作者:失魂引 | 监制:罗超
适配AI3(HW3)平台的FSD V14 Lite来了,但特斯拉车主们并不买账。
日前特斯拉CEO马斯克在X平台宣布,正在向搭载AI3平台的车型推送FSD V14 Lite版。此次推送面向早期访问客户,预计几周后会向更多车主推送更新。
自从去年10月FSD V14完整版向搭载AI4(HW4)平台的车型推送以来,搭载AI3平台的车主就一直盼望着自己的车也能升级为功能更强大的FSD V14。
然而FSD V14 Lite的到来,并非所有特斯拉车主都喜闻乐见。甚至在马斯克的推文下,有不少网友正在攻击他和特斯拉。
10万元只配残血版FSD?模型蒸馏难平民愤据特斯拉官方介绍,FSD V14 Lite加入了强化学习(RL)和离线模型,改进了主动和被动响应能力,覆盖导航处理、合并和分叉、行人互动、交通信号灯以及车辆切入场景;减少了失误减速情况,转向更平顺、车道居中功能稳定;引入了完整的自动泊车、倒车功能;支持选择停放地点,例如停靠在停车场、街道、车道、路边等位置。
此外,FSD V14 Lite也加入了驾驶风格自定义功能,越用越像老司机。
简单来说,FSD V14 Lite用模型蒸馏让AI3平台车型用上AI4平台V14的核心能力,补齐泊车、到达选项、速度配置,并全面提升响应与舒适性,让老车体验接近新车,基本功能完全一致,只是对神经网络做了压缩处理,没有删减功能模块。
然而一位网友却表现得十分激动,在马斯克推文下连发多条评论,吐槽自己花了那么多钱只买到一个比十年前的承诺逊色许多的版本,并对马斯克口吐芬芳。
(图源:X平台截图)
此人表示,如果没有FSD V14满血版,他还能接受FSD V14 Lite与十年前的宣传存在差距,同样花了大价钱却不能使用相同的版本,令他无法接受。这位网友的激烈言论自然引发了争吵,多位马斯克粉丝与其激烈辩论。
另一位网友言语相对平和地表示,消费者花了15000美元(约合人民币10.2万元),不是为了买残血版FSD,而是为了真正的自动驾驶,也就是FSD。与上面一位网友相同,该网友也遭遇了马斯克粉丝的反驳。一位马斯克粉丝回复他,如果买了AI4硬件的车型,是不是也能得到终身升级权限,这些根本不可能实现。
(图源:X平台截图)
在电车通(ID:dianchetong233)看来,一位X平台网友的观点很客观。他表示,当初购买特斯拉汽车,是基于特斯拉能够实现完全自动驾驶的承诺,消费者需要特斯拉兑现承诺,而不是因为硬件能力不足就食言而肥。
与国内车企推出的智驾方案不同,特斯拉FSD价格高昂,售价曾在5000美元到15000美元之间多次调整。问题是许多早早支持特斯拉,购买FSD的车主,至今没能体验到特斯拉承诺的「完全自动驾驶」。
特斯拉AI5芯片已经完成流片,或许明年就能落地商用,未来还会有AI6、AI7等平台。如果今天接受特斯拉不完成承诺,那么AI4平台车型用户谁来保证,究竟要到哪一代平台才能真正用上特斯拉和马斯克承诺的完全自动驾驶?
算力焦虑正在吞噬老款硬件不是特斯拉不愿意为AI3平台适配FSD V14满血版,而是AI3平台已经难以满足智驾系统越来越高的硬件要求。
AI3平台算力仅144TOPS,内存仅8GB,带宽仅48GB/s,面对FSD V14满血版端到端大模型+强化学习+离线预测的海量数据吞吐,可能会出现数据喂不进、模型跑不动、延迟超标等问题。不仅如此,AI4平台传感器的性能也得到了提升,更高的分辨率、更大的动态范围,是FSD V14能力全面飞跃的基础。
特斯拉出于安全性考虑,才会放弃为AI3平台推送FSD V14满血版。即便是残血版,AI3平台能够升级至FSD V14 Lite,可未来是否还能升级到更高级的版本,加入其他新功能尤未可知。
(图源:电车通摄制)
国内也有类似的情况,小鹏汽车推出第二代VLA时便表示,只有搭载双图灵芯片的车型才能升级到满血版,单图灵芯片和双Orin-X芯片的车型只能升级为第二代VLA蒸馏版。
值得一提的是,小鹏部分车型零售版还提供算力高达2250TOPS的三图灵芯片版本,为未来升级预留了足够的算力冗余。小鹏GX的Robotaxi版本更是搭载4颗图灵芯片,算力高达3000TOPS。
现阶段国内不少车企智驾版本存在分级,如比亚迪的天神之眼A/B/C、乾崑智驾ADS SE/Pro/Max/Ultra等,这些版本在能力上有所区分,小鹏的Ultra SE和Ultra两个版本均适配第二代VLA,差距不在今天而在未来。
本质上,智驾从「规则驱动」转向「端到端大模型驱动」,算力需求呈指数级增长,L2+只需要数百TOPS算力,L4需要数千TOPS算力,而现有老款硬件的算力天花板,早已被新时代的智驾算法甩在身后。
(图源:电车通摄制)
该不该为更高的算力、未来的升级潜力付出更多成本,是当前消费者需要考虑的问题。
电车通(ID:dianchetong233)认为,在购车时我们要结合自己的预算、换车周期综合考虑,换车周期越长,越需要为升级空间做准备。今天的「满血算力」可能3年后就成为「基础配置」。预算充足时,选择更高算力能大幅延长车辆的智驾生命周期,避免出现「刚买就过时」的情况。
换车周期较短的消费者,则可根据自己的预算判断。另外,算力本身也会影响车辆的保值率,当L3时代到来,无法升级到L3的车型,势必面临残值暴跌的局面。
对于算力的焦虑,甚至可能影响到非刚需群体的购车选择,让他们放弃短期内购车,静待车企推出算力更高,升级潜力更大的产品。
为了践行自己的承诺,让用户可放心买车,今年4月马斯克表示将在美国搭建一批迷你工厂,用来将AI3平台车辆升级到AI4或更高版本的平台。这项升级计划不仅会改装芯片和带宽,就连摄像头也会一并升级。
对于售价20万元以上的车型而言,若能付出几万元升级硬件,换来更强的智驾功能,短期内不用因智驾体验落后换车,显然是值得的。只是已经付出了5000美元到15000美元的消费者,是否会愿意再付出数万元升级硬件仍是个问题。
如何打破“落地即落伍”的短命诅咒?相较于后期改装,预埋高算力无疑是更低成本且消费者接受度更高的方案。过去,阻碍这一方案实施的最大难点,其实是成本高昂的算力芯片。正因如此,近几年国内外企业纷纷启动自研智驾芯片计划。
海外市场,特斯拉AI3、AI4、AI5芯片持续迭代,国内市场小鹏推出了自研图灵芯片,蔚来推出了自研神玑NX9031芯片,理想则拿出了马赫100芯片。不久前,全球新能源汽车销冠比亚迪,重磅发布了国内首款自研4nm智驾芯片璇玑A3,预计明年由高端子品牌腾势首发搭载。
(图源:电车通摄制)
通过自研芯片,车企能够有效降低智驾芯片的采购成本,从而实现「算力自由」,并为车辆预埋高算力,增加车辆的升级空间。
适配AI3平台的FSD V14 Lite推送风波,看似是特斯拉老车主与品牌之间的信任争议,实则是智能汽车行业高速迭代下,软件无限OTA与硬件生命周期有限的必然矛盾。曾经的完全自动驾驶宣传承诺,在指数级增长的大模型算力需求面前,被硬件代差彻底割裂,这也是无数老车主心生不满、产生被辜负感的核心原因。
这场由FSD V14 Lite掀起的算力焦虑,也为所有消费者敲响了购车警钟。智驾汽车早已不是「即买即用」的传统工业品,而是具备持续迭代能力的智能终端。车企扎堆自研智驾芯片、迭代高阶硬件平台,本质上也是为了打破外部芯片成本桎梏,让高算力硬件逐步普及,解决用户「买车即落伍」的痛点。
车企唯有坚持硬件前置预埋、保留充足迭代冗余,才能兑现长期OTA升级承诺;而消费者也需摒弃「重软件、轻硬件」的购车思维,不再为过度营销的「完全自动驾驶」概念盲目买单,而是以硬件算力、迭代潜力、长期适配性为核心选购标准。
封面图源:电车通摄制
End解锁车主专属圈,同系车主,都在这里!入群畅聊爱车,欢迎大家扫码添加小助理拉群!
FSD V14 Lite for the AI3 (HW3) platform has arrived, but Tesla owners aren't having it.
Tesla CEO Elon Musk recently announced on X that he is rolling out FSD V14 Lite to vehicles equipped with the AI3 platform. The update is initially targeting early access customers, with a broader rollout expected in the coming weeks.
Since the full version of FSD V14 was rolled out to AI4 (HW4) vehicles last October, AI3 platform owners have been eagerly awaiting an upgrade to the more capable FSD V14.
However, not all Tesla owners are thrilled about the arrival of FSD V14 Lite. In fact, many netizens have been attacking Musk and Tesla under his post.
According to Tesla's official introduction, FSD V14 Lite incorporates reinforcement learning (RL) and offline models, improving active and passive response capabilities. It covers navigation handling, merging and splitting, pedestrian interaction, traffic lights, and cut-in scenarios; reduces erroneous braking, delivers smoother steering, and stabilizes lane centering; introduces full auto parking and reverse features; and supports selecting parking locations such as parking lots, streets, driveways, and curbsides.
Additionally, FSD V14 Lite also adds a customizable driving style feature that becomes more refined the more you use it.
Simply put, FSD V14 Lite uses model distillation to bring the core capabilities of the AI4 platform's V14 to AI3 vehicles,补齐 parking, arrival options, and speed settings, while comprehensively improving responsiveness and comfort to make older cars feel closer to new ones. Basic functionality is fully consistent, with only neural network compression applied—no feature modules have been removed.
However, one netizen became extremely agitated, posting multiple comments under Musk's post complaining that he spent so much money only to receive a version far inferior to a promise made ten years ago, and vented his frustration at Musk.
(Source: X platform screenshot)
The user stated that without the full version of FSD V14, he could accept the gap between FSD V14 Lite and the 10-year-old marketing claims. However, paying the same high price yet being unable to use the same version is unacceptable. His heated remarks naturally sparked arguments, with several Musk fans debating him intensely.
Another netizen expressed a relatively calmer view, saying consumers paid $15,000 (approximately 102,000 RMB) not for a nerfed version of FSD, but for true autonomy—FSD as advertised. Like the previous commenter, he also faced pushback from Musk fans. One fan replied that even if you buy an AI4 hardware vehicle, lifetime upgrade access is not guaranteed either.
(Source: X platform screenshot)
From the perspective of Dianche Tong (ID: dianchetong233), one X user's viewpoint is quite objective. They argued that when buying a Tesla, consumers did so based on the promise of full self-driving achievement, and Tesla needs to deliver on that promise rather than backing out due to insufficient hardware capabilities.
Unlike intelligent driving solutions launched by domestic automakers, Tesla FSD comes at a premium price, with costs fluctuating multiple times between $5,000 and $15,000. The problem is that many early Tesla supporters who purchased FSD still have not experienced the "full self-driving" that Tesla promised.
Tesla's AI5 chip has already completed tape-out and may go into commercial production next year, with AI6 and AI7 platforms likely to follow. If today's acceptance means Tesla won't honor its promise, who will guarantee AI4 platform owners what future awaits? Which generation of platform will truly deliver the full self-driving that Tesla and Musk have promised?
It is not that Tesla is unwilling to adapt the full version of FSD V14 for the AI3 platform, but rather that AI3 can no longer meet the increasingly demanding hardware requirements of intelligent driving systems.
The AI3 platform offers only 144 TOPS of computing power, 8GB of memory, and 48GB/s bandwidth. Faced with the massive data throughput required by FSD V14's full version—end-to-end large models plus reinforcement learning and offline prediction—issues such as data bottlenecks, model sluggishness, and latency spikes may occur. Moreover, the AI4 platform's sensors also see performance improvements, with higher resolution and wider dynamic range serving as the foundation for FSD V14's comprehensive leap.
Tesla chose to abandon the full version for AI3 out of safety considerations. Even though the nerfed Lite version can upgrade AI3 to FSD V14 Lite, whether future upgrades to more advanced versions with additional new features are possible remains uncertain.
(Source: photographed by Dianche Tong)
A similar situation exists in China. When XPeng launched its second-generation VLA, it stated that only vehicles equipped with dual Turing chips can upgrade to the full version, while single Turing chip and dual Orin-X chip vehicles can only upgrade to the distilled version of the second-generation VLA.
Notably, some XPeng retail models also offer a triple-Turing chip version with computing power up to 2,250 TOPS, leaving ample room for future upgrades. The XPeng GX Robotaxi version is even equipped with four Turing chips, delivering up to 3,000 TOPS of computing power.
At present, many domestic automakers offer tiered versions of their intelligent driving systems, such as BYD's Heavenly Eye A/B/C and AITO's ADS SE/Pro/Max/Ultra. These versions differ in capability, and XPeng's Ultra SE and Ultra both support the second-generation VLA—the gap lies not in today's capabilities but in future potential.
Essentially, intelligent driving has shifted from "rule-driven" to "end-to-end large model-driven," causing computing demands to grow exponentially. L2+ only requires hundreds of TOPS, while L4 needs thousands of TOPS. The computing ceiling of existing older hardware has long been left behind by new-generation intelligent driving algorithms.
(Source: photographed by Dianche Tong)
Whether to pay more for higher computing power and future upgrade potential is a question consumers need to consider.
Dianche Tong (ID: dianchetong233) believes that when purchasing a vehicle, consumers should综合考虑 their budget and replacement cycle. The longer the expected ownership period, the more they need to plan for upgrade space. Today's "full-spec computing power" may become "basic configuration" in three years. When budget allows, choosing higher computing power can significantly extend the vehicle's intelligent driving lifespan and avoid the "obsolete right after purchase" scenario.
For consumers with shorter replacement cycles, decisions can be made based on personal budgets. Additionally, computing power itself affects resale value—when L3 arrives, vehicles unable to upgrade to L3 will inevitably face steep depreciation.
Anxiety over computing power may even influence purchasing decisions of non-essential buyers, causing them to delay purchases and wait for products with higher computing power and greater upgrade potential.
To honor its promises and give users confidence in their purchases, Musk announced in April that Tesla would build a series of mini factories in the US to upgrade AI3 vehicles to AI4 or higher platforms. This upgrade plan would include replacing chips and bandwidth, as well as upgrading cameras.
For vehicles priced above 200,000 RMB, paying tens of thousands of yuan for a hardware upgrade to gain stronger intelligent driving capabilities and avoid replacing the car due to outdated ADAS features in the short term would clearly be worthwhile. The question remains whether consumers who already paid $5,000 to $15,000 would be willing to spend tens of thousands more on hardware upgrades.
Compared to later retrofitting, pre-installing high computing power is undoubtedly a lower-cost and more consumer-friendly solution. In the past, the biggest obstacle to implementing this approach was the prohibitively expensive computing chips. That is precisely why automakers at home and abroad have been launching in-house intelligent driving chip programs in recent years.
In overseas markets, Tesla's AI3, AI4, and AI5 chips continue to iterate. In China, XPeng has introduced its self-developed Turing chip, NIO has launched its self-developed Shenji NX9031 chip, and Li Auto has unveiled its Mach 100 chip. Recently, global new energy vehicle sales champion BYD announced its first domestically developed 4nm intelligent driving chip, Xuanji A3, expected to debut on its high-end sub-brand Denza next year.
(Source: photographed by Dianche Tong)
Through in-house chip development, automakers can effectively reduce intelligent driving chip procurement costs, achieving "computing power freedom" and enabling vehicles to be pre-equipped with high computing power, thereby expanding upgrade potential.
The FSD V14 Lite rollout controversy for the AI3 platform may seem like a trust dispute between Tesla's older owners and the brand, but it actually reflects the inevitable conflict under the rapid iteration of the smart automotive industry: software unlimited OTA versus limited hardware lifecycle. The once-promised full self-driving commitment has been completely severed from reality by the computing power gap across hardware generations, which is the core reason for countless older owners' dissatisfaction and sense of betrayal.
This computing power anxiety, ignited by FSD V14 Lite, has sounded an alarm for all consumers. Smart driving vehicles are no longer traditional industrial products that are "buy and use"—they are intelligent terminals capable of continuous iteration. Automakers' rush to develop in-house intelligent driving chips and iterate high-end hardware platforms is fundamentally aimed at breaking free from the cost constraints of external chips, making high-computing-power hardware more accessible, and solving users' pain point of "buying a car that is already outdated."
Only by insisting on hardware pre-installation and retaining sufficient iteration headroom can automakers fulfill their long-term OTA upgrade commitments. Consumers should also abandon the "software over hardware" purchasing mindset, stop blindly paying for over-marketed "full self-driving" concepts, and instead make hardware computing power, upgrade potential, and long-term adaptability the core criteria for their purchases.
Cover image source: photographed by Dianche Tong
#FSD #Tesla #Musk #BYD #XPeng
