Last updated: September 2026
Key Takeaways
- No. DeepSeek V4.1-Flash is a 512GB-class model: a 552B backbone plus 196B of Engram memory, 510GB on disk, experts already in FP4. Quantization cannot bring it to 128GB.
- No desktop runtime exists on day one. V4 took 66 days to reach mainline llama.cpp, and V4.1 is a new architecture with more new parts.
- DeepSeek V4-Flash 0731 keeps the 128GB crown, and its API version was retired today. The weights on your disk are the only copy nobody can reroute.
No, you cannot run DeepSeek V4.1-Flash on a 128GB machine, and no consumer runtime can load it at any memory size yet. The official checkpoint is 510GB across 48 files, and the one box that could hold it, the 512GB M5 Ultra Mac Studio, does not ship until late October. "Flash" describes what the model costs DeepSeek to serve, not what it costs you to own.
That is the verdict. Below: the arithmetic behind it, runtime status as of September 10, 2026, what the hosted version costs, and what to run on the hardware you already have.
What DeepSeek V4.1-Flash is, and why "smallest" misleads
DeepSeek V4.1-Flash is a 552B-parameter mixture-of-experts model with a separate 196B-parameter conditional memory called Engram, native image input, a 1M-token context window, and an MIT license. DeepSeek published the weights and a technical report on Hugging Face on September 10, 2026.
The architecture is new. Forty layers are split into a 20-layer causal encoder and a 20-layer decoder, which is how the model activates only 8B parameters per token while reading input and 16B while generating. Each expert layer holds one shared and 384 routed experts, six of which fire per token.
DeepSeek calls it the smallest model in its new family. That is true relative to a V4.1-Pro that does not exist yet. Against the model it replaces, V4-Flash at 284B total and 13B active, V4.1-Flash nearly doubled the total while cutting the active count. Active parameters set speed. Total parameters set the memory bill. Our guide to what AI models actually run at home calls this the costliest misconception in local AI; this release is its sharpest example.
The gains are real on DeepSeek's own numbers against V4-Flash 0731: DeepSWE v1.1 74.2 versus 54.4, Terminal-Bench 3.0 30.0 versus 7.6, AutomationBench 54.8 versus 37.7. That is why 128GB owners are asking. The answer does not change.
How much RAM to run DeepSeek V4.1-Flash locally
About 475 GiB (510GB) as shipped, and no published precision gets a single machine under roughly 385 GiB without going below the quality floor DeepSeek already set.
The payload
The official repository holds 48 safetensors shards totaling 510.30GB, or 475.25 GiB, reproducible from the file listing without downloading a byte. The formats are mixed by design: dense and attention weights in FP8, routed experts in FP4, the two Engram tables in FP8. DeepSeek's reference inference code converts experts with an explicit FP4 flag and runs across eight GPUs.
Why quantization cannot rescue it
V4-Flash became a 128GB model because Unsloth could repack its native MXFP4 experts bit-exact into GGUF: a lossless 162GB build, and a 103GB 3-bit build a 128GB machine holds. The same repack on V4.1-Flash yields a 500GB-class file, because DeepSeek spent the FP4 step before you downloaded anything.
Going below FP4 on the experts is the compromise our June ledger measured on V4-Flash: the 2-bit build agreed with the full model on roughly 78 percent of top tokens. That cost applies here too, which is why no 2-bit path appears below as an answer.
Backbone versus Engram
The split matters because software can place the two halves differently. A community repacker who has read the checkpoint reports Engram as two lookup tables of roughly 94.6 GiB each at FP8, about 189 GiB together, leaving roughly 286 GiB for the backbone. Transcoding those tables to FP4, already published by the same repacker, brings Engram to about 100 GiB and the whole model to roughly 385 GiB.
On a GPU server, SGLang's V4.1 preview documents an opt-in option that keeps one shared copy of the Engram tables in host memory instead of on the GPUs; that is how a four-accelerator box serves it. A Mac has no such split: unified memory is one pool, and the whole payload has to fit in it.
Memory tier table: where V4.1-Flash lands
Only the 512GB M5 Ultra tier holds the model in one box, and nothing can run it there yet.
| Your memory tier | V4-Flash 0731 today | V4.1-Flash, September 10 |
|---|---|---|
| 24GB GPU (RTX 3090 or 4090) | No; below the 110GB floor | No |
| 64GB unified mini PC | No | No |
| 96GB multi-GPU workstation | Yes, 3-bit with CPU offload | No |
| 128GB unified (M5 Max, Ryzen AI Max+ 395) | Yes, 103GB 3-bit on mainline llama.cpp | No; the payload is 3.7x the machine |
| 256GB M5 Ultra | Yes, lossless 162GB build | No at any published precision without experts below FP4 |
| 512GB M5 Ultra (ships late October) | Yes | Fits: about 385 GiB with FP4 Engram, 475 GiB as shipped. No runtime, no measured speed |
| Thunderbolt 5 Mac Studio cluster | Yes | Fits; no software supports the architecture |
| GPU server, 4 to 8 datacenter accelerators | Yes | Yes, via SGLang preview: about 286 GiB on accelerators plus about 190 GiB of host RAM for Engram |
Sizes are arithmetic from the published formats and shard sizes, checked September 10, 2026. No measured GGUF or MLX build of V4.1-Flash exists; when one does, this table gets replaced with its ladder.
The KV cache is the one number that got smaller
V4.1-Flash stores 890 bytes of global KV cache per token, a quarter of V4-Flash and 437 times less than DeepSeek V1, per DeepSeek's announcement.
| Model (release) | Global KV per token | Global KV for 1M tokens |
|---|---|---|
| DeepSeek V1 (Nov 2023) | 389,120 bytes | About 389GB |
| DeepSeek V3.2 (Dec 2025) | 48,068 bytes | About 48GB |
| DeepSeek V4-Flash (Apr 2026) | 3,514 bytes | About 3.5GB |
| DeepSeek V4.1-Flash (Sep 2026) | 890 bytes | Under 1GB |
Global KV only. The sliding-window attention state is extra; V4.1-Flash bounds it by replaying a recent window instead of persisting it.
For a home machine, that history says the context window stopped being the memory problem: a full 1M-token context costs under 1GB of global KV, where V3.2 needed 48GB. The weights are now the entire bill, which is why compressing the cache changed nothing in the tier table.
Runtime status on day one
As of September 10, no released engine serves the deepseek_v41 architecture, and every consumer route is empty.
| Route | Status, September 10, 2026 |
|---|---|
| llama.cpp mainline | No deepseek_v41 architecture; no issue or pull request located |
| GGUF files | None. One repository exists with a README and no weight files |
| Unsloth, LM Studio, MLX | Nothing published for V4.1; all three carry V4-Flash |
| Ollama | No V4.1 entry located; its V4-Flash tags are cloud-only, so check the suffix if one appears |
| vLLM | Open pull requests filed on release day; none merged |
| SGLang | Preview container only, not a regular release; maintainer-verified on 4x GB300 and 4x MI350X |
| DeepSeek reference code | PyTorch, eight-way tensor parallel, dense FP8 and expert FP4 kernels |
Hosted access is live on the DeepSeek API under the model name deepseek-flash.
The precedent is worth a date. DeepSeek V4 was released April 24, 2026; mainline llama.cpp merged support in release b9840 on June 29, 66 days later, with calibrated Unsloth builds in early July. V4.1 adds the encoder-decoder split, a new sparse-attention scheme, Engram lookups, revised residual mixing, and a new speculative decoder, each needing kernels before a GGUF is useful. The V4 clock is the only evidence anyone has.
One footnote. A from-scratch MLX engine already runs V4-Flash on a 48GB Mac by streaming routed experts from SSD per token, at roughly 5 tokens per second. Engram is a lookup table, the kind of tensor that idea suits. Nobody has applied it to V4.1, and that path is SSD-bound: a research direction, not a tier.
The honest tradeoff: the cheapest frontier API you cannot house
The hosted version of V4.1-Flash is cheap, good, and already cheaper than V4-Flash was in August.
DeepSeek's pricing page, checked September 10, lists deepseek-flash at $0.15 per million uncached input tokens and $0.60 per million output off-peak, doubling to $0.30 and $1.20 in peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC, weekdays). Cache-hit input is $0.003 off-peak. In August the same page listed V4-Flash at $0.22 and $0.66. The new model costs less to call than the old one.
A third-party datapoint says the quality is there. OpenDesign, a design-tool company, reported on September 9 that the pre-release test checkpoint scored 81.2 on its everyday design tasks against 82.7 for GPT-6 Astra, at $0.051 per artifact versus $3.66. One vendor, one task domain, a checkpoint that expired the next day: a signal, not a verdict.
Here is the other side of the same week. DeepSeek retired the V4-Flash and V4-Flash-Vision-Exp API models on September 10; the old names still resolve, but V4.1-Flash answers them. From 04:00 UTC on September 14, every deepseek-v4-pro request is routed to V4.1-Flash too, until a V4.1-Pro exists. The model behind a name changed twice in one week for anyone who did not pin it. The V4-Flash 0731 weights on a local disk did not change and cannot be rerouted. That is the argument for owning weights; for this model, ownership starts at 512GB. Our DeepSeek V4 status ledger tracks each change with its source.
What to run instead today
Run DeepSeek V4-Flash 0731 on a 128GB machine. It is still the strongest open model that fits in one box you can buy today.
The 3-bit Unsloth build is 103GB with a 110GB floor and loads in mainline llama.cpp and LM Studio; the lossless 8-bit build is 162GB for 192GB and 256GB machines. Our DeepSeek V4-Flash reality check has the tier table and speed sources. On the PC side, the 128GB entry ticket is a Ryzen AI Max+ 395 mini PC; confirm 128GB in the listing title, because the same chassis ships with 64GB.
Check Price on Amazon: GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB)
On the Apple side, the 128GB M5 Max Mac Studio ships September 22 and the 512GB M5 Ultra follows in late October; our M5 Ultra wait-or-buy guide covers which tier earns its price. If you are waiting on V4.1-Flash specifically, wait for a measured run on the exact configuration you plan to buy, not a capacity estimate, including this one.
Below 128GB, run what fits well rather than what barely loads. Our best local AI models by VRAM tier guide maps the field from 8GB up, and the mini PC guide for local AI covers a dedicated, network-isolated box.
Verification ledger, and what flips this verdict
Everything above was checked against primary sources on September 10, 2026.
- Model card and file listing, deepseek-ai/DeepSeek-V4.1-Flash: 552B backbone, 196B Engram, 8B/16B active, MIT, 48 shards, 510.30GB.
- Reference inference README: eight-way tensor parallel, dense FP8, experts FP4.
- Hugging Face quantization list: one GGUF repository with no weights; two NVFP4 repacks, neither loadable in a released engine.
- llama.cpp: no deepseek_v41 architecture in mainline; no issue or pull request located; V4 merged in b9840 on June 29.
- SGLang: preview container only. vLLM: open pull requests, none merged.
- DeepSeek API pricing page and change log: model name, rates, retirements, September 14 reroute.
- Apple: M5 Ultra tiers of 96GB, 256GB, and 512GB; availability September 22; 512GB in late October.
Three events change the verdict, and this page will be updated on each: a mainline llama.cpp merge for deepseek_v41; a documented GGUF or MLX build with a size ladder; a measured run on a named 512GB machine with a tokens-per-second figure.
Frequently Asked Questions
How much RAM do you need to run DeepSeek V4.1-Flash locally?
About 475 GiB (510GB) as shipped, or roughly 385 GiB with the Engram tables converted to FP4. On a Mac that means the 512GB M5 Ultra, which arrives in late October and has no runtime for this model yet. A GPU server can split it: roughly 286 GiB on accelerators and about 190 GiB of Engram in host RAM, using SGLang's preview build.
Is DeepSeek V4.1-Flash an 8B model?
No. Eight billion is the number of parameters that do work per token while the model reads input; 16 billion work while it writes. The full model is a 552B backbone plus 196B of Engram memory, and all of it has to be stored where the runtime can reach it. Active parameters describe speed, not size.
Can a 512GB Mac Studio run DeepSeek V4.1-Flash?
On capacity, yes: the model fits in 512 GiB with roughly 125 GiB of headroom once Engram is stored in FP4, and under 40 GiB as shipped, before the OS and runtime take their share. On software, no: as of September 10 no MLX or llama.cpp build understands the architecture, and nobody has published a decode speed. Wait for a measured run before buying that tier for this model.
Is DeepSeek V4-Flash 0731 obsolete now?
Not on your hardware. It remains the best open model that runs on a single 128GB machine with mainline software, and its weights cannot be retired. Its hosted API was: since September 10, requests to deepseek-v4-flash are answered by V4.1-Flash. The local copy is the only version that still answers as V4-Flash.
What happens to deepseek-v4-pro API calls on September 14?
From 04:00 UTC on September 14, 2026, DeepSeek routes every deepseek-v4-pro request to V4.1-Flash and bills it at Flash rates, until a V4.1-Pro is released. If your integration depends on V4-Pro behavior, test before that date. The name keeps working while the model behind it changes.
