DeepSeek V4.1 Flash hardware requirements: GGUF, Mac and RAM
- Page updated
- Sources checked
- File sizes from Hugging Face, read October 18, 2026; DwarfStar and Unsloth 1017 files October 20
- Hardware only where DeepSeek, NVIDIA, DwarfStar, the llama.cpp pull requests or the GGUF and MLX publishers name it; the checker matches those routes, it computes no fit
- Jump to the checker ↓
DeepSeek-V4.1-Flash-1017 is DeepSeek's October 17, 2026 update to DeepSeek-V4.1-Flash (September 2026): the same multimodal mixture-of-experts model, 552B backbone parameters with 8B active per token in prefill and 16B in decode, plus 196B parameters of Engram lookup tables, a 1M-token context window and MIT-licensed weights. DeepSeek says it improves reasoning, coding, agentic and visual benchmarks; the API name and price stay the same.
Short answer: you need about 128 GB of memory; no consumer GPU on its own has a published route. The official weights are 510.3 GB (FP8 with FP4 experts). DwarfStar, a community engine, streams its 365.7 GB Q2 file of the September weights from SSD on one 128 GB Mac or DGX Spark. For the 1017 update every path is experimental: DwarfStar published a 1017 Q2 file on October 19 that its guides do not cover yet, and Unsloth's smallest 1017 GGUF is 233.0 GB, which Unsloth estimates needs 128–145 GB of RAM plus VRAM and which loads only in its unmerged llama.cpp branch or Unsloth Desktop. Ollama and LM Studio offer the model only in their clouds.
Which DeepSeek V4.1 Flash route matches your machine?
Open in a full page ↗Loading private checker…
- Developer
- DeepSeek Official
- Params · DeepSeek V4.1 Flash September release
- 552B backbone; 8B active per token in prefill, 16B in decode; Engram conditional memory 196B Official
- Context
- 1M tokens (1,048,576) Official
- Weights
- September 10 release and the October 17 update, 510.3 GB each, not gated Official
- License
- MIT Official
- llama.cpp
- Not in mainline Community
- DwarfStar · DeepSeek V4.1 Flash September release
- streams its 365.7 GB Q2 file of the September weights from SSD on one 128 GB Mac or DGX Spark Community
- Ollama
- Ollama and LM Studio offer the model only in their clouds Official
How big are the DeepSeek V4.1 Flash files?
File sizes as Hugging Face lists them (all parts of each repository or folder, decimal GB, read October 18, 2026; DwarfStar's 1017 file and Unsloth's files October 20). Every build keeps the 196B-parameter Engram tables, which several runtimes read from disk instead of memory, so a file size is not the memory a setup needs.
| Build | Weights | Size | Runs with | Source |
|---|---|---|---|---|
| Official (FP8, FP4 experts) | 1017 and September | 510.3 GB | DeepSeek's reference inference code | DeepSeek |
| UD-Q8_K_XL GGUF | 1017 | 508.4 GB | Unsloth's llama.cpp pull request #30817 branch, Unsloth Desktop | Unsloth |
| Q8_0 GGUF | 1017 | 508.0 GB | llama.cpp pull request #28696 branch | mrfakename |
| UD-Q4_K_XL GGUF | 1017 | 500.7 GB | Unsloth's llama.cpp pull request #30817 branch, Unsloth Desktop | Unsloth |
| Q4_K_M GGUF | 1017 | 444.7 GB | llama.cpp pull request #28696 branch | mrfakename |
| oQ4e MLX | 1017 | 432.0 GB | oMLX (Mac) | mlx-community |
| NVFP4 (current revision, MXFP4 Engram tables) | September | 429.0 GB | vLLM with pull request #61530 on NVIDIA Blackwell; the previous revision (FP8 Engram tables, about 492 GiB) loads in SGLang and stock vLLM | NVIDIA |
| UD-IQ3_XXS GGUF | 1017 | 394.7 GB | Unsloth's llama.cpp pull request #30817 branch, Unsloth Desktop | Unsloth |
| oQ3e MLX | 1017 | 379.9 GB | oMLX (Mac) | mlx-community |
| DwarfStar Q2 GGUF | September and 1017 (one file each; 1017 not yet in its guides) | 365.7 GB | DwarfStar only | antirez |
| Q3_K_M GGUF | 1017 | 347.3 GB | llama.cpp pull request #28696 branch | mrfakename |
| oQ2.5e MLX | 1017 | 342.7 GB | oMLX (Mac) | mlx-community |
| UD-Q2_K_XL GGUF | 1017 | 334.8 GB | Unsloth's llama.cpp pull request #30817 branch, Unsloth Desktop | Unsloth |
| oQ2e MLX | 1017 | 313.3 GB | oMLX (Mac) | mlx-community |
| UD-IQ2_XXS GGUF | 1017 | 300.5 GB | Unsloth's llama.cpp pull request #30817 branch, Unsloth Desktop | Unsloth |
| Q2_K GGUF | 1017 | 264.5 GB | llama.cpp pull request #28696 branch | mrfakename |
| UD-IQ1_M GGUF | 1017 | 250.4 GB | Unsloth's llama.cpp pull request #30817 branch, Unsloth Desktop | Unsloth |
| UD-IQ1_S GGUF | 1017 | 233.0 GB | Unsloth's llama.cpp pull request #30817 branch, Unsloth Desktop | Unsloth |
DwarfStar's Q2 file is 340.6 GiB, of which 152 GiB are main weights and 189 GiB are Engram tables it reads from disk; its Q4 file (September weights) is 483 GiB with 294 GiB of main weights. Of the 313.3 GB oQ2e build, 122.9 GB are Engram tables, which mlx-community says oMLX can serve from SSD instead of memory below oQ4. The GGUF publisher calls Q2_K the smallest usable size and suggests keeping the Engram tables on the CPU (-ot "engram_embd.weight=CPU"). Of Unsloth's 233.0 GB UD-IQ1_S, 110.6 GB is the Engram table, which it keeps on SSD via mmap; its other sizes (UD-IQ3_S 423.6 GB, UD-Q3_K_XL 485.1 GB) are in Unsloth's 1017 GGUF repository. DwarfStar's GGUF repository also added a 1017 "MXFP4-experts" GGUF in nine parts (501.6 GB) on October 19 with no documented setup. DwarfStar's download script points its Q2 target at the 1017 file but still checks the September file's checksum for it, so it rejects the download; the checker's commands download the file directly. Other community GGUF and MLX conversions exist; the table lists the ones whose runtime and sizes we checked.
What hardware can run DeepSeek V4.1 Flash?
Only what the sources name. RigCheck does not turn these file sizes into a fit estimate: the Engram tables, SSD streaming and the compressed context cache make file size a poor guide to memory. The one memory table is Unsloth's, for its own unmerged llama.cpp branch with the Engram table on SSD (RAM plus VRAM, or unified memory): 128–145 GB at 1-bit, 176–205 GB at 2-bit, 256–290 GB at 3-bit, 305 GB at 4-bit and 315 GB at 8-bit Unsloth estimate.
| Hardware | What the sources say | Source |
|---|---|---|
| One gaming GPU (RTX 4090 24 GB, RTX 5090 32 GB) in a PC | No published route for the card. Unsloth estimates 128–145 GB of RAM plus VRAM combined for its 1-bit 1017 builds in its unmerged llama.cpp branch (experimental) and names no PC card setup; DwarfStar's CUDA support for V4.1 is named only for the DGX Spark | Unsloth, DwarfStar |
| One 128 GB Mac or one DGX Spark | DwarfStar runs its Q2 file of the September weights with SSD streaming (--ssd-streaming); keep the GGUF on a fast local SSD. Its 1017 Q2 file is the same size but not yet in its guides (experimental). Unsloth says its 1017 UD-IQ1_S (233.0 GB) works on 128 GB RAM setups such as a MacBook, DGX Spark or RTX Spark, at the top of the 128–145 GB it estimates (runtime and context can need more), in its unmerged llama.cpp branch or Unsloth Desktop (experimental). On a DGX Spark, the unmerged llama.cpp pull request #28696 also reports its own Q2_K conversion of the September weights (experimental; no published file) | DwarfStar, Unsloth, llama.cpp PR #28696 |
| Two 128 GB Macs or two DGX Sparks | DwarfStar Q2 (September weights; its 1017 file experimentally) resident across both with tensor parallelism over RDMA: about 81 GiB of main weights per machine, the complete GGUF on each machine's disk | DwarfStar |
| Mac with 256 GB or more | DwarfStar: holds all Q2 main weights (September weights; its 1017 file experimentally); its Q4 file needs SSD streaming below 512 GB. Unsloth: its 1017 UD-IQ3_XXS (394.7 GB) works on 256 GB RAM setups such as a Mac Studio (unmerged llama.cpp branch) | DwarfStar, Unsloth |
| Mac with 512 GB (M3 Ultra) | DwarfStar: Q4 fully resident, tested on an M3 Ultra with 512 GB (September weights; DwarfStar has no 1017 Q4) | DwarfStar |
| Strix Halo PC with 128 GB (e.g. Framework Desktop) | Experimental: DwarfStar's guide says ROCm supports single-host V4.1 text with SSD streaming, and its README names Strix Halo systems for ROCm; it names no memory size, command or speed for V4.1 on ROCm | DwarfStar |
| GPU server | NVIDIA's NVFP4 checkpoint (September weights) serves on four GB300 GPUs: the current 429.0 GB revision with a vLLM build that includes pull request #61530, the previous revision (about 492 GiB) with SGLang or the stock vLLM image. DeepSeek's reference inference README converts the official weights for 8-way tensor parallelism in its example (MP=8); it names no GPU model or memory size. DeepSeek's DeepInfer runs V4.1 Flash (1017 by default) on NVIDIA SM100 GPUs or Huawei Ascend 950 NPUs (charts measured on 4 GPUs for prefill, 64 for decode; see Runtimes) | NVIDIA, DeepSeek |
The September weights and the 1017 weights are separate checkpoints of the same architecture. DwarfStar says it needs its own GGUF for V4.1 Flash. On October 19, 2026 it published a 1017 Q2 file of the same size and made it the default of its Q2 download target, whose description still names SSD streaming on one 128 GB Mac; its guides, QA record and GGUF description still cover only the September file, and it publishes no 1017 speed. RigCheck therefore shows DwarfStar's setups for 1017 as experimental.
Which runtimes run DeepSeek V4.1 Flash locally?
| Runtime | Status (checked October 18, 2026; DwarfStar, llama.cpp, Unsloth and Ollama October 20; DeepInfer October 21) | Source |
|---|---|---|
| llama.cpp | Not in mainline: llama.cpp master has a deepseek4 architecture but no V4.1 support, as the 1017 GGUF publisher also says. mrfakename's 1017 files need the branch from pull request #28696 (commit 2076412); text only, the vision tower cannot be converted yet. Unsloth's 1017 files need its own branch, pull request #30817, or the Unsloth Desktop app | mrfakename, Unsloth |
| DwarfStar (antirez/ds4) | Community engine with V4.1 Flash support: text and vision on Metal (Mac), text on CUDA for the DGX Spark; on ROCm, single-host text with SSD streaming (no V4.1 memory size or speed published). Uses its own calibrated GGUFs: Q2 and Q4 of the September weights; since October 19 also a Q2 of 1017, not yet covered by its guides | DwarfStar |
| oMLX (Mac) | mlx-community publishes oQ2e to oQ4e builds of 1017 made with oMLX v0.7.1, per its model cards; they name no hardware or memory size | mlx-community |
| DeepInfer (DeepSeek) | The framework behind DeepSeek's own services, open-sourced under Apache 2.0 for offline inference; V4.1 Flash is its only supported model and 1017 its default checkpoint. This release lacks the KV cache system, prefill/decode disaggregation and load balancing that DeepSeek says a complete inference service needs. Supported hardware: NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs only, on Linux with CUDA 13 (NVIDIA) or CANN 9.2 (Ascend). RigCheck's reading of NVIDIA's list: SM100 is compute capability 10.0, the B200 and GB200; the B300 and GB300 are 10.3, and the README does not say whether they count; the RTX 50 series (12.0) and the DGX Spark (12.1) are not SM100. DeepSeek's speed charts used 4 SM100 GPUs for prefill and 64 for decode | DeepInfer README, NVIDIA CUDA GPU list |
| vLLM, SGLang | NVIDIA tests its NVFP4 checkpoint on four GB300 GPUs. Its current revision stores the Engram tables as MXFP4, which needs vLLM pull request #61530 and is not validated on SGLang; the previous revision loads in the stock vLLM deepseekv41-flash-0909 image and in SGLang's dev-cu13-dsv41 build | NVIDIA |
| Ollama | Only deepseek-v4.1-flash:cloud and :1017-cloud, which run on Ollama's servers; no local build in the library | Ollama library |
| LM Studio | Its model page offers DeepSeek V4.1 Flash through LM Studio Cloud (hosted), with no local download | LM Studio |
DeepSeek ships no Jinja chat template with 1017: prompts are built with its encoding reference code or its deepseek-recipe libraries, so a runtime has to implement that format itself.
How fast is DeepSeek V4.1 Flash locally?
Every local number we found is for the September weights; none is for 1017 (DwarfStar and Unsloth publish no 1017 speed). Each applies only to the setup named:
- One DGX Spark, SSD streaming, automatic expert cache, 32K context allocated: a 3,241-token prompt prefilled at 93–96 tokens/s (September 13–14, 2026). DwarfStar's Spark guide gives no absolute single-session decode speed for this setup; eight sessions together reached 11.0 tokens/s in total (aggregate throughput, not the speed each client sees) DwarfStar.
- Two DGX Sparks, tensor parallel over a direct RoCE link: 21.8 tokens/s decode for one session (1K prompt, 64K context allocated, 2,048 teacher-forced decode tokens, no speculation; October 2, 2026); about 18 tokens/s at longer contexts DwarfStar.
- One Mac Studio M3 Ultra with 512 GB, DwarfStar Q4 fully resident (36,864-token allocation, medians of three sweeps, 128 output tokens per step): 18.69 tokens/s generation and 341.75 tokens/s prefill at a 4,096-token context, 18.11 tokens/s generation at 32,768. No figure is published for the Q2 file, a 128 GB Mac or the M5 Ultra DwarfStar QA record.
- One DGX Spark, llama.cpp pull request #28696 branch (unmerged) with its author's own Q2_K conversion: 2.8 tokens/s with GPU offload, 2.3 without, at a 2,048-token context (reported by October 10, 2026) llama.cpp PR #28696.
Measured DeepSeek V4.1 Flash on your own hardware, or tried and it failed? Send us the numbers: paste the benchmark output or log into the form (or email it). After checking it we add it to this page, labeled as a community report and shared under CC BY 4.0.
DeepSeek V4.1 Flash specs
| Parameters | 552B backbone; 8B active per token in prefill, 16B in decode; Engram conditional memory 196B September model card; 763.2B in the checkpoint with the MTP/DSpark head (about 14B) and the vision encoder Hugging Face, mlx-community |
|---|---|
| Architecture | Causal encoder-decoder: 40 layers (20 encoder, 20 decoder); 384 routed experts and 1 shared expert per MoE layer, 6 routed experts per token; Compressed Sparse Attention 2 with sliding-window layers; DSpark speculative decoding September model card; the 1017 config.json is identical |
| Context length | 1M tokens (1,048,576) Model card, config.json |
| Context memory | DeepSeek: the global KV cache takes 890 bytes per token with FP4 KV caching, about a quarter of DeepSeek-V4-Flash's September model card |
| Input / output | Text and images in, text out Model card |
| Weights | September 10 release and the October 17 update, 510.3 GB each, not gated Hugging Face |
| License | MIT DeepSeek |
Common setups
The checker's own answer for 8 common setups (graphics cards with 32 or 64 GB of system RAM, then Macs), in the finder's quick mode (DwarfStar's routes per DeepSeek checkpoint (documented for the September weights, experimental for 1017), not llama.cpp). Answers are published route matches, not measured setups; "Unknown" means the checker has no route or source for that hardware. Your own setup: the checker.
| Setup | DeepSeek V4.1 Flash September release | DeepSeek V4.1 Flash 1017 update |
|---|---|---|
| GeForce RTX 3060 (12 GB) + 32 GB RAM | – Unknown No published DeepSeek V4.1 Flash route checked here for this device. | – Unknown No published DeepSeek V4.1 Flash route checked here for this device. |
| GeForce RTX 4060 Ti (16 GB) + 32 GB RAM | – Unknown No published DeepSeek V4.1 Flash route checked here for this device. | – Unknown No published DeepSeek V4.1 Flash route checked here for this device. |
| GeForce RTX 3090 / 3090 Ti (24 GB) + 64 GB RAM | – Unknown No published route for this hardware (September weights). DwarfStar names CUDA support for V4.1 only on the DGX Spark, and no source publishes a setup for this card (the unmerged llama.cpp pull request #28696 reports a run only on a DGX Spark). RigCheck does not turn file sizes into a fit for this model, so this is not a "does not fit" answer, only the absence of a sourced route. | – Unknown No published route for this hardware (1017 weights). DwarfStar names CUDA support for V4.1 only on the DGX Spark, and no source publishes a setup for this card (the unmerged llama.cpp pull request #28696 reports a run only on a DGX Spark). Experimental for 1017: Unsloth estimates 128–145 GB of RAM plus VRAM combined for its 1-bit builds (from 233.0 GB, Engram table on SSD) in its unmerged llama.cpp pull request #30817 branch or Unsloth Desktop, but names no setup for this card, and this checker does not ask for system RAM. RigCheck does not turn file sizes into a fit for this model, so this is not a "does not fit" answer, only the absence of a sourced route. |
| GeForce RTX 5090 (32 GB) + 64 GB RAM | – Unknown No published route for this hardware (September weights). DwarfStar names CUDA support for V4.1 only on the DGX Spark, and no source publishes a setup for this card (the unmerged llama.cpp pull request #28696 reports a run only on a DGX Spark). RigCheck does not turn file sizes into a fit for this model, so this is not a "does not fit" answer, only the absence of a sourced route. | – Unknown No published route for this hardware (1017 weights). DwarfStar names CUDA support for V4.1 only on the DGX Spark, and no source publishes a setup for this card (the unmerged llama.cpp pull request #28696 reports a run only on a DGX Spark). Experimental for 1017: Unsloth estimates 128–145 GB of RAM plus VRAM combined for its 1-bit builds (from 233.0 GB, Engram table on SSD) in its unmerged llama.cpp pull request #30817 branch or Unsloth Desktop, but names no setup for this card, and this checker does not ask for system RAM. RigCheck does not turn file sizes into a fit for this model, so this is not a "does not fit" answer, only the absence of a sourced route. |
| Apple Silicon Mac (64 GB unified) | – Unknown No published DeepSeek V4.1 Flash route checked here for this device. | – Unknown No published DeepSeek V4.1 Flash route checked here for this device. |
| Mac with M5 Max, 40-core GPU (128 GB unified) | ! Matches DwarfStar's 128 GB route: Q2 with SSD streaming (September weights) (published route match) DwarfStar runs its Q2 file (365.7 GB) on one 128 GB Mac by streaming routed experts from SSD (--ssd-streaming), with the expert cache sized automatically. DwarfStar publishes no V4.1 speed for a 128 GB Mac (its one Mac figure is Q4 resident on a 512 GB M3 Ultra). Two 128 GB machines can instead hold it resident with tensor parallelism (set machines to two). Image input works on Metal with its separate 0.97 GB vision encoder. DwarfStar is a community engine, not DeepSeek's code; these are its documented setups, not a RigCheck estimate. Keep the GGUF on a fast local SSD (the 188.83 GiB of Engram tables are always read from the file). | ! Experimental: DwarfStar's September setup applied to its 1017 file: Q2 with SSD streaming (1017 weights) (published route match) DwarfStar runs its Q2 file (365.7 GB) on one 128 GB Mac by streaming routed experts from SSD (--ssd-streaming), with the expert cache sized automatically. Two 128 GB machines can instead hold it resident with tensor parallelism (set machines to two). DwarfStar documents and tests these setups with its September file only. On October 19, 2026 it published DeepSeek-V4.1-Flash-1017-IQ2_XXS-Q2_K-imatrix.gguf (365.7 GB, the size of its September Q2; its name gives the same IQ2_XXS / Q2_K mix) and pointed its Q2 download target at it, whose help text names SSD streaming on one 128 GB Mac, tensor parallelism on two, or a larger resident Mac. RigCheck applies the September setup to it. No 1017 speed or 1017 Q4 file is published; its 1017 MXFP4-experts GGUF (501.6 GB) has no documented setup. DwarfStar's vision encoder is from the September checkpoint; it does not say whether it works with the 1017 file. DwarfStar is a community engine, not DeepSeek's code; these setups are documented for its September file, not a RigCheck estimate. Keep the GGUF on a fast local SSD (the 188.83 GiB of Engram tables are always read from the file). Experimental alternative for 1017: Unsloth says its UD-IQ1_S GGUF (233.0 GB; its 110.6 GB Engram table stays on SSD via mmap) works on 128 GB RAM setups such as a MacBook, DGX Spark or RTX Spark, and estimates 128–145 GB of RAM plus VRAM for its 1-bit builds, the top of a 128 GB machine (runtime and context can need more); it runs only in Unsloth's unmerged llama.cpp pull request #30817 branch or Unsloth Desktop, with no published speed. |
| Mac Studio M3 Ultra (256 GB unified) | ! Matches DwarfStar: Q2 main weights fully resident on a 256 GB Mac (September weights) (published route match) DwarfStar says a 256 GB or larger Mac can hold all main weights of its Q2 file (151.77 GiB of the 365.7 GB file). Its Q4 file needs SSD streaming below 512 GB. DwarfStar publishes no V4.1 speed for this setup (its one Mac figure is Q4 resident on a 512 GB M3 Ultra). Image input works on Metal with its separate 0.97 GB vision encoder. DwarfStar is a community engine, not DeepSeek's code; these are its documented setups, not a RigCheck estimate. Keep the GGUF on a fast local SSD (the 188.83 GiB of Engram tables are always read from the file). | ! Experimental: DwarfStar's September setup applied to its 1017 file: Q2 main weights fully resident on a 256 GB Mac (1017 weights) (published route match) DwarfStar says a 256 GB or larger Mac can hold all main weights of its Q2 file (151.77 GiB of the 365.7 GB file). Its Q4 file needs SSD streaming below 512 GB and is September only. DwarfStar publishes no V4.1 speed for this setup (its one Mac figure is Q4 resident on a 512 GB M3 Ultra, September weights). Experimental for 1017: Unsloth says its 3-bit UD-IQ3_XXS (394.7 GB) works on 256 GB RAM setups such as a Mac Studio, the bottom of its 256–290 GB estimate for 3-bit, in its unmerged llama.cpp pull request #30817 branch or Unsloth Desktop. DwarfStar documents and tests these setups with its September file only. On October 19, 2026 it published DeepSeek-V4.1-Flash-1017-IQ2_XXS-Q2_K-imatrix.gguf (365.7 GB, the size of its September Q2; its name gives the same IQ2_XXS / Q2_K mix) and pointed its Q2 download target at it, whose help text names SSD streaming on one 128 GB Mac, tensor parallelism on two, or a larger resident Mac. RigCheck applies the September setup to it. No 1017 speed or 1017 Q4 file is published; its 1017 MXFP4-experts GGUF (501.6 GB) has no documented setup. DwarfStar's vision encoder is from the September checkpoint; it does not say whether it works with the 1017 file. DwarfStar is a community engine, not DeepSeek's code; these setups are documented for its September file, not a RigCheck estimate. Keep the GGUF on a fast local SSD (the 188.83 GiB of Engram tables are always read from the file). |
| Mac Studio M3 Ultra (512 GB unified) | ! Matches DwarfStar: Q2 or Q4 fully resident on a 512 GB Mac (September weights) (published route match) DwarfStar names a 512 GB Mac as the resident target for its Q4 file (518.6 GB in two parts; 294.15 GiB of main weights) and says full residency was tested on an M3 Ultra with 512 GB. Its Q2 file (151.77 GiB of main weights) also fits resident on any 256 GB or larger Mac. DwarfStar's QA record for Q4 fully resident on a 512 GB M3 Ultra (36,864-token allocation, medians of three sweeps): 341.75 tokens/s prefill and 18.69 tokens/s generation at a 4,096-token context, 18.11 tokens/s generation at 32,768. That is the only published Mac speed for V4.1: M3 Ultra with Q4, none for Q2 or the M5 Ultra. Image input works on Metal with its separate 0.97 GB vision encoder. DwarfStar is a community engine, not DeepSeek's code; these are its documented setups, not a RigCheck estimate. Keep the GGUF on a fast local SSD (the 188.83 GiB of Engram tables are always read from the file). | ! Experimental: DwarfStar's September setup applied to its 1017 file: Q2 main weights fully resident on a 512 GB Mac (1017 weights) (published route match) DwarfStar says a 256 GB or larger Mac can hold all main weights of its Q2 file (151.77 GiB of the 365.7 GB file). Its Q4 file needs SSD streaming below 512 GB and is September only. DwarfStar publishes no V4.1 speed for this setup (its one Mac figure is Q4 resident on a 512 GB M3 Ultra, September weights). Experimental for 1017: Unsloth says its 3-bit UD-IQ3_XXS (394.7 GB) works on 256 GB RAM setups such as a Mac Studio, the bottom of its 256–290 GB estimate for 3-bit, and estimates 315 GB of RAM plus VRAM for its 8-bit build (508.4 GB), in its unmerged llama.cpp pull request #30817 branch or Unsloth Desktop. DwarfStar documents and tests these setups with its September file only. On October 19, 2026 it published DeepSeek-V4.1-Flash-1017-IQ2_XXS-Q2_K-imatrix.gguf (365.7 GB, the size of its September Q2; its name gives the same IQ2_XXS / Q2_K mix) and pointed its Q2 download target at it, whose help text names SSD streaming on one 128 GB Mac, tensor parallelism on two, or a larger resident Mac. RigCheck applies the September setup to it. No 1017 speed or 1017 Q4 file is published; its 1017 MXFP4-experts GGUF (501.6 GB) has no documented setup. DwarfStar's vision encoder is from the September checkpoint; it does not say whether it works with the 1017 file. DwarfStar is a community engine, not DeepSeek's code; these setups are documented for its September file, not a RigCheck estimate. Keep the GGUF on a fast local SSD (the 188.83 GiB of Engram tables are always read from the file). |
Machine-readable: the same answers as JSON (CC BY 4.0, attribution "RigCheck").
Related hardware
Graphics card and Mac pages where this page's checker gives an answer, smallest memory first: graphics cards with 64 GB of system RAM, Macs with their unified memory, in the finder's quick mode (published route match; each page lists every model).
- Mac with M5 Max (128 GB): DeepSeek V4.1 Flash September release: Matches DwarfStar's 128 GB route: Q2 with SSD streaming (September weights); DeepSeek V4.1 Flash 1017 update: Experimental: DwarfStar's September setup applied to its 1017 file: Q2 with SSD streaming (1017 weights)
- Mac Studio M3 Ultra (256 GB): DeepSeek V4.1 Flash September release: Matches DwarfStar: Q2 main weights fully resident on a 256 GB Mac (September weights); DeepSeek V4.1 Flash 1017 update: Experimental: DwarfStar's September setup applied to its 1017 file: Q2 main weights fully resident on a 256 GB Mac (1017 weights)
- Mac Studio M3 Ultra (512 GB): DeepSeek V4.1 Flash September release: Matches DwarfStar: Q2 or Q4 fully resident on a 512 GB Mac (September weights); DeepSeek V4.1 Flash 1017 update: Experimental: DwarfStar's September setup applied to its 1017 file: Q2 main weights fully resident on a 512 GB Mac (1017 weights)
All graphics cards · All Macs · Any other machine
FAQ
How much RAM or VRAM does DeepSeek V4.1 Flash need?
About 128 GB at the least. The official weights are 510.3 GB. DwarfStar runs its 365.7 GB Q2 file of the September weights on one 128 GB Mac or DGX Spark by streaming most of it from SSD, or resident with about 81 GiB of main weights on each of two 128 GB machines. For 1017, Unsloth estimates 128–145 GB of RAM plus VRAM for its 1-bit GGUF (233.0 GB, the 110.6 GB Engram table on SSD) in its unmerged llama.cpp branch (experimental); the smallest 1017 GGUF we checked from another publisher is 264.5 GB (mrfakename Q2_K).
Can DeepSeek V4.1 Flash run on an RTX 4090 or RTX 5090?
Not with a published route for the card on October 20, 2026. Most of a model this size would sit in system RAM: Unsloth's table counts RAM plus VRAM, 128–145 GB for its 1-bit 1017 GGUF, but that build loads only in Unsloth's unmerged llama.cpp branch or Unsloth Desktop, and Unsloth names no PC card setup; DwarfStar names CUDA support for V4.1 only on the DGX Spark.
Can I run DeepSeek V4.1 Flash on a Mac?
The September V4.1 Flash, yes, with DwarfStar: its Q2 file streams from SSD on one 128 GB Mac and holds all main weights on a 256 GB Mac, and its Q4 file runs fully resident on a 512 GB M3 Ultra. For 1017 only experimental paths exist: DwarfStar's new 1017 Q2 file (not yet in its guides), Unsloth's 233.0 GB 1-bit GGUF, which Unsloth says works on a 128 GB MacBook (unmerged llama.cpp branch or Unsloth Desktop), and mlx-community's oMLX builds from 313.3 GB with no hardware named.
Is there a DeepSeek V4.1 Flash GGUF?
For 1017, Unsloth publishes UD-IQ1_S (233.0 GB) to UD-Q8_K_XL (508.4 GB) for its llama.cpp pull request #30817 branch, and mrfakename publishes Q2_K (264.5 GB), Q3_K_M (347.3 GB), Q4_K_M (444.7 GB) and Q8_0 (508.0 GB) for the pull request #28696 branch, text only; neither branch is merged. DwarfStar's own Q2 GGUFs (365.7 GB, one of each checkpoint; the 1017 one is not yet in its guides) and its September Q4 run only in DwarfStar.
Is DeepSeek V4.1 Flash on Ollama?
Only as deepseek-v4.1-flash:cloud and deepseek-v4.1-flash:1017-cloud, which run on Ollama's servers, not on your hardware (Ollama library, checked October 20, 2026).
Can DeepInfer run DeepSeek V4.1 Flash on my GPU?
Not on a gaming card, a DGX Spark or a Mac. DeepInfer, the inference framework DeepSeek open-sourced, whose only supported model is V4.1 Flash, runs only on NVIDIA SM100 GPUs (in RigCheck's reading of NVIDIA's list, compute capability 10.0: B200 and GB200) and Huawei Ascend 950 NPUs, under Linux only; an RTX 4090 is 8.9, the RTX 50 series 12.0 and the DGX Spark 12.1. DeepSeek measured it on 4 SM100 GPUs for prefill and 64 for decode. For local hardware, use the routes in the checker above.
What changed in DeepSeek-V4.1-Flash-1017?
DeepSeek calls it a minor update with better reasoning, coding, agentic and visual results and shorter outputs; its config.json and file size are the same as the September V4.1 Flash. The API name (deepseek-flash) and price are unchanged.
Sources
- Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash-1017 and DeepSeek-V4.1-Flash (model cards, config, file sizes, license; read October 18, 2026), the reference inference README; DeepSeek API change log (October 17, 2026)
- Unsloth: DeepSeek-V4.1-Flash run locally guide (1017 memory table, 128 GB and 256 GB setups, llama.cpp pull request #30817) and its 1017 GGUF file list (read October 20, 2026)
- mrfakename/DeepSeek-V4.1-Flash-1017-GGUF (GGUF sizes, llama.cpp branch); llama.cpp pull request #28696 (open, DGX Spark report); llama.cpp llama-arch.cpp (checked October 18, 2026; master and pull request status October 20)
- DwarfStar (antirez/ds4): model guide, DGX Spark guide, two-machine setup, QA record (M3 Ultra speeds), download script (1017 Q2 default since October 19, 2026) and its V4.1 Flash GGUF (September and 1017 files, read October 20, 2026)
- mlx-community oQ2e, oQ2.5e, oQ3e and oQ4e
- DeepSeek: DeepInfer (supported models and hardware, prerequisites, benchmark setup, license; read October 21, 2026) and NVIDIA: CUDA GPU compute capability (read October 21, 2026)
- NVIDIA: DeepSeek-V4.1-Flash-NVFP4 (size, GB300 vLLM / SGLang recipes)
- Ollama library: deepseek-v4.1-flash tags and LM Studio: DeepSeek V4.1 Flash (Ollama checked October 20, LM Studio October 18, 2026)