by DeepSeek
DeepSeek's April 2026 open-weight Flash model, retired on 10 September 2026 and replaced by DeepSeek V4.1 Flash.
DeepSeek retired V4 Flash on 10 September 2026. Its API name now routes to V4.1 Flash, billed at $0.15 input and $0.60 output per million tokens off-peak. Use deepseek-flash for new work. The open MIT weights stay on Hugging Face for teams that need the original April model, which scored 79% on SWE-bench Verified.
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0045 | $0.0006 | $0.0051 |
| Support reply | $0.0003 | $0.0002 | $0.0005 |
| One coding agent run | $0.030 | $0.012 | $0.042 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Retired by DeepSeek. Calls to the legacy name are billed at the current Flash rates listed in the plan tiers, with peak rates twice the off-peak ones. The vendor's pricing page (read 2026-11-09) lists no free tier. The replacement model is priced the same.
DeepSeek-V4 Flash, the 284B-parameter open-weight model released in April 2026, was retired by DeepSeek on 10 September 2026. Its legacy API name now serves DeepSeek V4.1 Flash with a 1M-token context window, and the MIT-licensed April weights remain available on Hugging Face for self-hosting.
Reads
Writes
Retired and routed to its successor
DeepSeek retired the model; its legacy API name now routes to the replacement and bills at Flash rates.
1M-token context
The 1,048,576-token window with a 384,000-token output cap carries over to the replacement model.
Open MIT weights
The April 2026 model, 284B parameters with 13B active per token, is published on Hugging Face under an MIT licence for self-hosting and fine-tuning.
Peak and off-peak pricing
Off-peak and peak rates apply; weekday-morning UTC peak hours cost twice as much as every other hour.
Cheaper than 75% of the 107 GA models with a published price, mid-pack on SWE-bench Verified (rank 17 of 40), and one of 42 that document a zero-data-retention option. Ranked against GA models; this record is not GA.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $0.262/M | #24 / 107 | peer median $1.71/M |
| Output speedtokens/s | 94 tok/s | #31 / 60 | peer median 94 tok/s |
| SWE-bench Verified% solved | 79% | #17 / 40 | peer median 76.8% |
| GPQA Diamond% correct | 88.1% | #32 / 64 | peer median 87.9% |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| DeepSeek | 1M | $0.15 / $0.60 | text → text | |
| DeepSeek | 1M | $0.15 / $0.60 | text, image → text | |
| DeepSeek | 1M | $1.32 / $3.96 | text → text | |
| Alibaba Cloud | 1M | $0.40 / $1.60 | text, image, video → text |
Training data
Pre-trained on 32 trillion tokens of curated web text, licensed datasets, and synthetic reasoning traces. Post-trained with two-stage pipeline: domain-expert SFT + GRPO, unified via on-policy distillation.
| Name | Endpoint | Auth |
|---|---|---|
| DeepSeek API (legacy name routes to V4.1 Flash) | https://api.deepseek.com | API Key |
Explicit reasoning budget in non-thinking mode
Set reasoning='high' or 'max' for complex tasks; use standard for routine queries.
Thinking modes add visible chain-of-thought; high/max improve accuracy on STEM by 3-5 points but cost 2-3x more tokens.
XML tags for structured input
<context>...</context><task>...</task>
Like Claude, V4 Flash follows XML structure more reliably than markdown for complex prompts.
Reasoning modes have different output styles
Workaround: Standard mode is concise; max mode is verbose. Test both for your use case.
Cache key resets if system prompt changes
Workaround: Fix system prompts early; changing them mid-deployment resets cache savings.
deepseek-v4-flash-20260424 · 24 Apr 2026
Initial V4 Flash release. 284B total and 13B active parameters, 1M context.
deepseek-v4.1-flash · 10 Sep 2026
Successor. V4 Flash retired the same day; legacy names routed to V4.1 Flash.
deepseek-v4.1-flash-1017 · 17 Oct 2026
V4.1 Flash upgraded to the 1017 build with API method and price unchanged.
Not as its own model. DeepSeek retired V4 Flash on 10 September 2026 when it released V4.1 Flash. The legacy name deepseek-v4-flash is still accepted, but it now runs the current Flash model at Flash rates, per DeepSeek's pricing page. The MIT-licensed weights remain on Hugging Face for self-hosting.
Billing follows the current Flash rate card. Input costs $0.15 per million tokens on a cache miss and a fraction of a cent on a cache hit, and output costs $0.60 per million, all at off-peak times. Peak pricing is double on weekday mornings UTC. DeepSeek's pricing page lists no free tier.
DeepSeek V4.1 Flash is the vendor's replacement, called as deepseek-flash. It keeps the 1M-token context and adds native image input. For more reasoning headroom at a higher price, DeepSeek V4 Pro is the larger sibling. Outside DeepSeek, Qwen3.7-Plus offers a similar long-context class with built-in vision.
DeepSeek's changelog says the old names are temporarily routed to V4.1 Flash after the retirement of V4 Flash. Outputs, latency and token usage can therefore differ from what you saw in the spring. Switch your configuration to deepseek-flash and re-run your own evaluation, because the vendor calls the routing temporary and may remove it.
Yes. DeepSeek's April 2026 weights are on Hugging Face and MIT-licensed, so you can self-host or fine-tune them without the API. It is a Mixture-of-Experts model with 13 billion of its 284 billion parameters active per token. Self-hosting is the only way to get the original April behaviour now that the API alias points at a newer model.
DeepSeek-V4 Flash was an open-weight Mixture-of-Experts language model that DeepSeek released on 24 April 2026 as the cheaper sibling of DeepSeek V4 Pro. DeepSeek retired it on 10 September 2026, when it shipped DeepSeek V4.1 Flash. The vendor's changelog says the old model names are temporarily routed to V4.1 Flash, and its pricing page (read on 9 November 2026) says requests to the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price. New projects should call deepseek-flash directly. This page is kept as a record of what V4 Flash was and how to migrate off it.
At launch the model had 284 billion total parameters with 13 billion active per token, a 1,048,576-token context window and a 384,000-token output cap, under an MIT licence. The model card on Hugging Face listed 79% on SWE-bench Verified, 88.1% on GPQA Diamond and 96.4% on HumanEval, and Artificial Analysis gave it an Intelligence Index of 47. Those scores belong to the retired April model and say nothing about what the alias serves today. The open weights are still on Hugging Face, so a team that pinned V4 Flash for reproducibility can still self-host it.
What changed for API users is the model behind the name and the price. The current Flash model takes image input natively, which V4 Flash did not, and the 17 October update (DeepSeek-V4.1-Flash-1017) kept the API method and price unchanged. Prices per million tokens are now $0.15 input on a cache miss, $0.003 on a cache hit and $0.60 output at off-peak rates, with peak rates exactly double. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. At launch the model cost $0.14 input and $0.28 output per million tokens, and DeepSeek's pricing page lists no free tier for the Flash model today.
The case for staying on the retired alias is thin. It protects nobody from change, because the name already points at a different model, and DeepSeek called the routing temporary. Anyone who needs the exact April behaviour has to self-host the weights. Teams that wanted a larger step up in reasoning should compare DeepSeek V4 Pro, and teams that need a different vendor can look at Qwen3.7-Plus or GPT-4o mini. The vision preview that shared this family, DeepSeek V4 Flash Vision Exp, was retired in the same announcement. Vendor details sit on the DeepSeek company page, and every retired model is listed under retired models.
Specifications on this page (context window, token prices, benchmark scores, licence) are read from DeepSeek's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.