DeepSeek-V4 Flash suits cost-sensitive teams running long-context research, document review, or high-volume agentic workflows who don't need frontier-level coding autonomy; it scores 96% on HumanEval for everyday code generation. It swaps in for pricier proprietary APIs, self-hosts under MIT license, but still trails Claude Opus 4.7 on complex multi-file coding.
DeepSeek's V4 Flash scores 88.1% on GPQA Diamond, the faster and cheaper sibling to DeepSeek-V4 Pro in the open-source Mixture-of-Experts V4 family. It ships a native 1,048,576-token context window, MIT-licensed for self-hosting, built for long-context and high-volume agentic work at a fraction of proprietary-API pricing.
Provider: DeepSeek · Family: DeepSeek-V4
Context window: 1,048,576 tokens · Max output: 384,000
Input modalities: text, image, tool-calls · Output: text, tool-calls
About DeepSeek-V4 Flash
DeepSeek-V4 Flash is an open-source Mixture-of-Experts language model released by DeepSeek on April 24, 2026, as part of the DeepSeek-V4 preview. The model activates only 13 billion of its 284 billion total parameters per token, making it far more efficient than dense equivalents while keeping frontier-class reasoning and coding ability. It is the faster, cost-optimized sibling to DeepSeek-V4 Pro (1.6T total / 49B active), both licensed under MIT for commercial deployment and self-hosting without vendor permission. The V4 series introduces Hybrid Attention, combining Compressed Sparse Attention and Heavily Compressed Attention, cutting per-token compute to roughly a tenth of V3.2's at 1M-token context. V4 Flash posts 88.1% on GPQA Diamond (graduate-level reasoning) and 96.4% on HumanEval (code generation), matching or beating most prior-generation proprietary models. Its Artificial Analysis Intelligence Index composite is 47, ahead of DeepSeek-V3.2 and roughly level with Claude Sonnet 4.6 in extended-reasoning mode, though behind V4 Pro's 52. On agentic coding it trails the closed frontier; see the comparison FAQ below for the exact SWE-bench gap against Claude Opus. Both thinking (standard, high, or max reasoning effort) and non-thinking modes run on the same model ID, toggled per request without separate routing logic, matching the pattern DeepSeek used for the deprecated deepseek-chat/deepseek-reasoner split. The open-weights release on Hugging Face (huggingface.co/deepseek-ai/DeepSeek-V4-Flash) supports self-hosting, fine-tuning, and research use with no vendor lock-in. Official weights ship in FP4+FP8 mixed precision at roughly 160GB for full precision; community GGUF quantizations bring that down to about 60GB at Q4_K_M with minimal quality loss. A single H100 80GB GPU at FP8 (or two H100s for higher throughput) runs V4 Flash in production, and community quantizations have pushed it onto single RTX 4090s at Q4. Architecturally, Compressed Sparse Attention applies 4x KV-cache compression along the sequence dimension, then sparse attention (top-1024 relevant tokens per query) plus a 128-token local window balances retrieval quality against compute; Heavily Compressed Attention applies 128x compression on a second attention head for long-range synthesis. At 1M tokens this drops KV-cache use to 7% of V3.2's. Manifold-Constrained Hyper-Connections (Sinkhorn-Knopp orthogonalization) stabilize training, and the DeepSeekMoE router uses Sqrt(Softplus) affinity scoring with hash-routed experts in the early layers. V4 Flash is text-only at launch: no native image, audio, or video input. DeepSeek's separate V4-Vision model family handles images at roughly 90 tokens per image versus 870 for competitors like GPT-4o, and audio/video support is announced for later releases. Tool use and function calling are fully supported through both OpenAI-compatible and Anthropic-compatible schemas, enabling agentic workflows with reliable structured output. DeepSeek's post-training combines supervised fine-tuning with group relative policy optimization in a two-stage pipeline: domain experts develop alignments independently, then models are unified through on-policy distillation. This is lighter-touch than Claude's Constitutional AI or GPT's RLHF-heavy approach, so V4 Flash is more permissive on edge cases and occasionally allows outputs stricter models refuse. Red-teaming details are not yet public; the model remains in preview status and its safety posture may evolve. DeepSeek's legacy API names deepseek-chat and deepseek-reasoner route to V4-Flash's non-thinking and thinking modes during a transition period, but both are fully deprecated after July 24, 2026, 15:59 UTC. Any production workload on those names must migrate to deepseek-v4-flash or deepseek-v4-pro before that cutoff. Azure AI Foundry, Google Vertex AI, AWS Bedrock, Fireworks, and Together AI all added V4-Flash within weeks of the public launch, reducing single-vendor dependency for enterprises. Best for cost-sensitive teams running long-context workloads (research analysis, document processing, contract review), open-source-first builders who want reproducible self-hosted deployments, and enterprises needing data residency outside the US. Skip it for mission-critical agentic coding at the frontier, for vision or voice-first products before V4-Vision matures, and for use cases with strict safety and alignment requirements.
Pricing
Prices follow DeepSeek's native API rate card. Prompt caching cuts cached input to $0.0028 per million tokens, a 98% discount versus fresh input. Third-party hosts (OpenRouter among them) list their own per-token rates, which can run slightly higher or lower than the native API. A batch-API discount has been mentioned by DeepSeek but is not yet confirmed.
Key Features
- 1M-Token Context, No Gating: Full 1,048,576-token context ships on both Flash and Pro at the same price tier, with a 384,000-token max output and 97-99% needle-in-haystack retrieval accuracy at full length.
- Reasoning Modes (Standard / High / Max): Toggle reasoning effort per request without switching model IDs; max mode adds visible chain-of-thought for STEM tasks, standard mode is fastest and cheapest for routine queries.
- Mixture-of-Experts Efficiency: Only a fraction of the network activates per token, so inference cost tracks the active-parameter count rather than the full model size, letting V4 Flash rival larger dense models at a fraction of the compute.
- Prompt Caching (24h TTL): Cached input tokens get a steep discount versus fresh input, useful for repeated system prompts, batch document analysis, and retrieval loops; the cache key resets if the system prompt changes.
- Native Tool Use, Dual API Format: Function calling works with both OpenAI-compatible and Anthropic-compatible schemas out of the box, so existing agent frameworks can point at V4 Flash without rewriting tool schemas.
Pros
- Ultra-low per-token pricing (see the cost FAQ) makes long-context and high-volume agentic workloads viable at a fraction of what proprietary frontier APIs charge.
- Fully open-weight under MIT: self-host, fine-tune, or deploy air-gapped without contacting DeepSeek, unlike API-only competitors.
- 1M-token context ships free on the base tier instead of gated behind a premium plan, which is unusual among frontier-scale models.
Cons
- Trails Claude Opus and GPT-5.5 on complex, multi-file agentic coding; see the comparison FAQ for the exact benchmark gap.
- No native image, audio, or video input at launch; DeepSeek's separate V4-Vision model or a third-party ASR/TTS service is required for those modalities.
- Lighter safety alignment than Claude or GPT: DeepSeek's post-training is less stringent, so it may allow edge-case outputs stricter models refuse.
Benchmarks
- humaneval: 96.4
- live bench: 91.6
- lmarena elo: 1300
- gpqa diamond: 88.1
- lmarena rank: 50
- swe bench verified: 79
- artificial analysis intelligence index: 47
- artificial analysis price blended per m: 0.21
- artificial analysis speed tokens per sec: 93.9
Frequently Asked Questions
What are DeepSeek-V4 Flash's pricing plans in 2026?
DeepSeek-V4 Flash costs $0.14 per million input tokens and $0.28 per million output tokens on DeepSeek's native API. Prompt caching drops cached input to $0.0028 per million, a 98% discount. Third-party hosts like OpenRouter, Azure AI Foundry, Vertex AI, and AWS Bedrock price it a little differently, so check current rates before you commit to one provider.
What do you get on DeepSeek-V4 Flash's free tier?
New accounts get 5 million tokens of free usage per 30-day period with no credit card required. Once those tokens are used or the 30 days lapse, billing switches to the standard per-token rate. The free tier applies to the native DeepSeek API only; third-party hosts set their own trial terms.
What should you use instead of DeepSeek-V4 Flash?
DeepSeek V4 Pro offers more reasoning headroom at a much higher per-token cost. GPT-4o mini is worth a look if you need native vision support today. Qwen3.7-Plus matches V4 Flash's 1M-token context class and adds built-in vision and video understanding. Pick V4 Flash itself when cost and long-context throughput matter more than multimodal input.
Is DeepSeek-V4 Flash better than DeepSeek V4 Pro?
On SWE-bench Verified, DeepSeek-V4 Flash scores 79% against Claude Opus 4.7's 93.9% and GPT-5.5's 94.6%, a real but not enormous gap for a model priced this far below both. For anything short of mission-critical multi-file refactors, the price difference usually outweighs the coding gap.
What does it take to start using DeepSeek-V4 Flash?
Grab an API key at platform.deepseek.com, then call the OpenAI-compatible or Anthropic-compatible chat endpoint with model id deepseek-v4-flash; the free tier covers your first 5 million tokens with no credit card. To self-host, download the MIT-licensed weights from Hugging Face and deploy on one 80GB H100-class GPU using FP8, or run a quantized GGUF build on a consumer RTX 4090.
Top Alternatives
- DeepSeek V4 Pro: V4 Pro posts an 80.6% SWE-bench Verified score, extra coding headroom that costs roughly 12x the per-token price of Flash. V4 Flash wins on plain cost and inference speed instead.
- GPT-4o mini: V4 Flash runs roughly 8x the context window at near-identical input pricing to GPT-4o mini. GPT-4o mini's edge is native image understanding today, since V4 Flash stays text-only until DeepSeek ships V4-Vision broadly.
- Qwen3.7-Plus: V4 Flash undercuts Qwen3.7-Plus on per-token pricing while matching its 1M-context class. Qwen3.7-Plus is the one to reach for if built-in vision and video understanding is a requirement today.