Where to Run DeepSeek V4 Flash Free and Cheap: Every Provider Compared (2026)

Start here: Third-party hosts still serve DeepSeek V4 Flash at the old flat rates or cheaper. As of August 22, 2026, the cheapest verified routes are OpenRouter's auto-router (as low as $0.06 input / $0.12 output per million tokens via StreamLake and Baidu Qianfan with 57 percent off) and DeepInfra ($0.09 / $0.18), both well below DeepSeek's own off-peak pricing ($0.22 / $0.66). If you want a low-friction agent trial with generous limits, OpenCode Go is $5 for the first month when you use referral code M6HEHM4JM5.

Last updated: August 22, 2026

Official Sources#

DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts model with 284B total parameters, 13B activated parameters, and a 1M-token context window. It is designed for fast inference and high-throughput coding workloads, released under an MIT license. On August 16, 2026, DeepSeek shifted its first-party API to peak and off-peak pricing, raising rates by 2x to 4x depending on the hour. Third-party providers have not all followed, which is why the cheapest access today is no longer direct from DeepSeek.

This post maps every verified route: the genuinely cheap ones, the official first-party tiers, free or near-free options, and local. Prices are per million tokens and were verified August 22, 2026. Pricing pages move, so treat the numbers as a snapshot.

What DeepSeek V4 Flash is, in one paragraph#

DeepSeek V4 Flash is DeepSeek's efficiency-tier open-weights coding model, released April 23, 2026 and refreshed July 31, 2026 as DeepSeek-V4-Flash-0731. It is a Mixture-of-Experts model with 284B total parameters, 13B activated per forward pass, a 1M-token context window, and an MIT license. It introduces a hybrid attention architecture combining Compressed Sparse Attention and Heavily Compressed Attention for efficient long-context processing. On coding benchmarks it delivers strong performance at a fraction of the cost and compute of closed models. For the full technical breakdown and cost math, see the DeepSeek V4 developer guide and the V4 economics post. This page covers access and pricing.

The cheapest hosted routes (verified August 22, 2026)#

For most developers the right answer is a hosted API. DeepSeek V4 Flash is MIT-licensed open weights, so any inference provider can serve it and compete on price. Here is the live picture, sorted by blended cost.

A few things worth knowing before you pick a row:

• OpenRouter is a router, not a host. Its endpoints API shows 16-plus providers serving deepseek/deepseek-v4-flash, and OpenRouter sends your request to the cheapest or fastest one that meets your constraints. You get failover and price competition without managing keys for each host. StreamLake and Baidu Qianfan are currently offering 57 percent off, which is why OpenRouter's blended rate sits near $0.06 input / $0.12 output. There is no :free variant for DeepSeek V4 Flash on OpenRouter despite catalog pages suggesting otherwise - the endpoints list for the free slug is empty. See the OpenRouter profile for the wider router picture.
• Quantization matters. The cheapest routes (DeepInfra, Sail Research, StreamLake) serve fp4 quantized weights; Together, Novita, and DeepSeek serve fp8 or higher precision. For coding agents the quality gap is usually small but real, and it is the reason the price differs. Test your own task before optimizing purely on price.
• Third-party hosts have not adopted peak/off-peak. Together AI, DeepInfra, and most OpenRouter providers still bill the old flat rate ($0.14 / $0.28 or cheaper), which undercuts DeepSeek's own off-peak pricing. This arbitrage will not last forever, but it is live right now.
• Blended price is lower than the table suggests. On a typical 3:1 input-output mix with aggressive caching, the cheapest routes (StreamLake, Baidu, DeepInfra) land near $0.07 to $0.10 per million tokens blended. OpenRouter's live pricing page shows weighted averages accounting for cache hit rates.

For the worked cost-per-task math versus closed models, the V4 economics post and the budget AI coding models comparison run the numbers.

Direct from DeepSeek: peak vs off-peak#

On August 16, 2026, DeepSeek replaced its flat-rate API pricing with a two-tier system: off-peak and peak. The change raised rates by 57 percent to 371 percent depending on the token type and hour. Here is the current structure as verified August 22, 2026 from the official pricing page:

Peak hours are 01:00-04:00 and 06:00-10:00 UTC, seven hours daily. Everything else is off-peak, and off-peak is exactly half the peak rate. Important change effective August 23, 2026: weekends (Saturdays and Sundays, Beijing Time) will be all off-peak, removing the peak windows on those days. This means if you schedule workloads for weekends and outside the two UTC windows on weekdays, you stay on the lower tier.

For context, the old flat rate (live until August 16) was $0.14 input and $0.28 output per million tokens. Off-peak output is now $0.66, a 136 percent increase. Peak output is $1.32, a 371 percent increase. Cache-hit pricing rose even more sharply - off-peak is 2.5x the old rate, peak is 5x. This hits agentic workloads hardest because they rely on caching to keep costs manageable.

The first-party API supports the full 1,048,576-token context, tool calling, structured output, and reasoning effort (high or xhigh). Use it when you want the reference behavior, predictable pricing directly from the source, or compliance requirements that rule out third-party providers. But if pure cost is the goal, third-party hosts are currently cheaper.

Free and near-free routes#

Three paths will run DeepSeek V4 Flash with little or no upfront cost today. None is unlimited, and the genuinely free ones are unstable or time-limited, so read the terms before wiring production workloads to them.

• OpenCode Go referral credits. If you want to test DeepSeek V4 Flash through an agent-first interface before committing to per-token APIs, OpenCode Go is the easiest place to start. Use referral code M6HEHM4JM5 and the subscription is $5 for the first month, then $10 per month. Go bundles DeepSeek V4 Flash with other open-weight models (GLM-5, Kimi K2, Qwen, MiniMax) under usage caps measured in requests per five-hour window. DeepSeek V4 Flash has the highest request limit in the catalog: 31,650 requests per five hours. Treat this as a low-friction trial rather than a production route - quotas reset, and Go is designed for evaluation and prototyping, not for scaling a production agent.
• OpenCode Zen free tier. OpenCode Zen (the team's public model gateway) lists a deepseek-v4-flash-free endpoint at https://opencode.ai/zen/v1/chat/completions. The free tier exists but is frequently exhausted. GitHub issues from July and August 2026 show persistent 429 FreeUsageLimitError responses even for minimal requests, suggesting the quota is either shared globally or resets unpredictably. The paid Zen rate is $0.14 input / $0.28 output per million tokens. Treat the free tier as promotional and likely unavailable; if it works when you test it, use it, but do not build production plans around it.
• Hugging Face Inference API limited access. Hugging Face occasionally opens free inference windows for high-profile open-weights models shortly after release. DeepSeek V4 Flash 0731 had a limited window in late July 2026. Free windows close, so check the model page for current status. When available, it is rate-limited and best suited for quick tests, not sustained workloads.

If you want free and you want it to last, the honest answer is local: download the weights and run them yourself. That is the local and self-host section below; if your hardware is more modest, our best local models hub is the starting point for laptop-class options.

Which route should you pick?#

• Cheapest production tokens: OpenRouter's auto-router or DeepInfra, accepting fp4 quantization. OpenRouter currently routes to StreamLake or Baidu Qianfan at 57 percent off for the lowest blended rate. Validate quality on your own task first.
• Stable flat rate without surprises: Together AI at $0.14 / $0.28, the old DeepSeek rate, with strong uptime. A known quantity.
• First-party reference with peak/off-peak scheduling: DeepSeek direct, $0.22 / $0.66 off-peak if you can schedule workloads outside 01:00-04:00 and 06:00-10:00 UTC. Weekends are all off-peak starting August 23.
• Agent-first trial with generous limits: OpenCode Go with referral code M6HEHM4JM5, $5 first month. Best for evaluating the model in a coding-agent loop without managing API keys.
• Self-host at scale or air-gapped: download the weights from Hugging Face and serve with vLLM or SGLang. Only worth it at steady high volume or when compliance requires on-premises inference.
• Truly offline or laptop-local: not practical. Even a 4-bit quant of V4 Flash needs multi-GPU memory. Step down to a smaller dense model for laptop inference.

Local and self-host#

Because the weights are MIT-licensed and on Hugging Face, you can run DeepSeek V4 Flash with no per-token cost at all - if you have the hardware.

• vLLM and SGLang. DeepSeek V4 Flash 0731 has first-class support for vLLM and SGLang, the two most common open serving stacks. The model card lists vllm serve "deepseek-ai/DeepSeek-V4-Flash-0731" as working out of the box. This only pays off above a high, steady token volume where amortized GPU cost beats per-token API pricing. Below that line, a hosted route is cheaper and far less operational work.
• Ollama. Ollama does not yet have an official GGUF build for DeepSeek V4 Flash 0731 as of August 22, 2026. Community builds may exist but are not verified here. At 284B total parameters, even a 4-bit quant needs multi-GPU high-RAM hardware, not a laptop.
• Transformers and llama.cpp. The Hugging Face model card documents native support for Transformers and experimental llama.cpp paths, plus Ascend NPU builds. These are self-host options for datacenter-class hardware, not consumer setups.

If your goal is genuinely local coding on modest hardware, a smaller dense model is the right tool - see the best local coding LLMs.

The Novita 90 percent Vercel Gateway promo (ended August 11)#

On August 4, 2026, Vercel AI Gateway announced a 90 percent discount on DeepSeek V4 Flash routed through Novita, dropping the effective rate to $0.014 input / $0.028 output per million tokens for Pro customers. The promo was time-limited and ended August 11, 2026. After that date, Novita's rate returned to the standard $0.14 / $0.28 through the gateway. The promo post is archived here for reference, but do not expect the 90 percent off rate to still be live. The standard Vercel AI Gateway blended rate for DeepSeek V4 Flash is currently $0.09 input / $0.18 output per million tokens, routing across multiple providers including Novita, DeepInfra, Fireworks, and others.

When not to self-host#

Self-hosting DeepSeek V4 Flash only makes economic sense at steady high volume. If you run fewer than several billion tokens per month, paying $0.09 to $0.22 per million on a hosted API is almost certainly cheaper than the amortized cost of GPU capacity, electricity, and operational overhead. The break-even line depends on your GPU access cost, but for most teams it sits well into the hundreds of dollars per month of sustained token spend. Below that line, use a hosted route. Above it, vLLM or SGLang on reserved capacity starts to pencil out.

The other reason to self-host is compliance: air-gapped environments, data residency requirements, or contractual restrictions on third-party inference. If that is the driver, cost is not the primary variable.

FAQ#

Is DeepSeek V4 Flash free?#

No permanent unlimited free hosted API exists. OpenCode Zen has a free tier that is frequently exhausted and unreliable. OpenCode Go is $5 the first month with referral code M6HEHM4JM5, then $10 per month. The weights themselves are free under an MIT license, so self-hosting has no per-token cost, but that requires datacenter-class GPU hardware. For practical purposes, the cheapest stable hosted access is DeepInfra or OpenRouter at $0.09 to $0.12 per million input tokens.

What is the cheapest way to use DeepSeek V4 Flash?#

On hosted APIs, the cheapest verified routes as of August 22, 2026 are OpenRouter's auto-router (as low as $0.06 input / $0.12 output per million tokens via StreamLake or Baidu Qianfan with 57 percent off) and DeepInfra at $0.09 input / $0.18 output. Both are fp4 quantized, so validate quality on your own task. Self-hosting is cheapest only at high sustained volume.

Can I run DeepSeek V4 Flash with Claude Code, Cursor, or OpenCode?#

Yes. DeepSeek V4 Flash exposes an OpenAI-compatible endpoint at https://api.deepseek.com, so any tool that accepts OpenAI-shaped APIs can point at it with a config change. Set OPENAI_BASE_URL=https://api.deepseek.com and OPENAI_API_KEY=your-deepseek-key, then use model=deepseek-v4-flash. Most coding agents (Claude Code, OpenCode, Cline, Cursor, GitHub Copilot) support custom OpenAI-compatible endpoints. The DeepSeek V4 developer guide covers setup for each tool.

Can I run DeepSeek V4 Flash locally on my laptop?#

Not practically. At 284B total parameters and 13B active, even a 4-bit quantized build needs multi-GPU high-RAM hardware. Ollama does not yet have an official GGUF build for V4 Flash 0731. For local coding on consumer hardware, a smaller dense model (Qwen3-8B, Llama 3.3-8B, DeepSeek-Coder-6.7B) is the right choice. See the best local coding LLMs.

What license is DeepSeek V4 Flash under?#

MIT. The weights are permissive with no regional or commercial restrictions. You can use, modify, deploy, and commercialize DeepSeek V4 Flash without a separate license from DeepSeek.

What are DeepSeek V4 Flash peak hours?#

Peak hours are 01:00-04:00 and 06:00-10:00 UTC, seven hours daily. Everything else is off-peak. Starting August 23, 2026, weekends (Saturdays and Sundays, Beijing Time) are all off-peak regardless of the clock. Off-peak rates are exactly half of peak rates. If you schedule workloads outside those windows, you stay on the cheaper tier.

Continue Reading#

• DeepSeek V4: The Developer's Guide to Flash and Pro - the full technical breakdown and setup guide
• DeepSeek V4 Economics: Cost, Quality, and the Frontier for Agentic Coding - the worked cost-per-task numbers behind the pricing table above
• Budget AI Coding Models Compared 2026 - side-by-side cost and quality for V4 Flash, GLM-5.2, Qwen3, and others
• Where to Run GLM-5.2 Free and Cheap - the same provider grid for Z.ai's open-weights coding model
• Where to Access Kimi K3 - Moonshot's 2.8T open-weights model access routes
• Where to Access AI Models in 2026 - the hub covering access routes, free tiers, and prices for every major model
• AI Coding Tools Pricing 2026 - the full pricing directory for coding agents, IDEs, and model APIs

Sources#

• DeepSeek API Pricing (official) - fetched August 22, 2026
• DeepSeek API Documentation (official) - fetched August 22, 2026
• Hugging Face: deepseek-ai/DeepSeek-V4-Flash-0731 - fetched August 22, 2026
• OpenRouter: deepseek/deepseek-v4-flash - fetched August 22, 2026
• Together AI Pricing - fetched August 22, 2026
• DeepInfra model catalog - referenced August 22, 2026
• Vercel AI Gateway: deepseek-v4-flash - fetched August 22, 2026
• OpenCode Go referral offer - fetched August 22, 2026
• DeepSeek V4 Flash peak/off-peak analysis (independent) - fetched August 22, 2026