
Hy4 flash Vision:74.8 MMMU-Pro,以及 52.0 为何依然成立
- Orca新OrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 每百万 tokens · 184 tok/s
- tencent新Tencent: Hy4 Flash2026-11-16$0.02 / $0.07 每百万 tokens · 675 tok/s
- Alibaba新Qwen: Qwen4 Max2026-11-16$1.50 / $4.50 每百万 tokens · 385 tok/s
- Google新Google: Gemini 4 Carbon2026-11-10$2.00 / $10.00 每百万 tokens · 87 tok/s
- Anthropic新Anthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 每百万 tokens · 150 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 每百万 tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 每百万 tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 每百万 tokens · 84 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 每百万 tokens · 121 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 每百万 tokens · 107 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 每百万 tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 每百万 tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 每百万 tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447智能
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 每百万 tokens · 213 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929智能
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 每百万 tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733智能
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 每百万 tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740智能
On 20 November 2026, an open pull request against SGLang proposed vision-language support for Hunyuan-4-Flash — the framework's name for Hy4 flash, the 220-billion-parameter model Tencent put on its API on 16 November and opened as weights a day later. Buried in the PR description, under a heading called "Accuracy / Tests", is the first MMMU-Pro number anyone has published for an instruct build of the model: 74.78%, averaged over three runs on the first 300 examples. Tencent's own model card reports 52.0 on MMMU-Pro for Hy4 flash. The two figures are 22.8 points apart and they should not be subtracted from each other, because they do not describe the same thing. Working out what each one does describe is most of what a reader needs before deciding whether Tencent Hy4 stays the model to call when an image is involved.
拉取请求实际上包含的内容
该 PR 是拆分早先一项综合变更的三个 PR 之一,并且明确堆叠在文本模型那一项之上。截至 11 月 21 日,它仍处于打开、未合并状态,包含一个提交,涉及八个文件(+1,883 / −13),三个 CI 槽位均显示“Skipped — add the run-ci label”。尚未完成任何代码所有者审查。它所基于的文本支持也仍处于打开状态。因此,这里没有任何内容已随已发布的 SGLang 构建发布;Hy4 flash 的 cookbook 条目仍然只描述文本服务,并明确表示该支持“尚未进入已发布的 SGLang wheel”。
这个 diff 要做的是把视觉塔端到端地接入服务链路。其中真正关键的部分有:
• ViT 及其投影器采用注意力张量并行进行分片,因此在多 GPU 节点上,图像塔不会成为单卡瓶颈。
• 多模态位置编码是交错式 MRoPE——按槽位给出 (t, h, w) 位置——并同时传入注意力层和 IndexPool rope;稀疏层用后者来选择上下文,而这正是如果只转换稠密路径就会悄然失效的部分。
• Engram 哈希被训练为还原位于填充值之下的图像和视频 token id,并跳过对它们进行哈希,从而使 n-gram 记忆表不会像对待文本那样消耗媒体 token。
• `language_model_only` 标志让同一引擎无需启用多模态路径即可服务文本模型。
最后这个细节才是对当前状况的诚实概括:代码的写法让视觉能力成为可选项,因为对于 12B-active 模型的大多数部署场景来说,它确实会是可选项。

为什么74.8不算22个点的跃升
公关稿中的数字与卡片上的数字之间有三处差异,而每一处都让这个数字朝着更有利的方向变动。
The first is the checkpoint. Tencent's card reports multimodal scores for base models only, and says so in as many words: "Multimodal results are reported only for Hy4 flash Base and Hy4 Base." Hy4 flash Base scores 52.0 on MMMU-Pro (Standard, 10 options, 4-shot); the larger Hy4 Base scores 59.6. The instruct checkpoints — the ones anyone actually deploys — have no published multimodal score at all. The PR measures the instruct model, so it is filling a genuine gap rather than beating a number it was compared against.
The second is the elicitation. The PR runs at temperature 0.9 and top-p 1.0 with reasoning_effort=high and a 32,768-token completion budget. Tencent's base evaluation is 4-shot, no chain of thought. On a ten-option multiple-choice task, an instruct model allowed to reason before answering will beat a 4-shot base model almost by construction; that gap is the post-training, not a vision improvement. Treat 74.8 as "the instruct model, asked properly" and 52.0 as "the base model, asked in the standard way". They are two different measurements of two different artefacts.
第三点是子集问题。PR 评估的是前 300 个示例,而不是完整集合,也没有说明这 300 个示例是如何抽取的。在 300 个条目、比例接近 0.75 的情况下,仅采样标准误就约为 2.5 个百分点,因此在对任何问题顺序效应进行校正之前,仅针对这一个数字的 95% 区间大约就是 ±5 个百分点。PR 自己的三次运行结果分别为 74.33、75.33 和 74.67——相差一个百分点,这比采样误差更窄,而这一点值得了解:固定子集上的运行间方差会低估换一组不同的 300 个问题时这个数字会变动多少。同一 PR 中的第二种配置——带推测解码代码但不启用推测的完整栈——在同一子集上平均为 74.00,并在三次运行中显示出 5.3 个百分点的差异。这两个数字都没有错。但两者都不是采购决策所需要的那种意义上的基准测试结果。
这个 PR 确实能证明的、而且这部分是扎实的,是流水线接线正确,而不只是能编译。在加载视觉塔的情况下,GSM8K 的 200 条样本得分为 97.00%——与纯文本分支报告的结果相差不到一个百分点,因此多模态路径没有拖累语言头。覆盖形状和颜色、OCR、柱状图和线条计数的合成图像提示都得到了正确回答。这是一个通过了的冒烟测试,也是对 PR 应当作出的正确断言:代码能跑。准确率数字只是运行它带来的副产品。
![A screenshot of the SGLang pull request 51937 page on GitHub, showing the title '[Feat] models: Hunyuan-4-Flash vision-language support', a single commit from branch hy4f/4-vl into hy4f/3-text-model, eight files changed, and an Accuracy / Tests section listing three MMMU-Pro runs at 74.33, 75.33 and 74.67 percent on the first 300 examples plus GSM8K at 97.00 percent.](https://cms.orcarouter.ai/api/media/file/3-4342.png)
那些并不表示视频的视频 token
PR 标题写的是图像和视频输入,diff 中带有 video_start_token_id、video_end_token_id 和 video_token_id,而且其中一个被修改的 processor 会处理视频帧采样。据此把 Hy4 flash 归为视频模型会很容易。不要这么做。同一个 diff 里还有一条明确的防护判断——“HYV4 vision does not support video token compression”——并且空间压缩和时间压缩标志都被设为 false。厂商的模型卡从未声称支持视频:它说该模型“原生接受文本和图像输入,生成文本”。而已经在为该模型提供服务的 vLLM recipe 设置了 --limit-mm-per-prompt '{"image":1,"video":0}',即每个 prompt 一张图像、零视频。
因此,视频常量是为一条两端都封闭的路径所做的管道铺设:框架能够识别这些 token id,从而不会因为它们而出错,同时禁用了真正的视频实现所需的压缩。任何在 PR 摘要中读到“图像和视频输入”,并打算围绕一个 220B 开放权重模型规划视频流水线的人,读到的都是路线图,而不是一项能力。
你今天可以实际运行的内容
具体情况因运行时而异,这才是实际要点。
• vLLM —— 已支持视觉路径。已发布的方案将类名指定为 HYV4FlashVLForConditionalGeneration,并且默认允许每个提示词附带一张图像。这是目前已经存在的路径。
• SGLang — 文本服务可通过 lmsysorg/sglang:hy4-flash 镜像使用,而其 cookbook 页面在任何地方都没有提到视觉、图像或多模态。视觉是上面提到的开放 PR;只有在技术栈合并并发布 wheel 后它才可用,在此之前不可用。
• Tencent's own API — image input was part of the launch description, and our own catalogue lists tencent/hy4-flash as text+image→text with a 1,048,576-token context, 65,536 output tokens and $0.0175 per million input against $0.07025 per million output. Those are catalogue values read today, not a re-quote of the RMB card from launch week, and the two do not agree — the mainland rates convert to roughly four times the input price. We did not send an image call to verify the vision leg for this piece, and the disagreement between the two serving frameworks is a good reason to test it yourself before you depend on it.

测试而非假设的原因在于,一个模型可能被宣传为多模态,其权重也是多模态的,但你碰巧调用的端点却仍然只支持文本。这正是 Hy4 flash 过去五天所处的状态:视觉编码器随检查点一同发布,而服务栈一直在一个 PR 接一个 PR 地追赶。
这让选择处于何种境地
If you need image understanding from the Hunyuan line today and you need it to be a decided question rather than an integration project, the answer is still Tencent Hy4. It went generally available on 4 November with vision and computer use as headline capabilities, and its documented behaviour includes reading fine-grained charts and diagrams rather than the single-image-per-prompt arrangement the flash recipe defaults to. It is also several times the price: $0.834 input and $2.501 output per million on the same catalogue. Hy4 flash at $0.0175 is a commodity-priced model with a vision path that is real but newly wired.
The decision rule that follows is unglamorous. Route text and document work to Hy4 flash now. Route image work to Tencent Hy4 until either Tencent publishes instruct-model multimodal scores or the SGLang stack merges and a released wheel can be tested — whichever comes first, and neither has happened as of 21 November. There is no reason to pick one and abandon the other: both are on one key here, provider list prices pass through at 0% markup, and a routing rule can send image-bearing requests to the flagship while text goes to the cheap model without a second contract or a code change. Automatic failover covers the case where the newer vision path turns out to be the one that breaks under real images.
接下来看什么
有三件事会改变这篇文章,而每一件都是可核查的。
关于堆叠合并。vision PR 叠加在一个本身尚未合并的 text PR 之上;两者都需要 review,而且 CI 在两个 PR 上都还没跑过。在 wheel 发布之前,无论 diff 里写了什么,任何 SGLang 用户都无法 serve Hy4 flash 图像。
Tencent publishing instruct-model multimodal numbers. The card's multimodal table covers base checkpoints only, and that is an unusual omission for a launch that led with multimodality. A full-set MMMU-Pro row for Hy4 flash instruct, run by Tencent or by an independent evaluator, would replace every estimate in this article with a measurement.
Artificial Analysis adding the model. There is no Hy4 flash entry on the leaderboard as of 21 November — the model URL 404s — so there is no third-party index, speed or price-normalised score to check any of this against. Launch-week coverage of the model leaned on Tencent's own WorkIndex suite and an 85-person internal blind test; neither touches vision.
在其中一项落地之前,74.8 只能出现在附带其限定条件的句子里:一个 instruct 检查点、300 个条目、高推理强度、三次运行、一位框架作者对尚未审查的 PR 所做的抽查。这是 Hy4 flash 视觉方面现有的最好数字。这个门槛比听起来要低,而诚实的解读是,该模型的视觉能力现在已被明确证明得到实现,且在很大程度上仍未得到测量。
本文中的对比1
根据本文内容识别
