AI 領域精選文章翻譯
by Simon Willison

原文連結: DeepSeek V4—almost on the frontier, a fraction of the price
2026 年 4 月 24 日
Chinese AI lab DeepSeek’s last model release was V3.2 (and V3.2 Speciale) last December. They just dropped the first of their hotly anticipated V4 series in the shape of two preview models, DeepSeek-V4-Pro and DeepSeek-V4-Flash.
DeepSeek 中文 AI 實驗室上一次發布模型還是 V3.2(以及 V3.2 Speciale),時間要追溯到去年 12 月。這次他們一口氣丟出備受期待的 V4 系列第一波預覽模型:DeepSeek-V4-Pro 與 DeepSeek-V4-Flash。
Both models are 1 million token context Mixture of Experts. Pro is 1.6T total parameters, 49B active. Flash is 284B total, 13B active. They’re using the standard MIT license.
兩個模型都支援 100 萬 token context,並採用 Mixture of Experts(MoE)架構。Pro 版共有 1.6T total parameters、49B active parameters;Flash 版則是 284B total、13B active。兩者都採用標準 MIT 授權。
I think this makes DeepSeek-V4-Pro the new largest open weights model. It’s larger than Kimi K2.6 (1.1T) and GLM-5.1 (754B) and more than twice the size of DeepSeek V3.2 (685B).
這讓我認為 DeepSeek-V4-Pro 已經是新的最大 open weights 模型。它比 Kimi K2.6(1.1T)和 GLM-5.1(754B)都大,甚至比 DeepSeek V3.2(685B)還大出兩倍多。
Pro is 865GB on Hugging Face, Flash is 160GB. I’m hoping that a lightly quantized Flash will run on my 128GB M5 MacBook Pro. It’s possible the Pro model may run on it if I can stream just the necessary active experts from disk.
Pro 在 Hugging Face 上是 865GB,Flash 是 160GB。我希望稍微量化過的 Flash 可以跑在 128GB 的 M5 MacBook Pro 上。若能只從磁碟串流必要的 active experts,Pro 也有可能在那台機器上跑起來。
For the moment I tried the models out via OpenRouter, using llm-openrouter:
目前我先透過 OpenRouter 測試這些模型,並使用 llm-openrouter:
llm install llm-openrouter
llm openrouter refresh
llm -m openrouter/deepseek/deepseek-v4-pro 'Generate an SVG of a pelican riding a bicycle'
Here’s the pelican for DeepSeek-V4-Flash:
這是 DeepSeek-V4-Flash 產生的鵜鶘:連結:
And for DeepSeek-V4-Pro:
這是 DeepSeek-V4-Pro 產生的版本:連結:
For comparison, take a look at the pelicans I got from DeepSeek V3.2 in December, V3.1 in August, and V3-0324 in March 2025.
拿來對照的話,可以看看我在 去年 12 月的 DeepSeek V3.2、去年 8 月的 V3.1,以及 2025 年 3 月的 V3-0324 產生的鵜鶘。
So the pelicans are pretty good, but what’s really notable here is the cost. DeepSeek V4 is a very, very inexpensive model.
所以鵜鶘表現其實相當不錯,但真正值得注意的是成本。DeepSeek V4 是一個非常非常便宜的模型。
This is DeepSeek’s pricing page. They’re charging $0.14/million tokens input and $0.28/million tokens output for Flash, and $1.74/million input and $3.48/million output for Pro.
這是 DeepSeek 的定價頁。Flash 的價格是每百萬 input tokens $0.14、每百萬 output tokens $0.28;Pro 則是每百萬 input tokens $1.74、每百萬 output tokens $3.48。
Here’s a comparison table with the frontier models from Gemini, OpenAI and Anthropic:
下面這張表是拿 Gemini、OpenAI 與 Anthropic 的前沿模型來對照:
| Model | Input ($/M) | Output ($/M) |
|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 |
| GPT-5.4 Nano | $0.20 | $1.25 |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 |
| Gemini 3 Flash Preview | $0.50 | $3 |
| GPT-5.4 Mini | $0.75 | $4.50 |
| Claude Haiku 4.5 | $1 | $5 |
| DeepSeek V4 Pro | $1.74 | $3.48 |
| Gemini 3.1 Pro | $2 | $12 |
| GPT-5.4 | $2.50 | $15 |
| Claude Sonnet 4.6 | $3 | $15 |
| Claude Opus 4.7 | $5 | $25 |
| GPT-5.5 | $5 | $30 |
DeepSeek-V4-Flash is the cheapest of the small models, beating even OpenAI’s GPT-5.4 Nano. DeepSeek V4 Pro is the cheapest of the larger frontier models.
DeepSeek-V4-Flash 是小型模型裡最便宜的一個,連 OpenAI 的 GPT-5.4 Nano 都被它壓過。DeepSeek V4 Pro 則是大型前沿模型裡最便宜的選項。
This note from the DeepSeek paper helps explain why they can price these models so low—they’ve focused a great deal on efficiency with this release, especially for longer context prompts:
DeepSeek 論文 裡有一段註記,能解釋為什麼他們可以把價格壓得這麼低——這次發表特別強調效率,尤其是長上下文提示:
In the scenario of 1M-token context, even DeepSeek-V4-Pro, which has a larger number of activated parameters, attains only 27% of the single-token FLOPs (measured in equivalent FP8 FLOPs) and 10% of the KV cache size relative to DeepSeek-V3.2. Furthermore, DeepSeek-V4-Flash, with its smaller number of activated parameters, pushes efficiency even further: in the 1M-token context setting, it achieves only 10% of the single-token FLOPs and 7% of the KV cache size compared with DeepSeek-V3.2.
在 100 萬 token context 的情境下,即使是 DeepSeek-V4-Pro,雖然啟用的參數更多,單 token FLOPs(以等效 FP8 FLOPs 計)也只相當於 DeepSeek-V3.2 的 27%,KV cache 也只需要 10%。而 DeepSeek-V4-Flash 因為啟用參數更少,效率又更進一步:在 100 萬 token context 設定下,單 token FLOPs 只有 DeepSeek-V3.2 的 10%,KV cache 只需要 7%。
DeepSeek’s self-reported benchmarks in their paper show their Pro model competitive with those other frontier models, albeit with this note:
DeepSeek 在論文中自行揭露的 benchmark 同樣收錄在這份 paper 裡,顯示 Pro 版可以和其他前沿模型競爭,不過也附了一段註解:
Through the expansion of reasoning tokens, DeepSeek-V4-Pro-Max demonstrates superior performance relative to GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks. Nevertheless, its performance falls marginally short of GPT-5.4 and Gemini-3.1-Pro, suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately 3 to 6 months.
透過擴展 reasoning tokens,DeepSeek-V4-Pro-Max 在標準推理 benchmark 上相較於 GPT-5.2 與 Gemini-3.0-Pro 展現出更高表現。不過,它的表現仍略低於 GPT-5.4 與 Gemini-3.1-Pro,表示其發展軌跡大約落後目前最先進的前沿模型 3 到 6 個月。
I’m keeping an eye on huggingface.co/unsloth/models as I expect the Unsloth team will have a set of quantized versions out pretty soon. It’s going to be very interesting to see how well that Flash model runs on my own machine.
我會持續關注 huggingface.co/unsloth/models,因為我預期 Unsloth 團隊很快就會釋出一批量化版本。那個 Flash 模型在我的機器上到底能跑多好,會非常值得觀察。
DeepSeek V4 這篇最有殺傷力的地方,不是它又多大,而是它把「前沿模型一定很貴」這件事打出裂痕。當 1M context、MoE、接近前沿的表現,能被壓到這種價格,接下來比的不只是哪家最強,而是誰能用最低的成本把能力穩定塞進產品。
這也會逼很多團隊重新算帳。以前只要把模型當黑盒接上去就好,現在得開始思考:長上下文值不值得、active parameters 要不要更克制、推論流水線能不能更省。模型競爭不再只是 benchmark 對決,而是整條系統誰比較會省。
更現實的是,這種價格會改變使用者的期待。當便宜模型已經夠好,大家就不會再容忍昂貴模型只是「大一點點更準」;真正的門檻會從能力差距,轉成整體效率與可部署性。