Blog / Model guides
DeepSeek V4 Flash requirements: 2x H200, measured boot time and cost per hour
Updated 22 August 2026. DeepSeek released the official 0731 build of V4 Flash on 31 July 2026 under the MIT license (changelog, weights). It adds the DSpark speculative-decoding module, about 9 GB more on disk according to the Hugging Face repository, and DeepSeek reports that it beats V4-Pro-Preview on all nine of its published benchmarks. On 16 August 2026 the DeepSeek API moved to peak and off-peak pricing: for Flash, $0.44 per million input tokens and $1.32 per million output tokens at peak, $0.22 and $0.66 off-peak, against the previous flat $0.14 and $0.28 (per the changelog).
The measurements below were taken on 11 and 18 August 2026 on the catalog build at those dates. They are not claimed for the 0731 build, which we will re-measure and report here.
DeepSeek V4 Flash is a 284B-parameter mixture of experts model that activates 13B parameters per token, released by DeepSeek under the MIT license. It is one of the largest models in the LLM Hangar catalog. This guide covers what it takes to run it privately in your own cloud account: the hardware it needs, how long a deployment takes, what it costs, and how to call it once it is up.
Hardware requirements
| Configuration | GPUs | Notes |
|---|---|---|
| Recommended | 2x NVIDIA H200 (141 GB each) | The configuration we measure below |
| Alternative | 4x NVIDIA H100 (80 GB each) | Same silicon generation, ranked behind the H200 shape |
| Weights on disk | 159.6 GB in 46 shards (fp8 with fp4 expert weights, as measured; the 0731 build adds about 9 GB for its DSpark module) | |
| Context length | 32,768 tokens served; the model's native context is 1M via YaRN | |
The GPUs must be Hopper-class or newer. The fp8 release does not run on A100s.
What we measured
Numbers below are from live deployments on 2x H200 through RunPod in a European region (Iceland), launched with LLM Hangar into our own account. Nothing here is extrapolated from a spec sheet. The build measured was the catalog build on those dates; the numbers are not claimed for the 0731 build, which we will re-measure.
| Metric | Measured |
|---|---|
| Cold boot, fresh instance to ready endpoint (2026-08-11) | 21 minutes 26 seconds |
| Cold boot, repeat run (2026-08-18) | about 21 minutes |
| Downloaded during boot | 159.6 GB of weights |
Throughput and time-to-first-token figures are only worth publishing together with the workload they were measured under. We will add ours after the next benchmark run: 4,096 input tokens, 1,024 output tokens, concurrency 1.
What it costs
Your cloud provider bills you directly for the GPUs. LLM Hangar is a flat subscription on top, so the infrastructure cost is just the GPU rate:
| Item | Observed |
|---|---|
| 2x H200 on RunPod, per hour | $7 to $14, rate fixed at pod creation |
| Around the clock, per month | roughly $5,100 to $10,200 |
| Business hours only (about 176 h per month) via a wake/sleep schedule | roughly $1,230 to $2,460 |
Those are the rates we observed at our own boots in August 2026, not a quotation. The dashboard shows your exact rate before you confirm anything.
Hard budget caps and auto-destroy are built in, so a forgotten deployment cannot run up an open-ended bill. See budget caps and auto-destroy.
Self-host or API after the August 16 price change
The API price change moves the break-even. Does the API cost of the tokens your instance serves in an hour exceed the hourly cost of the instance? Using only the 2x H200 rate we observed on RunPod ($7 to $14 per hour, midpoint $10.50) and the new API prices:
| Workload | Blended API price per 1M tokens | Break-even at $10.50/hr |
|---|---|---|
| 80% input, 20% output, peak | 0.8 x $0.44 + 0.2 x $1.32 = $0.616 | 17.0M tokens per hour, about 4,700 tokens/s sustained |
| 80% input, 20% output, off-peak | 0.8 x $0.22 + 0.2 x $0.66 = $0.308 | 34.1M tokens per hour, about 9,500 tokens/s sustained |
| Output only, peak | $1.32 | 8.0M tokens per hour, about 2,200 tokens/s sustained |
At the low end of the observed range ($7 per hour) the peak 80/20 break-even is about 3,200 tokens/s; at the high end ($14) it is about 6,300 tokens/s. The formula is: break-even tokens per hour = hourly instance cost divided by the blended API price per token. The unknown is the sustained throughput of 2x H200 serving this model at your concurrency. We have not published ours yet (see above), so this section gives you the arithmetic, not a verdict. Below the break-even, the API is cheaper per token and self-hosting is a privacy or control decision; above it, or with a wake/sleep schedule that only pays for the hours you use, self-hosting also wins on cost. The self-hosting cost calculator does this for any rate and throughput.
Current prices, live
Deploying it
- Connect your cloud account (AWS, Nebius, or RunPod).
- In the dashboard, choose DeepSeek V4 Flash from the catalog.
- Pick the 2x H200 shape and a region. EU-only residency is available.
- Review the cost estimate, set a budget cap, and confirm.
- Watch the deployment progress live. When it is ready you get a private endpoint and an API key.
Calling the endpoint
The endpoint is OpenAI-compatible, so existing clients work by changing the base URL:
from openai import OpenAI
client = OpenAI(
base_url="https://abc12345.gw.llmhangar.com/v1",
api_key="YOUR_API_KEY",
)
resp = client.chat.completions.create(
model="deepseek-v4-flash",
messages=[{"role": "user", "content": "Hello"}],
stream=True,
)More clients (curl, JavaScript, LangChain, Vercel AI SDK, editor configs) are in using your endpoint.
Why run it privately
Prompts and responses stay on an instance in your own account. The infrastructure belongs to you, your provider bills you for it directly, and when you delete the deployment we verify nothing is left running.
Frequently asked
Can DeepSeek V4 Flash run on a single H200?
Not in a configuration we qualify. The fp8 release is 159.6 GB on disk before any KV cache, so a single 141 GB H200 cannot hold it. The recommended shape is 2x H200; 4x H100 80 GB is the alternative. The GPUs must be Hopper-class or newer, because the fp8 release does not run on A100s.
How long does it take to boot?
21 minutes 26 seconds from a fresh instance to a ready endpoint on 2x H200 via RunPod, measured on 11 August 2026, and about 21 minutes on a repeat run a week later. 159.6 GB of weights download during the boot. Where the provider supports a prepared stage, later boots of the same model are much faster.
What does it cost per hour on RunPod or AWS?
On RunPod we observed $7 to $14 per hour for 2x H200, fixed at pod creation, which is roughly $5,100 to $10,200 a month around the clock. On AWS the only H200 shape is p5en.48xlarge with eight H200s, at a $63.30 per hour list price as of July 2026, so AWS only makes sense if you can use all eight. The live table above shows current rates.
Why do different sites give different VRAM numbers?
Because they assume different precisions, context lengths and concurrency. A bf16 copy of a 284B model is twice the size of the fp8 release; a 1M-token context needs far more KV cache than 32K; and serving many concurrent requests needs more memory than serving one. Our numbers are for the fp8 weights as published and 32,768 tokens of context served, and our throughput figures will be reported at concurrency 1 with 4,096 input and 1,024 output tokens.