DeepSeek V4.1-Flash: run the latest DeepSeek locally

DeepSeek V4.1-Flash is the current DeepSeek flagship open-weight release. It supersedes V4-Flash-0731 with native multimodal vision, an asymmetric 552B MoE architecture (8B active for input, 16B active for output), 1M-token context with up to 384K output tokens, and a dramatically compressed KV cache — 890 bytes per token globally, roughly 3.9× smaller than V4 Flash. DeepSeek released it September 8–10, 2026 (limited beta September 8, production September 10) under the MIT license. The October 17, 2026 checkpoint (V4.1-Flash-1017) delivers additional benchmark improvements and confirmed vLLM, SGLang, and Unsloth GGUF support. Use an R1 distill instead when you need DeepSeek-style reasoning on a laptop or 1 consumer GPU. The full R1 and V3 models remain 671B-class, multi-GPU options.

Last verified: 2026-10-20

The DeepSeek lineup

Model	What it is	Parameters	Context	Practical local hardware
DeepSeek-V4.1-Flash	Current flagship: native vision, asymmetric MoE, 3.9× smaller KV cache	552B total / 8B input active, 16B output active	1M (384K output)	2 high-memory systems; see hardware requirements below
DeepSeek-V4-Flash-0731	Superseded V4 Flash release with DSpark head	284B target / 13B active	1M	2 high-memory systems; see hardware guide
R1 distills	Qwen- and Llama-based models trained on R1 reasoning traces	1.5B–70B dense	128K	Laptop or 1 GPU with 24–48 GB of memory
DeepSeek-R1 / R1-0528	The full reasoning MoE	671B / 37B active	160K	Multi-GPU server; about 404 GB of `Q4_K_M` weights
DeepSeek-V3 / V3.1 / V3.2	The base-model line behind R1	671B-class MoE	160K	Multi-GPU server

The active-parameter count affects compute per generated token. It does not reduce the memory required to keep all experts resident. That is why a 552B V4.1-Flash still needs all weights resident even though only 8B or 16B activate per token.

Which DeepSeek should you run?

You want the current DeepSeek flagship: Use deepseek-ai/DeepSeek-V4.1-Flash. It adds native vision, asymmetric 8B/16B active parameters, and a 3.9× smaller KV cache over V4-Flash-0731. It is possible to self-host, but hardware capacity alone is not enough — review the requirements below.
You have a laptop or 1 consumer GPU: Start with an R1 distill. deepseek-r1:14b is a practical Ollama starting point on a 24 GB GPU. See DeepSeek-R1.
You need the original R1 reasoning model: Plan for a real multi-GPU server. The full 671b model is not a single-workstation pull.
You need a general DeepSeek base model: The V3 line remains 671B-class and multi-GPU. Use V3.1 tags in Ollama where they fit your deployment plan; verify current tags before standardizing one.
You need V4-Flash-0731 specifically: It is now superseded by V4.1-Flash. The DeepSeek API redirects old deepseek-v4-flash routes to V4.1-Flash. If you must stay on 0731, see the superseded section below.

DeepSeek V4.1-Flash

DeepSeek released V4.1-Flash on September 8–10, 2026 (limited beta September 8, production September 10). The official model card identifies it as the current flagship release that supersedes V4-Flash-0731. It describes a 552B-parameter mixture-of-experts model with a novel Causal-Encoder-Decoder architecture, asymmetric activation (8B active for input tokens, 16B active for output tokens), native multimodal vision (text + image input), 1M-token context with up to 384K output tokens, and MIT licensing.

October 17, 2026 checkpoint update. DeepSeek published an updated V4.1-Flash checkpoint (V4.1-Flash-1017) with improved benchmarks. vLLM, SGLang, and Unsloth GGUF builds are confirmed and available. Use the latest checkpoint when you download or pull the model.

What’s new vs V4-Flash-0731

V4.1-Flash delivers 6 major improvements over the July 31 release:

Native vision capability — the first DeepSeek Flash model with built-in multimodal image understanding, not just text
Asymmetric architecture — 8B active parameters for input processing, 16B for output generation, optimizing the compute/quality tradeoff
Dramatically compressed KV cache — 890 bytes global KV per token, roughly 3.9× smaller than V4 Flash
Reduced hardware footprint — 1/4 the HBM, 1/8 the SSD vs prior generation
Stronger benchmarks — beats V4 Flash 0731 on all 8 selected tests
Lower operating cost — approximately 75.5% cheaper to run

NVIDIA NVFP4 variant

NVIDIA published an optimized NVFP4 quantization at nvidia/DeepSeek-V4.1-Flash-NVFP4. This variant targets NVIDIA Blackwell and Hopper GPUs with FP4 precision, reducing memory requirements while preserving quality.

Serving support

vLLM and SGLang both support V4.1-Flash. Unsloth publishes GGUF builds at unsloth/DeepSeek-V4.1-Flash-GGUF, which enables llama.cpp, Ollama, LM Studio, and other GGUF-compatible backends. Check each runtime’s current release notes for the exact flags and tensor-parallel settings.

Identifiers for V4.1-Flash

Use these identifiers deliberately:

Identifier	Use it for
`deepseek-ai/DeepSeek-V4.1-Flash`	The upstream model id when you download or serve the official checkpoint
`nvidia/DeepSeek-V4.1-Flash-NVFP4`	The NVIDIA NVFP4 variant for Blackwell/Hopper GPUs
`deepseek-flash`	The official DeepSeek API model id (replaces `deepseek-v4-flash`)
`deepseek-flash`	A clear example Tokios deployment name; you can choose another public deployment name

The DeepSeek API now routes old deepseek-v4-flash requests to V4.1-Flash. If you registered a Tokios deployment named deepseek-v4-flash, update it to deepseek-flash to match the new API id.

A Tokios deployment is your public name. It is separate from the upstream model id. If you register deepseek-flash, clients send that in the model field while your runtime serves deepseek-ai/DeepSeek-V4.1-Flash.

Capabilities

Native vision. Built-in multimodal image understanding — text and image input, not just text.
Asymmetric compute. 8B active parameters for input processing, 16B for output generation, optimizing the compute/quality tradeoff.
Long context. 1M-token context window with up to 384K output tokens.
Coding and agentic work. Beats V4-Flash-0731 on all 8 selected benchmarks, including coding and agentic tasks.
Multilingual. Supports multiple languages for both input and output.
Compressed KV cache. 890 bytes per token globally, roughly 3.9× smaller than V4 Flash — reduces per-token memory overhead at long contexts.

Hardware fit

V4.1-Flash has a 552B-parameter backbone plus a ~196B Engram conditional memory module, for a total checkpoint of approximately 763B parameters. All parameters must stay resident — the asymmetric activation (8B/16B) reduces compute per token, not memory. The estimates below use community-verified checkpoint sizes. Check the model card and Unsloth GGUF builds for current sizes.

Precision	Estimated size	Hardware
BF16	~1.5 TB	8× H200 or equivalent
FP8	~511 GB	4× DGX Spark (TP=4) or 6× H200
NVFP4 + MXFP4 Engram	~400 GiB	4× DGX Spark (TP=4) or 2× Mac Studio M5 Ultra 512 GB
GGUF Q4_K_XL (Unsloth)	~501 GB	4× DGX Spark or 2× Mac Studio M5 Ultra 512 GB
GGUF IQ2_XXS (Unsloth)	~301 GB	3× DGX Spark or 1× Mac Studio M5 Ultra 512 GB
Mixed FP8 dense + MXFP4 experts	~160–210 GB	2× DGX Spark (TP=2) — community-tested

The NVIDIA NVFP4 variant targets Blackwell and Hopper GPUs. The mixed-precision path on 2 DGX Sparks (FP8 dense layers + MXFP4 experts, ~79–105 GiB per node) is documented with community benchmarks in the V4.1 Flash on 2 DGX Sparks guide.

Run it

vLLM (GPU serving)
SGLang
Ollama (GGUF)

vllm serve "deepseek-ai/DeepSeek-V4.1-Flash"

Check the model card for current flags and tensor-parallel settings for your hardware.

python -m sglang.launch_server --model-path "deepseek-ai/DeepSeek-V4.1-Flash"

Check the model card for current flags and tensor-parallel settings.

ollama run deepseek-v4.1-flash

Check the Ollama library for current tags. Unsloth GGUF builds are also available at unsloth/DeepSeek-V4.1-Flash-GGUF if you want to manually import a specific quantization.

Ollama and LM Studio now support V4.1-Flash through Unsloth GGUF builds. Check the Ollama library and LM Studio for current availability.

DeepSeek-V4-Flash-0731 (superseded)

V4-Flash-0731 is the previous DeepSeek Flash release, superseded by V4.1-Flash in September 2026. Use V4.1-Flash for new deployments unless you have a specific reason to stay on 0731.

DeepSeek released DeepSeek-V4-Flash-0731 on July 31, 2026. The official model card identifies it as the official V4 Flash release that supersedes the preview. It describes a 284B-target-model, 13B-active mixture-of-experts model with a bundled DSpark head, 1M-token context, low, high, and max reasoning effort, and tool calling. SGLang reports about 304B parameters when it counts the attached draft head. The weights are MIT-licensed.

V4-Flash-0731 is now superseded by V4.1-Flash. The DeepSeek API redirects old deepseek-v4-flash routes to V4.1-Flash. Use V4.1-Flash for new deployments unless you have a specific reason to stay on 0731.

The 0731 repository does not ship a Jinja chat template. It includes a dedicated encoding package for turning OpenAI-compatible messages into model input and parsing completions. Use a runtime recipe that supports this encoding, then test multi-turn chat, reasoning content, and tool calls before exposing the deployment.

Use these identifiers deliberately:

Identifier	Use it for
`deepseek-ai/DeepSeek-V4-Flash-0731`	The upstream model id when you download or serve the 0731 checkpoint
`deepseek-v4-flash`	The old DeepSeek API model id (now redirects to `deepseek-flash`)

A Tokios deployment is your public name. It is separate from the upstream model id. If you register deepseek-v4-flash, clients send that in the model field while your runtime serves deepseek-ai/DeepSeek-V4-Flash-0731.

How close is V4-Flash-0731 to GPT-5.6?

Treat comparisons as configuration-specific, not a general parity claim. Artificial Analysis reports Flash Max at 50 and GPT-5.6 Sol Low at 49 in its named comparison. That result compares Flash Max with Sol at its Low effort setting. It does not establish parity with GPT-5.6 Sol at higher effort, or across every workload.

Local paths for V4-Flash-0731

Configuration	Capacity conclusion	Maturity	What to know
1× DGX Spark or Strix Halo	128 GB holds the 103 GB 3-bit build with cache headroom, but not the 155–162 GB builds	Community-tested	Unsloth’s guide targets `UD-IQ3_XXS` for exactly this case. Cheapest way to run Flash at all; expect quality loss against near-lossless builds.
2× RTX PRO 6000 Blackwell	192 GB VRAM can hold 162–167 GB of weights with limited initial cache headroom	Community validation required	SGLang documents TP=2 for the earlier Flash checkpoint, but does not list 0731 as officially verified at TP=2. Start conservatively and validate your exact build.
2× DGX Spark	256 GB aggregate unified memory gives more room for weights and cache	Community-tested	NVIDIA documents clustering over ConnectX-7. Dual-Spark DeepSeek recipes target the preview checkpoint; rerun validation for 0731.
2× Strix Halo	256 GB aggregate unified memory can hold the 162 GB GGUF	Experimental	The distributed route uses llama.cpp RPC, which its maintainers describe as proof-of-concept, fragile, and insecure. Do not use it as a production serving design.
4× GB300	Official reference configuration	Verified	The 0731 model card’s reference configuration.

Every path needs disk space for the checkpoint, room for temporary downloads and runtime allocations, and measured KV-cache headroom. Start at 32K–128K context. Treat 384K and 1M as separate tuning targets, not default settings.

If you already run V4-Flash-0731, keep the serving endpoint on the head node’s loopback interface. Run 1 Tokios connector on that head node and point it only at the runtime’s /v1 endpoint. The hardware guide shows the topology and deployment mapping.

Quantization sizes for V4-Flash-0731

Unsloth publishes GGUF builds at unsloth/DeepSeek-V4-Flash-0731-GGUF. The sizes matter more than usual, because they are what decides whether one box is enough.

Build	File size	Memory to run it
`UD-Q8_K_XL`, lossless	162 GB	~169 GB
`UD-Q4_K_XL`, near lossless	155 GB	~162 GB
`UD-IQ3_XXS`, 3-bit	103 GB	110 to 135 GB
2-bit	102 GB	Fits the same class of machine, with more quality loss
1-bit	92 GB	Smallest published build; expect real degradation

Can a DGX Spark run DeepSeek-V4-Flash-0731?

Yes at 3-bit, on a single unit. Unsloth’s guide picks UD-IQ3_XXS precisely because it fits a 128 GB machine, and reports on NVIDIA’s developer forum run it on one Spark through llama-server, including at 262K context with smaller 2-bit builds. What one Spark cannot hold is either near-lossless build: UD-Q4_K_XL at 155 GB and UD-Q8_K_XL at 162 GB both exceed 128 GB. Those want two Sparks and the pooled memory in the table above.

Can Strix Halo run DeepSeek-V4-Flash-0731?

Yes, with the same 3-bit condition. A 128 GB Strix Halo holds the 103 GB build; anything near lossless needs a second machine, and that distributed route runs over llama.cpp RPC, which its own maintainers describe as proof-of-concept. Throughput here is bounded by memory bandwidth rather than capacity. The 13B active-parameter count works in your favour: per-token compute stays small even though all 284B parameters must remain resident.

Can a 24 GB GPU run DeepSeek-V4-Flash-0731?

No. The smallest published build is 92 GB, so no single consumer card is in range at any quantization, and offloading that much weight to system RAM surrenders the speed that makes Flash worth running. Use a DeepSeek-R1 distill instead — those are sized for one GPU.

The R1 and V3 family

V4.1-Flash does not replace every earlier DeepSeek option. The R1 distills remain the practical choice for small hardware. The full R1 and V3 line remain useful when you already operate a 671B-class serving cluster.

Model	Availability and local route	License
R1 distills	Ollama, LM Studio, vLLM, or SGLang; choose a size that fits 1 machine	MIT for Qwen-based distills; Llama terms for the Llama-based tags
Full DeepSeek-R1 / R1-0528	Multi-GPU server	MIT
DeepSeek-V3 / V3.1	Multi-GPU server; V3.1 has Ollama tags	MIT
DeepSeek-V3.2	Official Hugging Face checkpoint; multi-GPU server	MIT

See DeepSeek-R1 for distill sizes, context, and setup advice. See what size model fits your GPU? for the memory calculation behind the larger models.

Connect a DeepSeek deployment with Tokios

Once your chosen runtime is healthy, install the connector on the same machine as the runtime endpoint. The connector dials out to Tokios; it does not expose an inbound port. For a V4 cluster, run the connector only on the head node, where it can reach the cluster’s OpenAI-compatible /v1 endpoint.

Keep your local runtime running

Start DeepSeek V4.1-Flash in vLLM or SGLang and leave it bound to loopback. For V4-Flash-0731, use the serving recipe from the hardware guide.

Pair the Tokios connector

Sign in at tokios.com/console, open the Setup tab, and click Start pairing to get a one-time claim code (TKS-XXXX-XXXX). Run the installer on the machine that runs the model:

Windows

irm https://tokios.com/downloads/install/windows | iex; tokios-install -Code 'TKS-XXXX-XXXX'

macOS

curl -fsSL https://tokios.com/downloads/install/macos | sh -s -- --code TKS-XXXX-XXXX

Linux

curl -fsSL https://tokios.com/downloads/install/linux | sh -s -- --code TKS-XXXX-XXXX

Then approve the device on the Setup or Connectors tab.

In the Models tab, register the upstream model id under a public deployment name. For V4.1-Flash, register deepseek-flash and map it to deepseek-ai/DeepSeek-V4.1-Flash. For V4-Flash-0731, register deepseek-v4-flash and map it to deepseek-ai/DeepSeek-V4-Flash-0731. Clients send the deployment name in the model field, never the local id.

Create an API key

In the Keys tab, click Create key. Copy the sk-tok-… key now — it is shown only once.

Call it from anywhere

curl https://api.tokios.com/v1/chat/completions \
  -H "Authorization: Bearer sk-tok-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-flash",
    "stream": true,
    "messages": [
      {"role": "system", "content": "You are a helpful coding assistant."},
      {"role": "user", "content": "Audit this repository and propose a phased refactor."}
    ]
  }'

export ANTHROPIC_BASE_URL=https://api.tokios.com
export ANTHROPIC_API_KEY=sk-tok-YOUR_KEY
claude --model deepseek-flash

Caveats

Benchmarks are vendor-reported. DeepSeek published the V4.1-Flash comparisons on release day. Wait for independent evaluations before treating the scores as settled.
Hardware requirements are estimates. Specific quantization sizes for V4.1-Flash are not yet published. The hardware fit table uses the standard calculation (parameters × bytes per weight). Check the model card for current GGUF builds before committing hardware.
V4.1-Flash runtime support is confirmed. vLLM and SGLang both support V4.1-Flash. Unsloth publishes GGUF builds, enabling Ollama, LM Studio, and llama.cpp backends.
All 552B parameters stay resident. The asymmetric architecture (8B/16B active) reduces compute per token, not memory. Every expert weight must remain loaded regardless of the active count.
MIT license is permissive. The weights allow commercial use with no acceptable-use policy or user-count threshold, but read the license yourself before building on it.

FAQ

How much VRAM does DeepSeek V4.1-Flash need?

V4.1-Flash is a 552B-parameter model. Specific quantization sizes are not yet published at the time of writing. The asymmetric architecture (8B/16B active) and 3.9× smaller KV cache (890 bytes per token) suggest it may fit similar hardware to V4-Flash-0731 (110–169 GB depending on build), but check the model card for current GGUF builds and memory requirements. The NVIDIA NVFP4 variant targets Blackwell/Hopper GPUs with FP4 precision for reduced memory.

How much VRAM does DeepSeek-V4-Flash-0731 need?

Between roughly 110 GB and 169 GB, depending on which build you pull. The 3-bit UD-IQ3_XXS is 103 GB on disk and wants 110 to 135 GB of memory, which is why it fits a single 128 GB unified-memory machine. The near-lossless UD-Q4_K_XL is 155 GB and the lossless UD-Q8_K_XL is 162 GB, both of which need more than one such machine. All 284B parameters stay resident regardless of the 13B active count.

Is DeepSeek V4.1-Flash free to use?

The weights are MIT-licensed, which is unusually permissive for a model this capable and allows commercial use. Free to use is not the same as free to run: the 552B parameter count means significant hardware requirements. DeepSeek also operates a paid hosted API if you would rather not self-host.

Is DeepSeek-V4-Flash-0731 free to use?

The weights are MIT-licensed, which is unusually permissive for a model this capable and allows commercial use. Free to use is not the same as free to run: at 92 GB for even the smallest build, the hardware is the real cost. DeepSeek also operates a paid hosted API if you would rather not self-host.

What is the difference between V4.1-Flash and V4-Flash-0731?

V4.1-Flash supersedes V4-Flash-0731 with 6 major improvements: native vision capability (text + image input), asymmetric architecture (8B active for input, 16B for output vs 13B active in 0731), 3.9× smaller KV cache (890 bytes per token), 1/4 HBM footprint, 1/8 SSD footprint, and approximately 75.5% lower operating cost. V4.1-Flash beats 0731 on all 8 selected benchmarks. The DeepSeek API now redirects old deepseek-v4-flash routes to V4.1-Flash.

Can I run DeepSeek V4.1-Flash with Ollama?

Yes. Unsloth publishes GGUF builds at unsloth/DeepSeek-V4.1-Flash-GGUF, which enables Ollama and other GGUF-compatible backends. Check the Ollama library for current tags. vLLM and SGLang also support V4.1-Flash natively. For V4-Flash-0731, the R1 distills and DeepSeek-V3.1 do have Ollama tags — see DeepSeek-R1.

Can I run DeepSeek-V4-Flash-0731 with Ollama?

Not as the documented path. Unsloth’s guide drives the 0731 checkpoint through llama.cpp using the split GGUF files, and mainline llama.cpp carries the V4 architecture support. Check Ollama’s library for current availability rather than assuming a local tag exists. The R1 distills and DeepSeek-V3.1 do have Ollama tags — see DeepSeek-R1.

What is the difference between V4 Flash and an R1 distill?

Scale and purpose. V4.1-Flash is the current 552B mixture-of-experts release with native vision, 1M-token context, and asymmetric 8B/16B active parameters, aimed at multimodal and agentic work on multi-GPU systems. V4-Flash-0731 is the superseded 284B release. The R1 distills are small dense models that fit a single consumer GPU and trade capability for reach. If your hardware is one 24 GB card, the distill is not a compromise — it is the only one of the two that runs.

How do I reach a DeepSeek deployment from another machine?

Serve it locally, then pair the Tokios connector on the machine holding the runtime — for a cluster, on the head node only. The connector dials outward, so the box with the GPUs never needs an inbound port. Clients then call https://api.tokios.com/v1 with an sk-tok-… key and your deployment name in the model field.

V4 Flash local hardware

Compare 2 RTX PRO 6000s, 2 DGX Sparks, and 2 Strix Halo systems for V4-Flash-0731.

DeepInfer (datacenter)

DeepSeek’s open-source inference engine for V4.1-Flash on SM100 datacenter GPUs.

DeepSeek-R1

Choose and run an R1 reasoning distill on smaller hardware.

Choose a model by task

Compare local models for reasoning, coding, and agent work.

Match a model to your GPU

Understand weight, cache, and quantization memory requirements.

Get Started

Models

Concepts

Reference

DeepSeek V4.1-Flash: run the latest DeepSeek locally

The DeepSeek lineup

Which DeepSeek should you run?

DeepSeek V4.1-Flash

What’s new vs V4-Flash-0731

NVIDIA NVFP4 variant

Serving support

Identifiers for V4.1-Flash

Capabilities

Hardware fit

Run it

DeepSeek-V4-Flash-0731 (superseded)

How close is V4-Flash-0731 to GPT-5.6?

Local paths for V4-Flash-0731

Quantization sizes for V4-Flash-0731

Can a DGX Spark run DeepSeek-V4-Flash-0731?

Can Strix Halo run DeepSeek-V4-Flash-0731?

Can a 24 GB GPU run DeepSeek-V4-Flash-0731?

The R1 and V3 family

Connect a DeepSeek deployment with Tokios

Call it from anywhere

Caveats

FAQ

V4 Flash local hardware

DeepInfer (datacenter)

DeepSeek-R1

Choose a model by task

Match a model to your GPU

​The DeepSeek lineup

​Which DeepSeek should you run?

​DeepSeek V4.1-Flash

​What’s new vs V4-Flash-0731

​NVIDIA NVFP4 variant

​Serving support

​Identifiers for V4.1-Flash

​Capabilities

​Hardware fit

​Run it

​DeepSeek-V4-Flash-0731 (superseded)

​How close is V4-Flash-0731 to GPT-5.6?

​Local paths for V4-Flash-0731

​Quantization sizes for V4-Flash-0731

​Can a DGX Spark run DeepSeek-V4-Flash-0731?

​Can Strix Halo run DeepSeek-V4-Flash-0731?

​Can a 24 GB GPU run DeepSeek-V4-Flash-0731?

​The R1 and V3 family

​Connect a DeepSeek deployment with Tokios

​Call it from anywhere

​Caveats

​FAQ

V4 Flash local hardware

DeepInfer (datacenter)

DeepSeek-R1

Choose a model by task

Match a model to your GPU

The DeepSeek lineup

Which DeepSeek should you run?

DeepSeek V4.1-Flash

What’s new vs V4-Flash-0731

NVIDIA NVFP4 variant

Serving support

Identifiers for V4.1-Flash

Capabilities

Hardware fit

Run it

DeepSeek-V4-Flash-0731 (superseded)

How close is V4-Flash-0731 to GPT-5.6?

Local paths for V4-Flash-0731

Quantization sizes for V4-Flash-0731

Can a DGX Spark run DeepSeek-V4-Flash-0731?

Can Strix Halo run DeepSeek-V4-Flash-0731?

Can a 24 GB GPU run DeepSeek-V4-Flash-0731?

The R1 and V3 family

Connect a DeepSeek deployment with Tokios

Call it from anywhere

Caveats

FAQ