DeepSeek V4 Flash Free-Access Quick-Start Guide

Straight to the point: right now no platform offers a "completely free, unlimited, permanent" DeepSeek V4 Flash API — every free route has a quota or an expiry. But the free entry points are genuinely numerous: SenseNova is in open beta, OpenCode Zen is free for a limited time, NVIDIA has a free prototyping tier, and DeepSeek official plus Alibaba, Tencent, Qiniu and ModelScope all hand out signup credits. Claim a few of these and you have a few million tokens for coding and testing. Each platform below is written as "what you get → where to claim it → how to hook it up", with the base_url and model ID spelled out. At the end there's a minimal working setup for Codex and Claude Code, a cheat-sheet table and a common-error checklist.

First, why this model: DeepSeek V4 Flash is the model DeepSeek moved to its 0731 public beta on July 31, 2026 — a 284B-total / 13B-active MoE with a 1M-token context window, roughly $0.14/M input and $0.28/M output. It sits in the cheapest input tier right now, which is exactly why so many platforms are using it as a free-acquisition hook. For coding, testing and low-traffic use, the free quotas below are plenty.

1. SenseNova Token Plan (free beta)

What you get: free calls during the beta, no top-up and no card required; the quota refreshes on a 5-hour window. V4 Flash gets roughly 1,500 / 500 / 150 calls per 5 hours depending on the source (check the console). One account can hold up to 20 API keys that share a single quota pool. The beta is time-limited and may convert to paid later — grab it while it's here.

How to claim:

1. Go to platform.sensenova.cn and sign up with a +86 phone number (overseas numbers don't work right now).
2. Log in, open the "API Keys" page in the console and create a key.
3. Save the key, base_url and model ID — you'll need them in the next step.

Connect to Claude Code / Codex: base_url https://token.sensenova.cn/v1, model ID deepseek-v4-flash, OpenAI-compatible. Codex speaks OpenAI format, so it connects directly:

export OPENAI_BASE_URL=https://token.sensenova.cn/v1
export OPENAI_API_KEY=sk-your-sensenova-key
codex --model deepseek-v4-flash

Claude Code speaks the Anthropic protocol and this endpoint is OpenAI format, so you'll need a translation layer first (see Section 11).

2. OpenCode Zen (limited-time free)

What you get: a limited-time promotion, $0 with no card, 1M context. The quota isn't public — push it and you'll hit 429 / 503 / "Free usage exceeded". Heads up: free-tier conversations may be used for model training, so don't put sensitive company code through it.

How to claim:

1. Go to opencode.ai/zen and register with an email or GitHub account.
2. Generate an API key in the console (the page lists the free models).
3. Save the key; the free model is opencode/deepseek-v4-flash-free.

Connect to Claude Code / Codex: base_url https://opencode.ai/zen/v1, model ID opencode/deepseek-v4-flash-free (or just deepseek-v4-flash-free). Codex connects directly:

export OPENAI_BASE_URL=https://opencode.ai/zen/v1
export OPENAI_API_KEY=your-zen-key
codex --model opencode/deepseek-v4-flash-free

This is promotional free — usable today, but it can be pulled at any time.

3. NVIDIA NIM (free prototyping)

What you get: a free developer tier, no credit card; rate-limited to about 40 requests/minute (not token-metered), keys valid 12 months, no SLA. It supports tool calling, so it works well for agent-style workloads.

How to claim:

1. Go to build.nvidia.com and register a free developer account.
2. Search for deepseek-v4-flash, open the model page and click "Get API Key" to generate an nvapi--prefixed key.
3. Save the key; it's valid for 12 months and metered by requests, not tokens.

Connect to Claude Code / Codex: base_url https://integrate.api.nvidia.com/v1, model ID deepseek-ai/deepseek-v4-flash. Codex connects directly:

export OPENAI_BASE_URL=https://integrate.api.nvidia.com/v1
export OPENAI_API_KEY=nvapi-your-nim-key
codex --model deepseek-ai/deepseek-v4-flash

These three are the lowest-friction routes since none needs a card. Quick split: SenseNova suits mainland developers who want frequent, sustained use (fresh quota every 5 hours); OpenCode Zen suits occasional long-context runs (1M), but skip it for sensitive code; NVIDIA suits agent work thanks to tool calling, and 40 req/min is usually enough for personal use.

4. DeepSeek official (5M-token signup bonus)

What you get: 5 million tokens for new users (roughly the first month / 30 days). Official pricing is about ¥1 input / ¥2 output (¥0.02 cached input) per million tokens. DeepSeek announced an API price increase on Aug 6, so anything after the bonus is at the then-current rate.

How to claim:

1. Go to platform.deepseek.com and register with a phone number.
2. Open "API Keys", create a key — the credit lands automatically.
3. Direct access runs on api.deepseek.com: use the /anthropic path for Claude Code and the /v1 path for Codex.

Connect to Claude Code / Codex: DeepSeek official ships an Anthropic-compatible endpoint, so Claude Code connects directly:

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=your-deepseek-key
export ANTHROPIC_MODEL=deepseek-chat

For Codex, point at the OpenAI endpoint https://api.deepseek.com/v1 with model deepseek-chat. Official docs use deepseek-chat as the alias for V4 Flash — you'll see the mapping in the console, so just fill in that name.

5. Alibaba Bailian (separate 1M tokens for V4 Flash)

What you get: a separate 1M-token free quota for V4 Flash; the whole lineup combined gets 70M tokens / 90 days, auto-issued to new users with zero friction. One heads-up: Bailian's cached-input price is ¥0.20/M, 10x the official ¥0.02 — watch long-context costs.

How to claim:

1. Go to bailian.console.aliyun.com, register an Alibaba Cloud account and enable Bailian.
2. Find DeepSeek V4 Flash in the model plaza and enable it — the credit is granted automatically.
3. Copy the OpenAI-compatible endpoint and API key from the console.

Connect to Claude Code / Codex: use the OpenAI-compatible endpoint from the console, model ID deepseek-v4-flash (Flash and Pro have separate quotas). Codex:

export OPENAI_BASE_URL=endpoint-from-bailian-console
export OPENAI_API_KEY=your-bailian-key
codex --model deepseek-v4-flash

6. Tencent TokenHub (1M tokens / 90 days)

What you get: 1M tokens / 90 days, zero friction for new users.

How to claim:

1. Go to cloud.tencent.com and register a Tencent Cloud account.
2. Find DeepSeek V4 Flash in the model plaza (TokenHub) and enable it.
3. Copy the OpenAI-compatible endpoint and API key; the model name is deepseek-v4-flash.

Connect to Claude Code / Codex: use the OpenAI-compatible endpoint from the model plaza, model ID deepseek-v4-flash. Point Codex's OPENAI_BASE_URL at it:

export OPENAI_BASE_URL=endpoint-from-tencent-model-plaza
export OPENAI_API_KEY=your-tencent-key
codex --model deepseek-v4-flash

7. Qiniu Cloud (3M-token signup bonus)

What you get: 3 million tokens for new users (individuals only), activated when you create your first API key. Qiniu was among the first to serve the V4 Flash GA and measured roughly 170 tokens/s.

How to claim:

1. Go to qiniu.com/ai and register an individual account.
2. Create your first API key in the console — that step activates the 3M-token bonus.
3. Save the endpoint https://api.qnaigc.com/v1 and the model deepseek/deepseek-v4-flash-20260731.

Connect to Claude Code / Codex: base_url https://api.qnaigc.com/v1, model ID deepseek/deepseek-v4-flash-20260731. Qiniu also supports the Anthropic protocol, so Claude Code can connect directly using the Anthropic endpoint from the console plus your Qiniu key. Codex connects directly:

export OPENAI_BASE_URL=https://api.qnaigc.com/v1
export OPENAI_API_KEY=your-qiniu-key
codex --model deepseek/deepseek-v4-flash-20260731

8. ModelScope (daily credits)

What you get: after the revamp, 250 "MoLi" credits/day; V4 Flash costs ~2 credits per call, so ~125 calls/day. Another source says ~2,000 calls/day per account with 50–500/day per model (dynamically adjusted). No card needed.

How to claim:

1. Go to modelscope.cn, log in with a bound Alibaba Cloud account.
2. Create a key on the "API-KEY" page; daily credits are granted automatically.
3. Use the inference endpoint from the console, model name deepseek-v4-flash.

Connect to Claude Code / Codex: use the inference endpoint from the console, model ID deepseek-v4-flash (incl. the 0731 variant), OpenAI-compatible. Point Codex at it:

export OPENAI_BASE_URL=inference-endpoint-from-modelscope
export OPENAI_API_KEY=your-modelscope-key
codex --model deepseek-v4-flash

Among these five signup bonuses the pick is simple: DeepSeek official for the biggest bonus (5M) and direct Claude Code support; Qiniu for speed (170 tokens/s) plus an Anthropic endpoint; Alibaba and Tencent are worth grabbing if you already have cloud accounts; ModelScope wins on the daily refresh for long-term free use. Claim them all and you're holding five or six free keys.

9. Other platforms with small signup gifts

These platforms have smaller free gifts, or are just promotional prices, but they all serve deepseek-v4-flash and work fine as backup keys:

• Nous Research (Nous Portal): DeepSeek V4 Flash 0731 at 90% off for 7 days — not free, but genuinely cheap.
• OpenRouter (openrouter.ai): paid deepseek/deepseek-v4-flash at ~$0.0679 / $0.168 per M; free deepseek/deepseek-v4-flash:free at ~50 calls/day, 20 RPM. That free tier is contested — on 2026-07-14 someone measured a "no provider" response (empty endpoints), a suspected dead listing. Check the live OpenRouter page before relying on it; never treat it as "free and it works".
• TokenRiver (tokenriver.cn): native OpenAI/Anthropic/Gemini protocols, 1M Token signup gift (verify on login).
• AiHubMix (aihubmix.com): OpenAI-compatible, permanent free test layer on signup (low quota; confirm V4 Flash is inside the free layer).
• FlowBar (flowbarai.com): 50,000 trial tokens (30 days), base_url https://api.flowbarai.com/v1, overseas nodes need a proxy.
• EasyRouter (easyrouter.io): 400 free points; no mainland-China service.
• iFlytek Spark (xinghuo.xfyun.cn): Spark Lite permanently free, free quota for new logins, V4 Flash is in the catalog.
• 302.AI, DMXAPI, Meshs One, 4SAPI, Together AI: these five are paid alternatives that serve V4 Flash with no meaningful free tier — look again when you need them.

Setup is identical to the platforms above: they're all OpenAI-compatible, so Codex or LiteLLM both work. For base_url, FlowBar is https://api.flowbarai.com/v1; the rest give you the endpoint in their consoles, and the model name is uniformly deepseek-v4-flash.

10. Don't waste your time: these "free" tiers don't include V4 Flash

When you see "DeepSeek free", first check whether it's actually V4 Flash. These will be a wasted trip:

• Volcano Ark: the 500K-token new-user gift is R1-only and explicitly excludes V4; V4 Flash only has a limited-time half price (until 2026-08-27).
• Baidu Qianfan: no new-user free quota found for V4 Flash; the platform is enterprise-agent oriented, not worth hunting.
• Cloudflare Workers AI: the free tier (10,000 Neurons/day) covers only the R1 distill, not V4 Flash.
• SambaNova: the free tier only has V3.1 / V3.2, no V4 Flash.
• Groq / Cerebras: neither hosts V4 Flash, only other DeepSeek versions — don't sign up for it.

11. How to use it: one minimal working setup

Once you've claimed a free key, wiring it into a tool is the real point. Below are two paths: Codex direct (OpenAI format, covers most free platforms) and Claude Code (direct only when the platform ships an Anthropic endpoint).

11.1 Codex: one config for every OpenAI-compatible platform

Codex (@openai/codex) natively speaks the OpenAI Chat Completions format, so it can connect to all the free endpoints above — the only differences are base_url and model. Install it first:

npm install -g @openai/codex

On macOS / Linux, environment variables are the quickest, using SenseNova as the example:

export OPENAI_BASE_URL=https://token.sensenova.cn/v1
export OPENAI_API_KEY=sk-your-sensenova-key
codex --model deepseek-v4-flash

The Windows PowerShell equivalent:

$env:OPENAI_BASE_URL="https://token.sensenova.cn/v1"
$env:OPENAI_API_KEY="sk-your-sensenova-key"
codex --model deepseek-v4-flash

To pin a platform permanently, write ~/.codex/config.toml and supply the key via OPENAI_API_KEY:

# ~/.codex/config.toml
model = "deepseek-v4-flash"
provider = "openai"
base-url = "https://token.sensenova.cn/v1"

After that just run codex. To switch platforms, change base-url and model to the cheat-sheet values; nothing else moves. For a one-shot task without the interactive UI, use codex exec: codex exec "fix the type errors in this repo" — it shares the same env vars and config.toml.

11.2 How to confirm it actually works

After changing config, probe with curl before you fight inside the tool. It's the same for every OpenAI-compatible platform:

curl https://token.sensenova.cn/v1/chat/completions \
  -H "Authorization: Bearer sk-your-sensenova-key" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"hi"}]}'

A choices[0].message.content in the response means it works. 401 is a wrong key, 404 is a wrong model ID, 429 means quota or rate limit — see Section 13 for fixes.

Prefer Python? The same script tests every platform — just swap three parameters:

from openai import OpenAI

client = OpenAI(
    base_url="https://token.sensenova.cn/v1",
    api_key="sk-your-sensenova-key",
)
resp = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Introduce yourself in one sentence"}],
)
print(resp.choices[0].message.content)

Swap base_url, api_key and model for each platform in the cheat sheet and run through them one by one — which key works becomes obvious fast.

11.3 Claude Code: direct only when the platform ships an Anthropic endpoint

Claude Code speaks the Anthropic Messages protocol and reads three env vars by default: ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_MODEL. DeepSeek official and Qiniu ship Anthropic endpoints, so you can fill these in directly:

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=your-deepseek-key
export ANTHROPIC_MODEL=deepseek-chat

The Windows PowerShell equivalent:

$env:ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
$env:ANTHROPIC_AUTH_TOKEN="your-deepseek-key"
$env:ANTHROPIC_MODEL="deepseek-chat"

To route the main model and sub-agents through V4 Flash too, set these as well:

export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-chat
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-chat
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-chat
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-chat

Reopen the terminal, launch Claude Code and run /status — if the Base URL and Model read correctly, you're live.

11.4 Want Claude Code on the other platforms? Put LiteLLM in front

SenseNova, NVIDIA, OpenCode Zen, Bailian, Tencent and ModelScope only give OpenAI-format endpoints, so filling base_url straight into Claude Code won't work. The simplest fix is a local LiteLLM proxy that translates the OpenAI endpoint into the Anthropic protocol; point Claude Code at the proxy:

# config.yaml (LiteLLM)
model_list:
  - model_name: deepseek-v4-flash
    litellm_params:
      model: openai/deepseek-v4-flash
      api_base: https://token.sensenova.cn/v1
      api_key: sk-your-sensenova-key

litellm --config config.yaml

Then set ANTHROPIC_BASE_URL to http://localhost:4000 and ANTHROPIC_AUTH_TOKEN to the LiteLLM key. The full walkthrough is in this site's LiteLLM guide.

To route all your free channels into Claude Code at once, list several models in one config and switch by name:

model_list:
  - model_name: sensenova-flash
    litellm_params:
      model: openai/deepseek-v4-flash
      api_base: https://token.sensenova.cn/v1
      api_key: ${SENSENOVA_KEY}
  - model_name: nvidia-flash
    litellm_params:
      model: openai/deepseek-ai/deepseek-v4-flash
      api_base: https://integrate.api.nvidia.com/v1
      api_key: ${NVIDIA_KEY}
  - model_name: qiniu-flash
    litellm_params:
      model: openai/deepseek/deepseek-v4-flash-20260731
      api_base: https://api.qnaigc.com/v1
      api_key: ${QINIU_KEY}

export ANTHROPIC_BASE_URL=http://localhost:4000
export ANTHROPIC_AUTH_TOKEN=sk-litellm-key
export ANTHROPIC_MODEL=sensenova-flash

Switching models is then a one-variable change to ANTHROPIC_MODEL. If you'd rather not hand-write config, a GUI switcher like CC Switch has provider presets built in (see this site's CC Switch guide).

11.5 A sensible claiming order

To claim everything efficiently: start with DeepSeek official (5M, direct Claude Code), then SenseNova (free beta, direct Codex) and Qiniu (3M), and finish with the Bailian, Tencent and ModelScope credits as backups. That gives you free keys across different protocols, so if one dies another is still standing.

11.6 Per-platform Codex config: only two lines differ

When pinning ~/.codex/config.toml long-term, the only difference between platforms is base-url and model. For example:

# NVIDIA NIM
model = "deepseek-ai/deepseek-v4-flash"
provider = "openai"
base-url = "https://integrate.api.nvidia.com/v1"

# OpenCode Zen
model = "opencode/deepseek-v4-flash-free"
provider = "openai"
base-url = "https://opencode.ai/zen/v1"

# DeepSeek official (OpenAI endpoint)
model = "deepseek-chat"
provider = "openai"
base-url = "https://api.deepseek.com/v1"

Keys always come from the OPENAI_API_KEY env var. For the rest (Bailian, Tencent, Qiniu, ModelScope), swap those two lines for the cheat-sheet values — the structure is identical. On Windows the config file lives at %USERPROFILE%\.codex\config.toml.

11.7 Rotating between several free keys

Free quotas are inherently unstable, so keep several keys and rotate. In Codex, switch by changing the env vars and starting a new session:

unset OPENAI_BASE_URL OPENAI_API_KEY
export OPENAI_BASE_URL=https://api.qnaigc.com/v1
export OPENAI_API_KEY=your-qiniu-key
codex --model deepseek/deepseek-v4-flash-20260731

When a key throws 429 or runs dry, move to the next — no reinstall needed. Same idea in Claude Code: swap ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN.

If you only plan to use one tool, the conclusion collapses to a sentence: Codex only — set up base-url and model for any platform per 11.1; Claude Code only — register DeepSeek official and set the three env vars per 11.3; that's the least-effort combo.

12. Cheat sheet

Notes: Codex connects directly to every OpenAI-compatible platform; only Claude Code needs the distinction. "Direct" means the platform ships an Anthropic endpoint; "via LiteLLM" means a translation proxy is required. Most platforms want a phone number or real-name verification — register with real info to avoid being flagged.

13. Common errors and fixes

• 401 Unauthorized / Invalid API key: the key is mistyped or has a stray space. Some keys carry prefixes (sk-, nvapi-) — copy them exactly from the console.
• 404 / model not found: wrong model ID. They're not uniform across platforms — deepseek-v4-flash, opencode/deepseek-v4-flash-free, deepseek-ai/deepseek-v4-flash, deepseek/deepseek-v4-flash-20260731 all differ. Copy from the cheat sheet.
• 429 Too Many Requests: you've hit a quota or rate limit. SenseNova is a 5-hour window, NVIDIA ~40 req/min, OpenRouter free ~50/day — wait for the window to reset or switch keys.
• 503 / "Free usage exceeded": common with limited-time promos like OpenCode Zen; the free quota is spent or the promo ended. Normal.
• no provider / empty endpoints: the OpenRouter :free dead-listing problem — "Free" on a directory page doesn't mean someone is actually serving it. Switch to the paid tier or another platform.
• Claude Code errors against an OpenAI endpoint: Claude Code doesn't speak OpenAI protocol. Confirm you're on the platform's /anthropic endpoint, or go through LiteLLM.
• Connection timeout: overseas nodes (FlowBar, some relays) need a proxy; timeouts on a mainland network are expected. Use a proxy or switch to a mainland-direct platform (SenseNova, Qiniu, Bailian, Tencent, ModelScope).

14. A few shared gotchas

• "Free" ≠ unlimited: every free tier has a quota, an expiry, or both. SenseNova refreshes on 5h, ModelScope daily, DeepSeek's 5M is first-month/30 days, Bailian/Tencent 1M expire in 90 days, Qiniu's 3M activates on first-key creation and is individuals-only.
• Your data is the price: free tiers like OpenCode Zen state conversations may improve their models. Keep sensitive company code off the free tiers.
• It can disappear any day: OpenCode Zen is a limited-time promo and NVIDIA has no SLA — treat these as nice-to-have, not production dependencies.
• Region locks: SenseNova only takes +86 numbers; FlowBar's nodes are overseas and need a proxy; EasyRouter excludes mainland China.
• Console numbers win: exact per-window counts (SenseNova calls per 5h, ModelScope daily counts) differ by source — the console is the source of truth.
• Free terms change faster than pricing: this article is a snapshot around Aug 12, 2026 — always confirm against the platform's current page.

15. FAQ

• Can I re-claim a signup bonus after it's used up? Signup gifts are usually one-time; window- or daily-based ones (SenseNova, ModelScope) refresh automatically.
• Can I register multiple accounts on one platform? Technically yes, but most platforms enforce one account per person (Qiniu individuals-only, SenseNova needs +86), and mass registration risks flagging — not recommended.
• Will my free key just stop working? Promo tiers (OpenCode Zen, NVIDIA prototyping) can change anytime — normal, which is why the article keeps saying to hold several keys.
• Can I run production on the free tier? Not recommended — no SLA, can vanish anytime. Production belongs on DeepSeek official or a solid paid relay.
• How long does the quota last? Roughly: a V4 Flash conversation is a few thousand tokens, so 1M tokens covers a few hundred medium-length conversations — one to two weeks of personal coding/testing. The 5M official bonus is more generous. Long-context analysis is the real drain.
• How do I make the free quota last longer? Lean on cached input, cap output length, and don't stuff whole files into context. V4 Flash cached input is only ¥0.02/M — re-reading the same material with cache hits saves a lot on long tasks.

16. What to do next

1. Register one or two platforms and claim your keys (DeepSeek official + SenseNova is the recommended starter — one gives direct Claude Code, the other direct Codex).
2. Set up Codex or Claude Code per Section 11 and confirm with curl before you start working.
3. When a quota runs out or throws 429, rotate to the next key — don't fight one dead path.
4. Bookmark the cheat sheet in Section 12; free quotas change fast, so check back in a while.