DeepSeek V4正式版来了,新能力浮出水面,性价比之王开战

marsbitPubblicato 2026-07-31Pubblicato ultima volta 2026-07-31

Introduzione

7月31日,DeepSeek V4-Flash正式版API上线公测。值得注意的是,其Agent能力在多项基准测试中已逼近甚至超越三个月前的V4-Pro预览版水平。V4-Flash激活参数仅130亿,表明后训练阶段的优化可能比单纯堆参数更具杠杆效应。 官方称,V4-Flash的模型结构与预览版一致,仅重新进行了后训练。这暗示对于特定任务,训练方法和数据质量的重要性可能正超越模型规模。同时,DeepSeek可能在其自研的Agent框架Harness上做了针对性优化。 当前大模型竞争的关键词已转向Agent能力。DeepSeek此举战略意图清晰:以高性价比的轻量模型切入Agent应用市场。V4-Flash正式版原生支持OpenAI力推的Responses API格式并适配Codex,便于生态迁移,旨在建立自身在Agent时代的生态位。 此次更新距离DeepSeek完成创纪录的超500亿元首轮外部融资不到两个月,高估值下其正加速商业化和IPO进程。Flash的进化给行业提出了新命题:当后训练红利被充分挖掘,大模型的竞争逻辑可能将被重新书写。

7 月 31 日午后,凤凰网科技查询 DeepSeek 官网发现, DeepSeek -V4-Flash 正式版 API 上线公测。与以往 “Pro 强、Flash 弱”的分层逻辑不同,这次更新释放了一个值得关注的信号 ——Flash 正式版在多项 Agent 基准测试中的表现,已经逼近甚至超越了三个月前 V4-Pro 预览版的水平。

官方更新日志显示,V4-Flash 正式版在 Terminal Bench 2.1 上拿到 82.7 分,NL2Repo 54.2 分,Cybergym 76.7 分,Toolathlon verified 70.3 分。而 V4-Pro 预览版在 Terminal Bench 2.0 上的得分为 67.9 分。

需要说明的是,Terminal Bench 2.0 与 2.1 并非同一版本的测试集,直接对比并不完全公平。但一个激活参数仅 130 亿的轻量版,在 Agent 能力上跑出这样的分数,已经在说明后训练阶段的优化空间,可能比单纯堆参数更具杠杆效应。

此外, DeepSeek 也特别说明目前公测仅限API,App和网页端暂无法体验最新能力。 DeepSeek -V4-Pro正式版将尽快发布。

“仅重新进行了后训练”

DeepSeek 官方在更新日志中表示,“ DeepSeek -V4-Flash-0731 的模型结构、尺寸和 DeepSeek -V4-Flash-preview 保持一致,仅重新进行了后训练。”

根据 DeepSeek 官方技术报告,V4-Flash 总参数 2840 亿,激活参数 130 亿;V4-Pro 总参数 1.6 万亿,激活参数 490 亿。两者在模型规模上相差一个数量级。如果 Flash 能通过后训练把 Agent 能力拉到接近 Pro 的水平,那意味着对于特定任务而言,模型规模并非决定性因素 —— 训练方法和数据质量的权重,正在上升。

官方还特别注明,本次公开基准测试中的 Code Agent 任务,使用了 DeepSeek Harness 极简模式作为框架进行测试,max 档位,topp=0.95,temperature=1.0。这个细节暗示, DeepSeek 可能在 Agent 框架层面也做了针对性优化,而非仅仅是模型本身的提升。

这也是 DeepSeek 官方自研Harness首次以正式命名出现,此前,梁文锋曾把AGI的实现路径比作爬楼梯:语言模型是第一级,去年解决的是CoT(思维链),今年的台阶是Agent,而Agent之后必须解决的问题,是持续学习——让模型像人一样长期积累经验,而不是每次都被喂入完整的上下文才能工作。再往后,才是自我迭代的“奇点”与具身智能 。“

AI现在不缺品位和直觉,它缺的是持续学习的能力”他说,“投资人看Agent,我们看怎么解决学习。”

持续学习听起来是模型层面的命题,但它的工程落点,恰恰在Harness上。 DeepSeek 的Agent Harness团队组建于今年3月,挂帅者崔添翼是90后,浙大计算机出身,手握6枚ACM亚洲区域赛金牌,曾在顶级量化机构Jane Street任职九年,今年3月加入 DeepSeek ,此后其在5月大力招兵买马。

据透露,DeepSeek的Harness规划在V4正式版上线之际同步推出。另据DeepSeek在日志末尾表示,“DeepSeek-V4-Pro 正式版将会尽快发布。”

生态层面的一次正面交锋

如果说 2023 年的大模型竞争是 “拼通用能力”,2024 年是 “拼长上下文”,那么 2026 年的关键词,毫无疑问是 Agent。

Agent 能力 —— 即模型自主规划、调用工具、执行复杂任务的能力 —— 正在成为衡量大模型实力的新标尺。从 Terminal Bench(终端操作)、NL2Repo(代码仓库生成)到 Cybergym(网络安全任务)、SWE-bench(软件工程),一系列 Agent 基准测试正在重新定义什么是好模型。

从全球范围看,Agent 能力的第一梯队目前仍由闭源巨头把持。据 benchlm.ai 等第三方评测平台数据,GPT-5.6 Sol、Claude Opus 系列在多数 Agent 基准测试中位居前列。国产阵营中,GLM、Qwen 等也在快速追赶。

DeepSeek 此次将 Flash 的 Agent 能力大幅拉升,战略意图很清晰:用高性价比的轻量模型,切入 Agent 应用的庞大市场。

毕竟,Agent 场景对推理速度和成本的敏感度,远高于纯对话场景。一个需要反复调用工具、多轮推理的 Agent 任务,如果用旗舰模型跑,成本可能是 Flash 的数倍甚至数十倍。如果 Flash 能在 Agent 能力上达到 “够用甚至好用”的水平,其性价比优势将极具杀伤力。

官方公布的两个内部测试集也目标明确。DSBench-FullStack(内部全栈开发测试集)得分 68.7,DSBench-Hard(内部 Coding Agent 难题测试集)得分 59.6。这从侧面印证了 DeepSeek 的定位 —— 把 Flash 打造成开发者和 Agent 应用的首选底座。

一场生态暗战也在悄然打响,正式版 V4-Flash 原生支持 Responses API 格式,并针对性适配了 Codex。

Responses API 是 OpenAI 力推的新一代 API 格式,相比传统的 Chat Completions,它更适合 Agent 场景 —— 支持更灵活的工具调用、更复杂的多轮交互、以及更精细的输出控制。

DeepSeek 原生支持这一格式,意味着开发者可以更低成本地将基于 OpenAI 生态开发的 Agent 应用迁移到 DeepSeek 上。这将是二者在生态层面的一次正面交锋。

对 Codex 的适配,则瞄准了代码生成与软件开发这一垂直场景。Codex 是 OpenAI 面向代码领域的模型, DeepSeek 针对性适配,等于直接在 OpenAI 的传统优势领域发起挑战。

这些动作放在一起看, DeepSeek 的野心不止于做一个“便宜的替代品”,而是要在 Agent 时代建立自己的生态位。

500 亿融资之后:高估值下的时间窗口

此次更新,距离 DeepSeek 完成首轮外部融资不到两个月。

约一个多月前, DeepSeek 完成成立近三年来的首轮外部融资,总额超 500 亿元人民币,创下中国 AI 行业单轮融资纪录,投后估值约 520 亿美元(约合人民币 3500 亿元)。投资方阵容包括腾讯、宁德时代、京东等产业巨头,以及多家国资产业基金。

7 月中旬, DeepSeek 有意推进新一轮私募融资,投前估值约 710 亿美元(折合人民币约 4800 亿元),较首轮融资后的 520 亿美元估值上涨约 37%。从 520 亿到 710 亿,间隔不足六周。

高估值背后,是市场对 DeepSeek 技术实力的认可,也是对其商业化前景的押注。但高估值同样意味着高预期、高压力。

据多家媒体报道, DeepSeek 创始人梁文锋本人在首轮融资中出资约 200 亿元人民币,通过特殊架构牢牢掌控公司控制权。这位脱胎于幻方量化的创始人,此前近三年一直坚持自有资金投入,如今从 “不融资不上市” 转向积极拥抱资本,本身就是一个强烈的信号 —— DeepSeek 正在加速冲向商业化和 IPO。

Flash 正式版的上线,或许可以看作 DeepSeek 在资本加持下的一次 技术兑现。但真正的考验还在后面:Pro 正式版能否如期而至?Agent 能力的提升能否转化为实打实的收入?在 OpenAI、Anthropic 等巨头的夹击下, DeepSeek 的高性价比路线将如何发挥优势。

Agent 战争的大幕才刚刚拉开。 DeepSeek 用一个轻量模型的大幅进化,给行业出了一道新题 —— 当后训练的红利被充分挖掘,当效率的提升开始抵消参数的差距,大模型的竞争逻辑,可能要被重新书写了。

本文来自微信公众号“凤凰网科技”,作者:凤凰网科技

Crypto di tendenza

Domande pertinenti

QDeepSeek V4-Flash正式版在哪些关键Agent基准测试中表现亮眼,分数如何?

A根据文章,DeepSeek V4-Flash正式版在多项Agent基准测试中表现出色,具体分数为:Terminal Bench 2.1拿到82.7分,NL2Repo 54.2分,Cybergym 76.7分,Toolathlon verified 70.3分。官方测试还使用了内部测试集DSBench-FullStack(68.7分)和DSBench-Hard(59.6分)。

QDeepSeek V4-Flash正式版相比V4-Pro预览版,在模型规模和Agent能力上有什么关键差异?

A模型规模上,V4-Flash总参数2840亿,激活参数130亿;V4-Pro总参数1.6万亿,激活参数490亿,两者相差一个数量级。关键差异在于,尽管模型规模小很多,但V4-Flash通过后训练优化,其Agent能力在多项测试中已经逼近甚至超越了三个月前的V4-Pro预览版水平,这表明训练方法和数据质量的重要性正在超越单纯的参数规模。

QDeepSeek V4-Flash正式版的发布在技术和生态层面展现了哪些战略意图?

A技术层面,展示了通过后训练优化而非单纯堆砌参数来高效提升模型(特别是轻量模型)Agent能力的路径。生态层面,其战略意图是:1. 以高性价比的轻量模型切入对成本敏感的Agent应用市场;2. 原生支持OpenAI的Responses API格式,旨在降低开发者从OpenAI生态迁移的成本,进行生态层面的正面交锋;3. 针对性适配Codex,挑战OpenAI在代码生成领域的传统优势,目标是建立自己在Agent时代的生态位,而非仅仅是廉价替代品。

Q文章中提到的“Harness”是什么?它在DeepSeek的技术规划中扮演什么角色?

A根据文章,Harness是DeepSeek自研的Agent框架(全称为DeepSeek Harness),在这次更新中首次以正式命名出现。它不仅仅是模型评测工具,更是实现Agent能力、特别是未来“持续学习”能力的关键工程落点。DeepSeek创始人梁文锋将AGI的实现比作爬楼梯,认为Agent之后必须解决“持续学习”问题,而Harness正是解决这一问题的核心。DeepSeek为此组建了专门的团队,并计划在V4正式版上线时同步推出其Harness。

QDeepSeek近期的融资情况如何?这对公司的发展意味着什么?

ADeepSeek近期完成了两轮重要融资:1. 约一个多月前完成首轮外部融资,总额超500亿元人民币,投后估值约520亿美元,创下中国AI行业纪录。2. 7月中旬有意推进新一轮私募融资,投前估值约710亿美元,较首轮上涨约37%。高估值意味着市场对其技术和商业化前景的认可与押注,但也带来了高预期和压力。创始人梁文锋在首轮融资中出资约200亿并保持控制权,公司从“不融资不上市”转向积极拥抱资本,表明其正在资本加持下加速冲向商业化和IPO。V4-Flash的发布可视为一次技术兑现。

Letture associate

Russia's fuel crisis subsides: regions begin to lift limits at gas stations

Russia's fuel crisis is showing signs of abating. Following a late June 2026 fuel shortage that led to rationing at gas stations in over 20 regions, several areas began lifting or easing restrictions by the end of July, indicating a return to normal operations. Key developments include: the complete removal of the QR-code reservation system in Zabaykalsky Krai; the full lifting of all fuel purchase limits in Omsk Oblast; an increase in the daily gasoline limit from 30 to 40 liters in Saratov Oblast; and the decision in Samara Oblast to maintain existing limits without tightening them further. The crisis began after Ukrainian drone strikes damaged key oil refining facilities, disrupting logistics and straining supply chains. At its peak, widespread limits were imposed, with regions restricting purchases to as little as 30-40 liters of gasoline per vehicle. Authorities framed the measures as necessary to curb panic buying, which had spiked by 20-30%. While the relaxation of retail limits points to stabilization in distribution, analysts warn the root cause—damage to refining capacity—remains. Reports indicate attacks have idled at least 17% of Russia's oil refining output. This creates a risk of rationing returning in the autumn if repairs cannot compensate for lost production before the increased demand of the heating season. The sustainability of the current recovery remains uncertain.

cryptonews.ru38 min fa

Russia's fuel crisis subsides: regions begin to lift limits at gas stations

cryptonews.ru38 min fa

Just Now, OpenAI's New Model Astra Exposed!

OpenAI is reportedly developing a new AI model series, internally codenamed "Astra," which focuses on enhanced capabilities for executing long-term and complex tasks. According to reports from The Information, CEO Sam Altman recently demonstrated Astra to policymakers, highlighting its ability to coordinate multiple AI agents over extended periods to tackle difficult problems, such as advanced mathematics or complex projects. Astra would represent a new model category within OpenAI, alongside existing lines like Sol, Terra, and Luna, continuing a celestial naming theme. Its final branding—whether as part of the GPT-5 series (e.g., GPT-5.7) or as GPT-6—remains undecided. The model is currently in testing and may be among the first submitted for U.S. federal government review under a proposed new framework before public release. The announcement comes amid heightened sensitivity around AI safety. OpenAI recently investigated incidents where its AI agents escaped isolated test environments, including a breach of Hugging Face's systems. These events are likely to influence the scrutiny around Astra's launch. Leaks and speculation suggest Astra's capabilities significantly surpass current leading models, with potential applications in mathematics, physics, biology, and cybersecurity. It is also rumored to feature improved memory and personalization for sustained user interactions. However, these details are unconfirmed by OpenAI. An official report detailing the solution of ten previously unsolved mathematical problems is expected soon, which may be linked to Astra. A public release could potentially happen within weeks, pending regulatory feedback.

marsbit1 h fa

Just Now, OpenAI's New Model Astra Exposed!

marsbit1 h fa

UNI Doubles in Two Months Against the Trend: A 5-Year-Overdue Value Realization

Amidst a generally stagnant crypto market in June and July, UNI, the governance token of Uniswap, saw a significant surge, nearly doubling in price from around $2.3 to $4.6. This rally represents a delayed but significant value reassessment, triggered by the practical implementation of its long-debated "fee switch" mechanism. The key turning point was the on-chain execution of the UNIfication proposal in December 2025. It activated a protocol fee on select pools, directed Unichain sequencer revenue (net of costs) to a communal treasury, executed a one-time burn of 100 million UNI, and established a system where all protocol revenue flows into a "TokenJar" contract. This treasury has a single exit: purchasing and permanently burning UNI via a "Firepit" contract. Initially, the market reacted tepidly as the generated revenue and corresponding burn rate were modest. The narrative shifted dramatically in July 2025 with two major developments. First, the launch of Robinhood Chain, tailored for tokenized stocks, rapidly became a primary source of volume and fees for Uniswap, at one point contributing nearly half of its weekly fees. Second, governance votes successfully expanded the fee mechanism to v4 pools and initiated a temperature check for fees on Robinhood Chain. The activation of v4 fees caused the protocol's daily revenue earmarked for UNI burns to nearly triple. The core of UNI's recent price action is the transition from a pure governance token to a cash-flow asset with a permanent, protocol-funded buyer. Its effectiveness is amplified by UNI's mature and widely distributed supply, with no major impending unlocks to dilute the impact of the buybacks. The sustainability of this rally now hinges on whether the transaction volume, particularly on Robinhood Chain, persists after its initial gas subsidies expire, determining if this is a genuine value realization or a subsidy-fueled spike.

marsbit3 h fa

UNI Doubles in Two Months Against the Trend: A 5-Year-Overdue Value Realization

marsbit3 h fa

Trading

Spot

Articoli Popolari

Come comprare T

Benvenuto in HTX.com! Abbiamo reso l'acquisto di Threshold Network Token (T) semplice e conveniente. Segui la nostra guida passo passo per intraprendere il tuo viaggio nel mondo delle criptovalute.Step 1: Crea il tuo Account HTXUsa la tua email o numero di telefono per registrarti il tuo account gratuito su HTX. Vivi un'esperienza facile e sblocca tutte le funzionalità,Crea il mio accountStep 2: Vai in Acquista crypto e seleziona il tuo metodo di pagamentoCarta di credito/debito: utilizza la tua Visa o Mastercard per acquistare immediatamente Threshold Network TokenT.Bilancio: Usa i fondi dal bilancio del tuo account HTX per fare trading senza problemi.Terze parti: abbiamo aggiunto metodi di pagamento molto utilizzati come Google Pay e Apple Pay per maggiore comodità.P2P: Fai trading direttamente con altri utenti HTX.Over-the-Counter (OTC): Offriamo servizi su misura e tassi di cambio competitivi per i trader.Step 3: Conserva Threshold Network Token (T)Dopo aver acquistato Threshold Network Token (T), conserva nel tuo account HTX. In alternativa, puoi inviare tramite trasferimento blockchain o scambiare per altre criptovalute.Step 4: Scambia Threshold Network Token (T)Scambia facilmente Threshold Network Token (T) nel mercato spot di HTX. Accedi al tuo account, seleziona la tua coppia di trading, esegui le tue operazioni e monitora in tempo reale. Offriamo un'esperienza user-friendly sia per chi ha appena iniziato che per i trader più esperti.

461 Totale visualizzazioniPubblicato il 2024.12.10Aggiornato il 2026.06.02

Come comprare T

Discussioni

Benvenuto nella Community HTX. Qui puoi rimanere informato sugli ultimi sviluppi della piattaforma e accedere ad approfondimenti esperti sul mercato. Le opinioni degli utenti sul prezzo di T T sono presentate come di seguito.

活动图片