返回首页

对比阅读

对比阅读:Qwen3 Hits 6250 token/s on RTX 5090: Open Source Drops Inference Costs Another 50% 与 Qwen3 在 RTX 5090 跑出 6250 token/s — 开源社区把 AI 推理成本又压低 50%

AEN
Qwen3NvidiaOpen Source LLMs·

Qwen3 Hits 6250 token/s on RTX 5090: Open Source Drops Inference Costs Another 50%

A Reddit post shows: on the RTX 5090, a new compressed version of Qwen3 8B reaches 6250 tokens/s (roughly 6,250 characters generated per second) — 50% faster than traditional Q4 compression of the same size, and 4–7% faster than other competing NVFP4 compressed versions.

What this is

The poster is a member of the unsloth community (an open-source project specializing in model compression and fine-tuning), who released a new version of Qwen3 8B. "Quantization" here can be understood as "slimming down an AI model while preserving as much capability as possible" — the common approach is compressing model parameters from high-precision to low-precision numbers, making it run faster and use less VRAM.

What's special this time is NVFP4. This is Nvidia's new 4-bit floating-point format introduced on the latest Blackwell architecture (RTX 5090 being the consumer-facing representative), expressing numbers with fewer bits while maintaining sufficient precision — currently the newest generation of hardware-accelerated format. The post also mentions this version includes an "MTP speculative decoding head" — a technique that lets AI "guess a few characters ahead when generating text, and skip ahead if correct." With recommended settings, it can be roughly 15% faster still.

Worth noting: this is a community contribution, not an official release from Alibaba or Nvidia.

Industry view

Supporters see this as a key signal: the open-source community is responding to new hardware formats at remarkable speed, with high-quality solutions emerging shortly after Blackwell's launch. In the cost structure of model inference (running AI), "how to efficiently leverage new hardware" is becoming the new battleground. For the same input, higher hardware utilization means lower per-inference cost.

But there are reservations. First, this result is highly dependent on the specific RTX 5090 generation hardware — of limited relevance to users still on 30/40 series cards. Second, the benchmark data comes from community self-testing, lacking independent reproduction. Third, "50% speedup" typically picks a favorable baseline, so real-world business gains will be smaller.

More notably, there's another signal here: Qwen3 (Alibaba's Tongyi series) remains one of the Chinese large models most frequently "modded" by the global open-source community, showing that Chinese models have non-trivial acceptance among overseas developers — an underestimated fact.

Impact on regular people

For enterprise IT: the hardware barrier to local deployment of large models (running AI on your own servers rather than calling the cloud) keeps dropping. The same graphics card can run larger or faster models, and the cost structure of private deployment is being rewritten.

For working professionals: local AI assistants (offline, running on your own computer) will become increasingly usable, with more tool options for coding and document editing — but it still requires some technical skill to set up today.

For consumer markets: no direct impact in the short term, but long-term, the pricing pressure on cloud AI services comes from efficiency gains like this — cheaper inference means room for subscription price reductions.

BZH
Qwen3通义千问Nvidia·

Qwen3 在 RTX 5090 跑出 6250 token/s — 开源社区把 AI 推理成本又压低 50%

一条 Reddit 帖子显示:在 RTX 5090 上,一个新的 Qwen3 8B 压缩版本推理速度达到 6250 token/s(即每秒生成约 6250 个字),比同体积的传统 Q4 压缩版本快 50%,也比其他同类 NVFP4 压缩版本快 4-7%。

这是什么

发帖人是 unsloth 社区(专门做模型压缩和微调的开源项目)成员,发布了一个 Qwen3 8B 的新版本。这里的「量化」(quantization)可以理解为「把 AI 模型瘦身但尽量保留能力」——常见手段是把模型参数从高精度数字压成低精度,让它跑得更快、占用显存更小。

这次的特别之处在于 NVFP4。这是 Nvidia 在新一代 Blackwell 架构(RTX 5090 是消费级代表)上推出的 4 位浮点格式,用更少的比特数表达数字但保持足够精度,是目前最新一代的硬件加速格式。帖子还提到这个版本包含了「MTP 投机解码头」——一种让 AI 生成文本时「先猜几个字、猜对就跳过」的技术,按推荐设置还能再快约 15%。

需要说明的是,这是社区贡献,不是阿里或 Nvidia 的官方发布。

行业怎么看

支持方认为这是关键信号:开源社区对新硬件格式的跟进速度极快,Blackwell 推出后短期内就有高质量方案出现。模型推理(让 AI 跑起来算)的成本结构里,「如何高效利用新硬件」正在成为新竞争点。同样的输入,硬件利用率越高,单次推理成本越低。

但也有保留意见。一是这条成绩高度依赖 RTX 5090 这一代特定硬件,对仍在用 30/40 系显卡的用户意义有限;二是 benchmark 数据来自社区自测,缺少独立复现;三是「50% 加速」通常选取对自己有利的基线,真实业务场景下的提升幅度会打折扣。

更值得注意的是另一层信号:Qwen3(阿里通义系列)依然是全球开源社区最被频繁「魔改」的中文大模型之一,说明中国模型在海外开发者群体里接受度并不低,这是个被低估的事实。

对普通人的影响

对企业 IT:本地部署大模型(在自有服务器上跑 AI 而非调用云端)的硬件门槛继续降低,同一张显卡能跑的模型越来越大或越来越快,私有化部署的成本结构正在被改写。

对个人职场:本地 AI 助手(不联网、在自己电脑上跑)会越来越可用,写代码、改文档这类工具选择会更丰富,但目前仍需要一定技术能力才能用起来。

对消费市场:短期内对普通消费者没有直接感受,但长期看,云端 AI 服务的价格压力来自这类效率提升——推理变便宜,订阅费就有下调空间。