DeepSeek-V4-Flash: specs, benchmarks, and how to run it

AI This article was created with the help of AI.

DeepSeek-V4-Flash specs and MoE architecture

DeepSeek-V4-Flash is a 284-billion parameter sparse Mixture-of-Experts (MoE) model engineered for high-throughput reasoning and long-context processing. Built by DeepSeek-AI, which publishes the V4 checkpoints on Hugging Face, the architecture activates only 13 billion parameters per forward pass, delivering the execution speed of a mid-sized dense network alongside the parameter capacity of a frontier foundation model. By combining sparse routing with a compact activation path, the model addresses the primary bottleneck in enterprise inference: the massive VRAM footprint and compute overhead typically required to serve high-capability models.

Precision mechanics and architectural optimizations

The model weights implement a hybrid precision structure designed to balance numerical stability with memory efficiency. MoE expert parameters are quantized natively to FP4 precision, while the remaining attention, norm, and router parameters stay in FP8. This mixed FP4 and FP8 precision layout puts the native build at roughly 146 GB on disk, compared with about 284 GB for the FP8-only quantized release. During forward execution, the network uses Manifold-Constrained Hyper-Connections (mHC), a refinement of conventional residual connections that constrains the residual mapping onto a specific manifold to enhance the stability of signal propagation across layers while preserving model expressivity.

Both V4 models were pre-trained on more than 32T diverse, high-quality tokens using the Muon optimizer, which DeepSeek credits with faster convergence and greater training stability. From an infrastructure standpoint, the native FP4 expert build requires NVIDIA B200 GPUs for the FP4 expert weights, for example an HGX B200 node. On Hopper-generation hardware there is no FP4 path, so teams run the FP8-only quantized release with tensor parallelism across all eight H100 GPUs in the HGX system, which avoids failed FP4 kernel launches and out-of-memory errors during model load.

Performance benchmarks and agentic capabilities

DeepSeek published the official V4-Flash release on 31 July 2026, a re-post-trained version of the April preview with the architecture and parameter count unchanged and the gains coming from a new post-training pipeline focused on coding, agents, reasoning, and tool use. The chat template exposes three reasoning effort modes: Non-think for fast, intuitive responses, Think High for explicit chain-of-thought analysis and planning, and Think Max for maximum reasoning effort. In Think Max mode, the model allocates extended test-time compute to generate internal reasoning traces before returning its final output, significantly lifting accuracy across software engineering and code generation benchmarks.

Coding, mathematical reasoning, and terminal execution

Empirical evaluations show that DeepSeek-V4-Flash comes close to significantly larger foundation models on rigorous coding assessments. On LiveCodeBench, a benchmark designed to mitigate test-set contamination through fresh competitive programming problems, DeepSeek-V4-Flash achieves a Pass@1 score of 91.6 in Think Max mode. On the Terminal Bench 2.1 agentic evaluation, which measures an agent's ability to work in a live shell environment, the official 0731 release scores 82.7, against 61.8 for the V4-Flash preview and 72.1 for V4-Pro (Preview), on vendor-reported numbers.

In the same Think Max mode, the model records 94.8 Pass@1 on the February 2026 HMMT mathematics competition set and 86.2 EM on MMLU-Pro. For autonomous software engineering pipelines, these metrics translate into reliable tool calling and structured output parsing. The model is designed to be driven through an OpenAI-compatible message format, with dedicated encoding and parsing scripts for its reasoning-mode output, which lets long-horizon agents run multi-file refactoring tasks without breaking argument signatures or command syntax.

Processing the 1-million token context window

Handling million-token context windows in production typically introduces severe GPU memory constraints. In standard multi-head attention (MHA) or grouped-query attention (GQA), key-value (KV) cache memory scales linearly with sequence length. Agentic workloads make this worse, because they carry system instructions, tool outputs, retrieved context, code, logs, memory, and multi-step reasoning traces across a workflow, and as context windows grow, attention and KV cache become major bottlenecks.

Hybrid attention mechanics: CSA and HCA

To bypass the KV cache bottleneck, the DeepSeek-V4 Flash architecture employs a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to dramatically improve long-context efficiency. CSA compresses KV entries through dynamic sequence compression and then sparsifies the attention matrices, while HCA applies much more aggressive compression by consolidating KV entries across sets of tokens into a single compressed entry.

The efficiency headline in the DeepSeek-V4 technical report is stated for the Pro variant: at a one-million-token context, DeepSeek-V4-Pro requires only 27 percent of the single-token inference FLOPs and 10 percent of the KV cache of DeepSeek-V3.2. Flash shares the same hybrid attention stack, so it inherits the same class of savings. For enterprise engineering teams managing long-context inference infrastructure, that efficiency unlocks practical use cases: ingesting entire code repositories, indexing quarterly financial reports, and parsing hundreds of thousands of lines of terminal execution logs within a single inference call without sharding state across multiple distributed databases.

Contrasting DeepSeek-V4-Flash with DeepSeek-V4-Pro

Selecting between DeepSeek-V4-Flash and DeepSeek-V4-Pro requires evaluating the trade-off between parametric knowledge depth and operational throughput. While both models share the V4 architecture stack and support a native 1-million token context window, their scale and active parameter footprints differ substantially. DeepSeek-V4-Pro contains 1.6 trillion total parameters with 49 billion active parameters per forward pass, making it the flagship model for frontier reasoning. In contrast, Flash operates with 284 billion total parameters and 13 billion active parameters.

Reasoning parity versus raw knowledge retention

In structured evaluations that emphasize algorithmic reasoning and tool calling, DeepSeek-V4-Flash in Think Max mode closes much of the performance gap with DeepSeek-V4-Pro. Flash achieves comparable Pass@1 rates on standard coding benchmarks like LiveCodeBench, making it an exceptional engine for code refactoring, continuous integration agents, and structured data extraction. However, as noted in the technical evaluation of DeepSeek-V4-Pro, the larger 1.6T parameter model retains a measurable advantage on raw world-knowledge tasks, broad multi-discipline synthesis, and complex domain-specific question answering (such as Simple-QA and Chinese-SimpleQA).

For cost-sensitive enterprise adopters, Flash functions as a workhorse model for high-frequency workflows. Applications that require sub-second generation latencies, high concurrency, or continuous background processing benefit from Flash's 13B active parameter profile. Teams can route standard agent interactions and test executions to Flash while reserving the 1.6T Pro variant for high-stakes strategic reasoning or open-ended analytical prompts.

Pay-per-token pricing on serverless inference

Serving large MoE models on self-managed infrastructure often leads to poor resource utilization. When engineering teams provision dedicated GPU instances to handle bursty traffic, hardware sits idle between request spikes, driving up the total cost of compute. Serverless inference avoids that by decoupling model execution from dedicated hardware provisioning. Billed strictly on active usage, DeepSeek-V4-Flash costs $0.15 per million input tokens and $0.30 per million output tokens.

Cost optimization and provisioning economics

At $0.15 per million input tokens and $0.30 per million output tokens, DeepSeek-V4-Flash costs roughly a twelfth of the per-token price of DeepSeek-V4-Pro ($1.75 input and $3.50 output per million tokens). There are no monthly base platform fees, no minimum seat licenses, and zero spend on idle GPU cycles. Modelling those figures side by side shows how per-token economics benefit development teams that process high-volume agent logs or run periodic batch evaluation jobs.

By adopting serverless endpoints for long-context workloads, engineering organizations eliminate the operational complexity of managing CUDA drivers, configuring vLLM distributed runners, or sizing tensor parallel groups across multi-GPU nodes. The billing model tracks exact token consumption, providing predictable unit economics as request volume scales.

Hosting region for the Flash endpoint

The DeepSeek-V4-Flash endpoint is hosted in eu-north1, inside the European Union. Lyceum runs its European data centres in Spain, Paris and the Nordics, and DeepSeek-V4-Flash is served from them, so EU data residency and GDPR framing apply to this model. Hosting is a per-model fact rather than a platform-wide one: a small number of catalogue models are served globally and say so on their own pages, so check the individual model before assuming its region.

Data routing and network security

When calling the DeepSeek-V4-Flash endpoint, traffic routes directly to compute nodes within the UK facility. This ensures deterministic network paths and predictable round-trip latency for applications operating across the UK and Western Europe. Enterprise security and governance parameters remain strictly enforced across all infrastructure zones.

• Deterministic EU routing: Requests land directly in eu-north1 without unmapped multi-region hops.
• Zero data retention: Prompts and generated completions are processed in volatile GPU memory and are never written to persistent databases or used for model training.
• Transient session caching: In-memory prompt caching persists only in GPU VRAM for the duration of active sessions (a few minutes at most) before eviction.
• Transport-layer encryption: All API calls enforce end-to-end TLS 1.3 encryption across transit channels.

By publishing exact data centre regions on a per-model basis, we ensure that compliance and engineering teams have full visibility into request geography, allowing them to match workload requirements with their specific compliance boundaries.

How to call it via the OpenAI-compatible API

Integrating DeepSeek-V4-Flash into existing software stacks requires no custom client libraries. The endpoint is 100 percent compatible with the standard OpenAI Python and Node.js SDKs. To transition an existing application, developers need only update the base URL to point to the serverless endpoint, supply their API key, and set the target model string to deepseek-ai/DeepSeek-V4-Flash.

Python implementation example

The following code snippet demonstrates how to initialize the client and submit a chat completion request with custom system parameters and reasoning control:

from openai import OpenAI client = OpenAI( base_url="https://api.lyceum.technology/api/v2/external/serverless", api_key="your-lyceum-api-key", ) response = client.chat.completions.create( model="deepseek-ai/DeepSeek-V4-Flash", messages=[ {"role": "system", "content": "You are an expert systems engineer. Analyze the provided log trace for CUDA OOM errors."}, {"role": "user", "content": "Inspect this stack trace and suggest tensor parallel adjustments: [paste trace]"} ], temperature=0.2, max_tokens=1024, ) print(response.choices[0].message.content)

Production considerations and deployment summary

• Architecture: 284B MoE with 13B active parameters per token, optimized for high-throughput code analysis and agent execution.
• Context & Economics: 1-million token context powered by CSA/HCA hybrid attention, priced at $0.15 input / $0.30 output per million tokens.
• Integration: OpenAI SDK drop-in replacement hosted in eu-north1 with zero data retention.

Run it on serverless inference.