A generated title card reading 'Hy4 flash vision: two checkpoints, two numbers', showing two cards side by side: 74.8 for the Hy4 flash instruct model on a 300-item MMMU-Pro slice from SGLang PR 51937, and 52.0 for the Hy4 flash base checkpoint on Tencent's model card, separated by a 'not a gain' divider.
Guides & Insights

Hy4 flash Vision: 74.8 MMMU-Pro, and Why 52.0 Still Stands

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

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 T​encent 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. T​encent'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 T​encent Hy4 stays the model to call when an image is involved.

What the pull request actually contains

The PR is one of three that split an earlier omnibus change, and it is explicitly stacked on top of the text-model one. As of 21 November it is open, unmerged, carries a single commit across eight files (+1,883 / −13), and all three CI slots read "Skipped — add the run-ci label". No code-owner review has been completed. The text support it builds on is also still open. So nothing here has shipped in a released SGLang build; the cookbook entry for Hy4 flash still describes text serving only and says outright that support "is not in a released SGLang wheel yet".

What the diff does is wire the vision tower into the serving path end to end. The pieces that matter:

• The ViT and its projector are sharded across attention tensor parallelism, so the image tower is not a single-rank bottleneck on a multi-GPU node.

• Multimodal position encoding is interleaved MRoPE — per-slot (t, h, w) positions — carried into both the attention layers and the IndexPool rope that the sparse layers use to select context, which is the part that would quietly break if only the dense path were converted.

• Engram hashing is taught to restore image and video token ids that sit under padding values and to skip hashing them, so the n-gram memory tables do not consume media tokens as if they were text.

• A language_model_only flag lets the same engine serve the text model without standing up the multimodal path at all.

That last detail is the honest summary of where this sits: the code is written so that vision is optional, because for most deployments of a 12B-active model it will be.

A generated scoreboard titled 'Hy4 flash vision — the scoreboard' listing six rows: MMMU-Pro instruct 74.8 unaudited on a 300-item slice; MMMU-Pro base checkpoint 52.0 four-shot; vision encoder SigLIP2-So400m with 27 layers; video input token ids only with compression off; vLLM vision serving shipping today; SGLang vision serving an open, unreviewed pull request, with a footer noting the two MMMU-Pro figures are different measurements and not comparable.

Why 74.8 is not a 22-point jump

Three things separate the PR's figure from the card's, and each one moves the number in the flattering direction.

The first is the checkpoint. T​encent'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. T​encent'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.

The third is the subset. The PR evaluates the first 300 examples, not the full set, and does not say how those 300 were drawn. At 300 items and a proportion near 0.75, the sampling standard error alone is about 2.5 points, so a 95% interval on that single figure is roughly ±5 points before any question-ordering effect. The PR's own three runs came in at 74.33, 75.33 and 74.67 — a spread of one point, which is narrower than the sampling error, and that is worth knowing: run-to-run variance on a fixed subset understates how much the number would move on a different 300 questions. A second configuration in the same PR, the full stack with the speculative-decoding code but no speculation, averages 74.00 on the same subset and shows a 5.3-point spread across its three runs. Neither figure is wrong. Neither is a benchmark result in the sense a procurement decision needs.

What the PR does establish, and this part is solid, is that the pipeline is wired correctly rather than merely compiling. GSM8K on 200 items scores 97.00% with the vision tower loaded — within a point of what the text-only branch reported, so the multimodal path is not degrading the language head. Synthetic image prompts covering shapes and colours, OCR, a bar chart and a line count were answered correctly. That is a smoke test that passed, and it is the right claim to make about a PR: the code works. The accuracy number is a side effect of running it.

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.

The video tokens that do not mean video

The PR title says image and video input, the diff carries video_start_token_id, video_end_token_id and video_token_id, and one of the modified processors handles video frame sampling. It would be easy to file Hy4 flash as a video model on that basis. Don't. The same diff contains an explicit guard — "HYV4 vision does not support video token compression" — with both spatial and temporal compression flags set to false. The vendor's card never claims video: it says the model "natively accepts text and image inputs, generates text". And the vLLM recipe that already serves the model sets --limit-mm-per-prompt '{"image":1,"video":0}', one image and no video per prompt.

So the video constants are plumbing for a path that is closed at both ends: the framework recognises the token ids so it does not choke on them, and disables the compression a real video implementation would need. Anyone reading "image and video input" in a PR summary and planning a video pipeline around a 220B open-weight model is reading a roadmap, not a capability.

What you can actually run today

The picture differs by runtime, which is the practical takeaway.

• vLLM — already serves the vision path. The published recipe names the class HYV4FlashVLForConditionalGeneration and allows one image per prompt by default. This is the route that exists now.

• SGLang — text serving exists through the lmsysorg/sglang:hy4-flash image, and the cookbook page for it contains no mention of vision, image or multimodal anywhere. Vision is the open PR above; it becomes available when the stack merges and a wheel ships, and not before.

• T​encent'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.

A screenshot of the OrcaRouter catalogue page for tencent/hy4-flash, showing the model described as the cost-efficient variant of Tencent Hunyuan Hy4 with a 1M-token context window, text and image input with text output, 65,536 maximum output tokens, and a published price of $0.02 input and $0.07 output per million tokens.

The reason to test rather than assume is that a model can be advertised multimodal, be multimodal in its weights, and still be text-only on the endpoint you happen to call. That is exactly the state Hy4 flash has been in for five days: the vision encoder shipped with the checkpoints, and the serving stacks have been catching up one PR at a time.

Where this leaves the choice

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 T​encent 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 T​encent 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.

What to watch next

Three things would change this piece, and each is checkable.

The stack merging. The vision PR sits on top of a text PR that is itself unmerged; both need review, and CI has not run on either. Until a wheel ships, no SGLang user can serve Hy4 flash images regardless of what the diff says.

T​encent 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 T​encent 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 T​encent's own WorkIndex suite and an 85-person internal blind test; neither touches vision.

Until one of those lands, 74.8 belongs in a sentence with its qualifiers attached: an instruct checkpoint, 300 items, high reasoning effort, three runs, one framework author's spot check on a PR that has not been reviewed. It is the best number that exists for Hy4 flash's vision side. That is a lower bar than it sounds, and the honest reading is that the model's vision is now demonstrably implemented and still largely unmeasured.

Compared in this article1

Detected from this article