temperature 1.0 and top_p 1.0, with an output budget of at least 256K tokens for long agentic runs.
To run DeepSeek V4.1 Flash 1017 552B locally: The weights are the binding constraint, and nothing about them got smaller with this update. DeepSeek's native release is 510.3GB across 48 shards, and this site counts all of it at every quantization level, exactly as it does for the September build, so the page's ladder shows the same size from Q4_K_M down to Q2_K. At an 8k context the total is about 571.6GB and at the full 1M tokens about 574.1GB, because the compressed attention holds a full-window KV cache to about 2.3GB (against about 85.9GB if the same 40 layers stored one entry per token). No single GPU or unified-memory machine this site tracks holds that, a 512GB Mac Studio M3 Ultra included, so the realistic native setup is a datacenter node: eight 80GB H100s or four 180GB B200s clear the total with room for the full context, while four 141GB H200s (564GB) do not. About 203GB of the file is Engram tables (large banks of learned memory the model looks up by token rather than computes with), which DeepSeek's technical report says its own serving stack keeps in host memory and prefetches as needed, so a node smaller than this site's total may work in practice; this site has not verified such a setup. Community conversions of the new weights appeared within hours of release (an NVFP4 repository of 429.4GB and an MLX build of 432.0GB, with a GGUF repository still empty), none verified here, and the September builds do not carry over because the weights changed. Until one is verified, DeepSeek's API is the simplest way to use this model. As a MoE model, inference speed depends on active parameters (16B) rather than total size.
Every score here is DeepSeek's own, and no independent lab had re-scored this build when this page was written, so read the gains as vendor-reported. On the table DeepSeek published with the release, the new build beats the September one on every row the two share, and the gains are uneven: Terminal-Bench 4.0 rose from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and Humanity's Last Exam without tools from 36.8 to 42.6, while DeepSWE v1.1, where the September build already scored 74.2, moved only to 77.0. The vision rows rose too (Chartography 78.9 to 83.0, BabyVision 89.6 to 92.5), though those are with-tools runs. Two cautions on reading the table. Terminal-Bench 4.0 is a newer suite than the Terminal-Bench 2.1 the September card quoted, so its 31.2 and 42.3 cannot be compared with that card's 90.6. And the 1017 card has no GPQA Diamond row: the 90.9 this site lists for the build is the September card's, carried over so the two rank level, not a score DeepSeek measured on 1017.
DeepSeek V4.1 Flash 1017 VRAM at 1M context: about 2.3 GB of KV cache, not 86 GB
On most models, filling a million-token window is where a local setup falls apart: the KV cache (the running memory of everything already read) grows with every token and swamps the weights. DeepSeek's technical report for this architecture gives the global KV cache as 890 bytes per token, which taken literally would be under 1 GB for the whole window. This site instead scales its own estimate for DeepSeek V4 Flash 0731 down by the report's roughly 4x, which lands higher, so the figures here lean cautious. The 1017 update keeps the September build's architecture, so its cache is the same.
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 2.3 GB of KV cache on this site's math, and under 1 GB if DeepSeek's 890-byte figure is taken literally. The same 40-layer stack storing one entry per token would want about 85.9 GB instead, more than the weights of most models this site tracks. Either way, context is not the constraint on this model. The weights are (see the verdict below).
VRAM at each quantization
Numbers here are computed at 8k context. Because KV cache grows linearly with context length, expect higher totals at longer sequence lengths.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 2208.0 GB | 0.02 GB | 2473.0 GB |
| BF16 | 1104.0 GB | 0.02 GB | 1236.5 GB |
| FP16 | 1104.0 GB | 0.02 GB | 1236.5 GB |
| Q8_0 | 586.8 GB | 0.02 GB | 657.2 GB |
| Q6_K | 510.3 GB | 0.02 GB | 571.6 GB |
| Q5_K_M | 510.3 GB | 0.02 GB | 571.6 GB |
| Q4_K_Mrec | 510.3 GB | 0.02 GB | 571.6 GB |
| Q3_K_M | 510.3 GB | 0.02 GB | 571.6 GB |
| Q2_K | 510.3 GB | 0.02 GB | 571.6 GB |
| NVFP4cuda | 510.3 GB | 0.02 GB | 571.6 GB |
Shown at 8k context with FP16 KV cache. NVFP4 needs a CUDA GPU to run. Toggle TurboQuant in the calculator to view compressed KV cache numbers.
Benchmarks
DeepSeek V4.1 Flash 1017 vs the September build: biggest gains on terminal and repo-level work
Same weights, same size, same VRAM bill: the only thing the October 17 update changed is post-training, so the useful question is what it bought. DeepSeek published a side-by-side table with the release, and the new build is ahead on every row both cards share. The gains are lopsided, though, and they follow what DeepSeek says it tuned for: terminal and repository-level agent tasks, not general knowledge.
DeepSeek's Hugging Face card for the 1017 build and its API changelog, both 17 October 2026, run on DeepSeek's own harnesses (Claude Code for Terminal-Bench 4.0, Mini-SWE-V2 for DeepSWE v1.1). Vendor-reported, and no independent lab had re-run them when this page was written. Apex Bench is left out because DeepSeek's two pages name it differently and disagree on the September figure.
The two largest jumps are Terminal-Bench 4.0, from 31.2 to 42.3, and ProgramBench, from 20.3 to 29.1. DeepSWE v1.1 and NL2Repo-Bench moved least among the coding rows, each by 2.8 points (DeepSWE from 74.2 to 77.0), and the September build was already strong on both. If you use this model for long terminal sessions the difference should show; for plain chat there is little here to chase. Terminal-Bench 4.0 is a newer suite than the 2.1 version the September card quoted, so these numbers cannot be set against that card's 90.6.
GPUs that run DeepSeek V4.1 Flash 1017 552B natively (0)
No single GPU in our list fits this model at Q4 with 8k context. Browse all GPUs or try multi-GPU setups.
Cloud GPU Rental
Cheapest way to run DeepSeek V4.1 Flash 1017 552B: 8× A100 80GB at Q6_K, ~54 tok/s, from $9.52/hr.
Pay by the hour · no contract · pods start in about a minute.
Rent this setup on RunPod ↗ (+$5 signup credit)Prices as of Oct 17, 2026. Rates change; check RunPod for current pricing.
Affiliate link: CanItRun may earn a commission. Doesn't affect the fit calculation above.
Notes
Released October 17, 2026 as a minor update to DeepSeek V4.1 Flash; DeepSeek's API changelog says V4.1 Flash has been upgraded to it, so the deepseek-flash model name now serves this build. The shape did not change: the Hugging Face config is identical to the September 10 release's and the weights are the same 510.3GB across the same 48 shards, so the VRAM requirements are unchanged and every gain comes from post-training. DeepSeek reports gains across its own table, led by Terminal-Bench 4.0 (31.2 to 42.3), ProgramBench (20.3 to 29.1) and Humanity's Last Exam without tools (36.8 to 42.6), and says the build finishes tasks faster with shorter outputs. These are DeepSeek's own runs, and no independent lab had re-scored this build when this page was written. The GPQA Diamond 90.9 on this record is the September card's, carried over (not re-run) so the build ranks level with its sibling, because the 1017 card has no GPQA row; Terminal-Bench is left empty because the card reports Terminal-Bench 4.0, a different suite from the Terminal-Bench 2.1 this site tracks. The catch for local use is the weights: the native release is about 510GB, this site counts all of it at every quantization level, and no community conversion of the new weights has been verified.
Compare DeepSeek V4.1 Flash 1017 552B with other models
How to run DeepSeek V4.1 Flash 1017 552B locally
Q4_K_M needs 571.6 GB: needs multiple datacenter-class GPUs (80 GB+ each).
vLLM or SGLang on a multi-GPU datacenter node (DeepSeek's own commands, native checkpoint)
# vLLM, from DeepSeek's Hugging Face card
vllm serve "deepseek-ai/DeepSeek-V4.1-Flash-1017"
# SGLang, also from the card
python3 -m sglang.launch_server \
--model-path "deepseek-ai/DeepSeek-V4.1-Flash-1017" \
--host 0.0.0.0 --port 30000
# The card sets no parallelism flag. To span several GPUs, add:
# vLLM: --tensor-parallel-size 8
# SGLang: --tp 8
# Sampling: temperature 1.0, top_p 1.0
# Output budget: DeepSeek recommends at least 256K tokensWhy this quantization? There is no ladder to climb yet. DeepSeek's native checkpoint, 510.3GB across 48 shards, is the only build of the 1017 weights this site has verified, so every row of the quantization table shows that same size and Q4_K_M is a placeholder label here, not a file you can download. About 203GB of it is Engram lookup tables that DeepSeek's own serving stack keeps in host memory. Community conversions of the new weights appeared within hours of release but are unverified, and the September builds were made from different weights, so none of them carry over.
Who is DeepSeek V4.1 Flash 1017 552B for?
Teams with a multi-GPU datacenter node (eight 80GB H100s or four 180GB B200s clear this site's total at full context; four 141GB H200s do not) who want DeepSeek's newest open agent model on their own hardware, and anyone content to call DeepSeek's API instead of hosting it. It is not a single-workstation model today.
Best for
- Long-horizon coding and terminal agents, the area DeepSeek says the 1017 update targeted (Terminal-Bench 4.0 from 31.2 to 42.3 on its own table)
- Agent workflows in DeepSeek's own harness or in Tencent's CodeBuddy and WorkBuddy, the tools DeepSeek names as optimization targets
- Reading screenshots, charts and documents through native image input, with DeepSeek-reported gains on Chartography and BabyVision
- Million-token context work: on this site's math the compressed attention holds a full window to about 2.3GB of KV cache
- Self-hosted deployment under MIT, with no revenue or attribution strings attached
Not ideal for
- Any single GPU or Mac: the native release totals about 572GB even at an 8k context, past a 512GB Mac Studio M3 Ultra
- llama.cpp, Ollama or LM Studio users: mainline llama.cpp had no V4.1 support as of 17 October 2026, and there is no verified GGUF of the 1017 weights
- Anyone who needs independently verified numbers: every score is DeepSeek's own, and no lab had re-run them when this page was written
- Anyone who needs a quantized checkpoint they can load today: the community conversions are hours old and unverified
Should you run DeepSeek V4.1 Flash 1017 locally, today?
Only on a datacenter node, and only on DeepSeek's own checkpoint. The update did not make the model any smaller: the native release is 510.3 GB, and this page's quantization table counts all of it at every rung, so a Q4_K_M row shows the same size as Q2_K. That totals about 571.6 GB at an 8k context and about 574.1 GB at the full million tokens, past even a 512 GB Mac Studio M3 Ultra, the largest single unified-memory machine this site tracks. Eight 80 GB H100s or four 180 GB B200s clear it with room for full context, and four 141 GB H200s do not. No verified quantized build of the new weights exists yet, and mainline llama.cpp had no V4.1 support as of 17 October 2026. If the question is whether to move from the September build, DeepSeek's own table says yes for agent work, with the caveat that it is DeepSeek's table. For everyone else, the API's deepseek-flash name is the simplest way in, and DeepSeek says its pricing is unchanged.
Frequently asked questions
- What are the VRAM requirements for DeepSeek V4.1 Flash 1017 552B?
- DeepSeek V4.1 Flash 1017 552B requires approximately 571.6 GB of VRAM at Q4_K_M quantization, 657.2 GB at Q8, and 1236.5 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 V4.1 Flash 1017 552B have?
- DeepSeek V4.1 Flash 1017 552B has 552 billion total parameters, but only 16 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 V4.1 Flash 1017 552B run on a 16 GB GPU?
- No. At Q4_K_M, DeepSeek V4.1 Flash 1017 552B needs 571.6 GB of VRAM, more than 16 GB. You will need a multi-GPU server.
- Can DeepSeek V4.1 Flash 1017 552B run on a 24 GB GPU?
- No. Even at Q4_K_M, DeepSeek V4.1 Flash 1017 552B needs 571.6 GB. Consider a multi-GPU server with 572 GB+ of combined VRAM.
- What is the smallest quantization for DeepSeek V4.1 Flash 1017 552B that fits in 24 GB of VRAM?
- DeepSeek V4.1 Flash 1017 552B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 571.6 GB at Q2_K.
- What GPU do I need to run DeepSeek V4.1 Flash 1017 552B locally?
- You need a multi-GPU server. At Q4_K_M, DeepSeek V4.1 Flash 1017 552B needs 571.6 GB VRAM, more than any single consumer GPU. That's roughly 8x 80 GB datacenter GPUs (H100, A100, or similar) pooled together.
- Can I run DeepSeek V4.1 Flash 1017 552B in the cloud?
- Yes. The cheapest RunPod setup we found that holds DeepSeek V4.1 Flash 1017 552B entirely in GPU memory at 8k context is 8× A100 80GB at Q6_K, at roughly 53.9 tokens per second and from $9.52 per hour for the whole pod. Prices as of Oct 17, 2026; RunPod rates change, so check their pricing page for the current figure.
- Does DeepSeek V4.1 Flash 1017 need all 510 GB in GPU memory?
- On this site's numbers, yes: it counts the whole 510.3 GB file, which totals about 571.6 GB at an 8k context and about 574.1 GB at the full 1,048,576 tokens once KV cache and overhead are added. Eight 80 GB H100s or four 180 GB B200s clear that, and four 141 GB H200s (564 GB) fall just short. DeepSeek's own design is less strict. About 203 GB of the file is Engram lookup tables that its technical report says the serving stack keeps in host memory and prefetches as needed, so a smaller node may work in practice. This site has not verified a setup like that, and context is not the issue either way: the compressed attention holds the whole window to about 2.3 GB of KV cache.
- What changed between DeepSeek V4.1 Flash and the 1017 update?
- Only the post-training. The configuration file is identical to the September 10 release's and the weights are the same 510.3 GB, so the architecture, the 1M-token context window, the image input and the VRAM requirements are all unchanged. DeepSeek reports gains across its own table, led by Terminal-Bench 4.0 (31.2 to 42.3), ProgramBench (20.3 to 29.1) and Humanity's Last Exam without tools (36.8 to 42.6), and says the build finishes tasks faster with shorter outputs. The API's deepseek-flash name now points to the new build.
- Can DeepSeek V4.1 Flash 1017 run on a Mac Studio or a single GPU?
- Not on this site's numbers. The native release totals about 571.6 GB even at an 8k context, past a 512 GB Mac Studio M3 Ultra and every single GPU tracked, the 288 GB B300 included. Community conversions of the new weights, an NVFP4 repository of 429.4 GB and an MLX build of 432.0 GB, appeared within hours of release but are unverified, and this site has no independent speed measurements for this build yet.
- Does DeepSeek V4.1 Flash 1017 work with llama.cpp, Ollama, and LM Studio?
- Not yet, as far as this site could verify. As of 17 October 2026, mainline llama.cpp had no DeepSeek V4.1 support (pull request 28696 on ggml-org/llama.cpp was still open, with its latest activity on 13 October), and a GGUF repository for the 1017 weights was empty on release day. DeepSeek's own card gives vLLM and SGLang commands for the native checkpoint. The September build's community conversions were made from different weights and do not apply.