The documented DeepSeek V4 Flash context window is up to 1M tokens, with up to 384K tokens of output, as of July 31, 2026. The open-weight configuration specifies 1,048,576 maximum positions. Those numbers describe capacity, not guaranteed recall accuracy: input plus generated output must fit in the window, clients may set lower limits, and long-context quality must be tested on your workload.
Context snapshot — verified July 31, 2026
Field Current value API context Up to 1M tokens Open-weight config max_position_embeddings: 1048576Maximum output Up to 384K tokens Constraint Input + output must fit within context Current version DeepSeek-V4-Flash-0731, API public betaInput modality Text only Sources: Models & Pricing and official config.json.
Independent site — not affiliated with DeepSeek.
1M versus 1,048,576
DeepSeek's product documentation uses the rounded phrase “1M context,” while the model configuration uses 1,048,576 positions. These are compatible descriptions at different levels of precision, not two separate model editions. User-facing copy should say “up to 1M tokens”; technical capacity planning may cite the exact configuration value.
The maximum output is 384K, not 1M. A request cannot combine a full 1M-token input with another 384K tokens of output because both must fit within the context budget. Reserve output space when constructing prompts. Endpoint and mode details are in DeepSeek V4 Flash API.
Capacity is not reliable recall
A context window answers “how much can the model accept?” It does not answer “will the model retrieve every buried fact, follow every old instruction, or use tools reliably at that length?” DeepSeek's technical report presents first-party long-context experiments, but those are vendor evaluations and do not guarantee your repository, document mix, or multi-turn agent state.
The corpus includes a July community report claiming reliable behavior near 100K, drift around 250K, and clear retrieval failures by 500K. That is a single, unreplicated user report, not a supported product limit. It is useful only as a reason to test instead of assuming 1M effective recall.
How to test your effective context
Use a task-shaped evaluation rather than one theatrical “needle in a haystack” prompt:
- Build fixed context sizes—for example 32K, 100K, 250K, 500K, and near the client limit.
- Test fact retrieval, instruction retention, cross-file synthesis, and tool use separately.
- Repeat each task across multiple positions and runs; record false confident answers, not only exact misses.
- Compare single-turn documents with multi-turn agent histories. They stress different failure modes.
- Record client, API surface, model ID, served revision, reasoning effort, and date.
- Measure tokens, latency, cache usage, retries, and task acceptance.
This also prevents a common benchmark error: treating Preview results such as MRCR 1M as proof of 0731 performance under a different harness. For versioned results and limitations, see DeepSeek V4 Flash benchmarks.
Client limits and local-memory costs
The service can advertise 1M while a specific agent integration chooses a lower client-side value. Check the active client's configuration and provider metadata rather than assuming it exposes the full server limit.
For local inference, context length also changes KV-cache memory and runtime behavior. The official number does not prove that a particular GPU, quantization, runtime branch, or workstation can run 1M tokens. This launch set intentionally does not publish device promises because the collected cases mix Preview and 0731 artifacts, custom patches, aggressive quantization, and different memory layouts.
Context cost and caching
Large repeated prefixes can benefit from context caching, but only reported cache hits receive the lower input rate. A long context that misses cache or triggers retries can still be expensive. See DeepSeek V4 Flash pricing for the dated rates and worked examples.
Compressing context is not automatically safe either: summaries can delete the very constraint or code path needed later. Prefer retrieval and explicit working sets, keep invariant instructions short, and re-inject critical acceptance criteria before high-risk changes.
For the dated model-family overview and release status, see DeepSeek V4 Flash.
FAQ
What is the DeepSeek V4 Flash context length?
Up to 1M tokens in the API documentation; the official 0731 configuration lists 1,048,576 positions.
Can it output one million tokens?
No. The documented maximum output is up to 384K tokens, and input plus output must remain within the context window.
Does 1M context mean perfect recall across 1M tokens?
No. It is a capacity specification. Reliable retrieval and instruction following require workload-specific tests.
Can I send a 1M-token PDF directly?
Not as a PDF or document block. The current API is text-only, so text must be extracted first; see Is DeepSeek V4 Flash multimodal?.
Will every Claude Code or API client expose the full 1M?
Not necessarily. Clients and providers may enforce smaller limits; inspect the active configuration.
Return to the complete DeepSeek V4 Flash overview for price, API, download, size, benchmarks, and local-use guidance.
Independent informational website. Not affiliated with, endorsed by, or sponsored by DeepSeek. DeepSeek and related marks belong to their respective owner. Last verified July 31, 2026.