DeepSeek-V4-Flash-0731: run the latest DeepSeek locally

If you want DeepSeek’s latest open-weight model, start with deepseek-ai/DeepSeek-V4-Flash-0731. It supersedes the V4 Flash preview, adds substantially stronger agentic behavior, supports low, high, and max reasoning effort, and bundles the DSpark speculative-decoding module. It is a 284B-target, 13B-active MoE that can fit across 2 high-memory systems, but it still needs careful runtime and memory planning. 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. For the complete 0731 hardware comparison and serving paths, see DeepSeek V4 Flash 0731 local hardware.

​

The DeepSeek lineup

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 13B-active V4 Flash still needs roughly 162–167 GB for its weights before cache and runtime overhead.

​

Which DeepSeek should you run?

• 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 the current V4 Flash model: Use deepseek-ai/DeepSeek-V4-Flash-0731. It is possible to self-host, but hardware capacity alone is not enough. Choose a path by runtime maturity in the local hardware guide.

​

DeepSeek-V4-Flash-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.

​

Why 0731 is the relevant release

DeepSeek’s own agentic evaluation shows a large step up from the preview. The model card reports 82.7 on Terminal Bench 2.1, 54.4 on DeepSWE, and 70.3 on Toolathlon-Verified. Those results used DeepSeek Harness in minimal mode, max reasoning effort, temperature = 1.0, and top_p = 0.95. Treat them as evidence for the exact evaluated setup, not guaranteed local throughput or quality. Use these identifiers deliberately:

​

How close is Flash 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

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.

​

Quantization sizes decide how many machines you need

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.

​

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 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. 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.

1. Pair the connector from the Setup tab in the Tokios dashboard.
2. Register a deployment in Models. For example, register deepseek-v4-flash and map it to the upstream model id your server exposes.
3. Create a scoped sk-tok-… API key in Keys.
4. Call https://api.tokios.com/v1 with your deployment name in model.

For a V4 Flash cluster, run the connector only on the head node, where it can reach the cluster’s OpenAI-compatible /v1 endpoint. Follow the local hardware guide for the serving examples, then register a deployment.

​

FAQ