DeepSeek-V4-Flash-Vision-Exp: How Its Benchmarks Stack Up vs Opus 4.8

DeepSeek released DeepSeek-V4-Flash-Vision-Exp, its first experimental multimodal model in the V4 family, adding visual modules to the existing DeepSeek-V4-Flash architecture and scoring 36.5 Pass@1 on ApexBench versus 26.2 for its predecessor, DeepSeek-V4-Flash-0731. The model beats Anthropic's Opus 4.8 on ZeroBench (35.0 vs 34.0 Pass@5) and edges it on Agents' Last Exam (27.3 vs 25.7), but Opus 4.8 leads on most text agent benchmarks including NL2Repo, DSBench-Hard, and Cybergym. DeepSeek published the benchmarks on its model card, noting the 0731 baseline scores on ApexBench and Agents' Last Exam came from a model that ignored multimodal input, and the model supports vLLM and SGLang with speculative decoding via DSpark.

DeepSeek-V4-Flash-Vision-Exp: How Its Benchmarks Stack Up vs Opus 4.8 DeepSeek's new multimodal model scores on ApexBench, ZeroBench, and text agent tasks, compared directly against Opus 4.8 and its own predecessor. What is DeepSeek-V4-Flash-Vision-Exp? DeepSeek-V4-Flash-Vision-Exp is DeepSeek’s first experimental multimodal model in the V4 family. It’s built on the existing DeepSeek-V4-Flash architecture, with visual modules added and continued training applied to give the model image understanding on top of its existing text and agentic capabilities. DeepSeek released it as open weights, and it’s already picked up nearly 18,000 downloads on Hugging Face. The pitch from DeepSeek is straightforward: better multimodal agent performance than its predecessor, without giving up ground on the text-only agent tasks that DeepSeek-V4-Flash was already tuned for. The benchmark table DeepSeek published backs that framing up in some places and complicates it in others, especially once Anthropic’s Opus 4.8 is added to the comparison. TL;DR DeepSeek-V4-Flash-Vision-Exp adds vision to DeepSeek-V4-Flash through added visual modules and continued training, rather than being a from-scratch multimodal build. Multimodal agent scores jumped noticeably over the previous DeepSeek-V4-Flash-0731 release , including a large gain on ApexBench 36.5 vs 26.2 Pass@1 . The model beats Opus 4.8 on ZeroBench Pass@5 , scoring 35.0 against Opus’s 34.0, and edges it slightly on Agents’ Last Exam 27.3 vs 25.7 . Opus 4.8 still leads on most text agent benchmarks , including NL2Repo, DSBench-Hard, and Cybergym, sometimes by double-digit margins. Text-only agent performance stayed roughly flat compared to the prior DeepSeek-V4-Flash release , which is the claim DeepSeek makes about not regressing on non-vision tasks. A footnote in DeepSeek’s own numbers matters : the 0731 baseline scores on ApexBench and Agents’ Last Exam came from a model that ignored the multimodal parts of the input entirely, so those comparisons aren’t apples to apples. The model ships with support for vLLM and SGLang , including speculative decoding via a method DeepSeek calls DSpark, aimed at production-style serving rather than just research use. Remy is new. The platform isn't. Remy is the latest expression of years of platform work. Not a hastily wrapped LLM. What benchmarks did DeepSeek publish? DeepSeek’s model card lays out two groups of benchmarks: text agent capabilities and multimodal agent capabilities. Three models are compared side by side: DeepSeek-V4-Flash-Vision-Exp the new release , DeepSeek-V4-Flash-0731 its immediate predecessor , and Anthropic’s Opus 4.8 as the external reference point. On the text agent side, the benchmarks cover things like Terminal Bench 2.1, NL2Repo, Cybergym, DeepSWE, Toolathlon-Verified, DSBench-Hard, and AutomationBench Public . These are tasks built around coding, tool use, and multi-step automation rather than image understanding. On the multimodal side, DeepSeek reports ApexBench Pass@1 , Agents’ Last Exam, Chartography, and ZeroBench Pass@5 . These benchmarks specifically test how well a model reasons over visual input as part of agentic tasks, not just static image captioning or classification. DeepSeek evaluated its own models using what it calls DeepSeek Harness in “minimal mode,” running at maximum reasoning effort with temperature set to 1.0 and top p at 0.95. That detail matters for anyone trying to reproduce the numbers: the scores reflect a specific harness and sampling configuration, not necessarily what you’d get running the model with default settings in a different framework. How does DeepSeek-V4-Flash-Vision-Exp compare to its predecessor? The clearest story in the benchmark table is the jump in multimodal agent performance between DeepSeek-V4-Flash-Vision-Exp and DeepSeek-V4-Flash-0731. On ApexBench, the new model scores 36.5 Pass@1 versus 26.2 for the 0731 release, a gain of more than 10 points. On Agents’ Last Exam, it improves from 25.2 to 27.3. Chartography and ZeroBench don’t even have comparison numbers for the 0731 model, marked with a dash in DeepSeek’s table, which suggests those benchmarks weren’t run on it at all or weren’t meaningful given that older model’s lack of real vision training. That last point is worth sitting with. DeepSeek includes a footnote clarifying that for ApexBench and Agents’ Last Exam, DeepSeek-V4-Flash-0731 “ignores the multimodal elements in the input.” In practice, that means the 0731 baseline was answering these multimodal benchmark questions using only the text portions of the prompts, since it didn’t have vision capability built in yet. So the improvement isn’t just a model getting better at an existing skill. It’s a model gaining a skill it didn’t meaningfully have before. On the text agent side, the differences between the two DeepSeek models are small. Terminal Bench 2.1 goes from 82.7 to 83.9. DeepSWE actually improves from 54.4 to 59.3. NL2Repo climbs from 54.2 to 57.7. Cybergym dips slightly, from 76.7 to 75.3. Toolathlon-Verified rises from 70.3 to 75.9. Taken together, this supports DeepSeek’s claim that adding vision capability didn’t cost the model much, if anything, on pure text agent tasks. If anything, several text benchmarks improved alongside the vision gains. How does it compare to Opus 4.8? This is where the picture gets more mixed, and it depends heavily on which category of benchmark you’re looking at. One coffee. One working app. You bring the idea. Remy manages the project. On text agent tasks, Opus 4.8 generally leads. It scores 85.0 on Terminal Bench 2.1 versus 83.9 for DeepSeek-V4-Flash-Vision-Exp, a modest gap. The gap widens elsewhere: Opus scores 69.7 on NL2Repo against DeepSeek’s 57.7, and 71.7 on DSBench-Hard against DeepSeek’s 63.6. Cybergym also favors Opus, 78.3 to 75.3. The one text benchmark where DeepSeek’s new model pulls ahead is DeepSWE, where it scores 59.3 against Opus’s 58.0, a narrow win but a win nonetheless. On multimodal agent tasks, the results flip in interesting ways. ApexBench still favors Opus, 39.4 versus 36.5, a real gap of nearly three points. But on ZeroBench Pass@5 , DeepSeek-V4-Flash-Vision-Exp actually beats Opus 4.8, scoring 35.0 to Opus’s 34.0. On Agents’ Last Exam, DeepSeek also edges ahead, 27.3 to 25.7. Chartography is close, with Opus slightly ahead at 65.0 versus DeepSeek’s 64.3. So the honest summary is: Opus 4.8 remains stronger on most text-heavy agentic benchmarks, and it still leads on the single largest multimodal benchmark in the table ApexBench . But DeepSeek-V4-Flash-Vision-Exp isn’t just closing the gap on multimodal tasks, it’s outright surpassing Opus on two of the four multimodal benchmarks DeepSeek chose to report, ZeroBench and Agents’ Last Exam. Is this a meaningful benchmark comparison, or a curated one? Any time a lab publishes benchmarks against a rival’s model, it’s fair to ask how the comparison was chosen. DeepSeek picked seven text agent benchmarks and four multimodal benchmarks. Opus 4.8 wins five of the seven text benchmarks and one of the four multimodal ones outright ApexBench, by a clear margin . DeepSeek’s new model wins on DeepSeek’s DeepSWE benchmark, ZeroBench, and Agents’ Last Exam. That’s not a table stacked entirely in DeepSeek’s favor. If it were pure marketing, you’d expect DeepSeek to either drop ApexBench where it loses by the widest margin among multimodal tests or omit the text agent section where Opus wins more often than not. The fact that both appear, with Opus ahead in several categories, suggests DeepSeek is at least presenting a reasonably complete picture rather than only the wins. The footnote about the 0731 baseline ignoring multimodal input is also a point in favor of the table’s credibility. DeepSeek could have left that detail out and let the multimodal comparison look like a bigger apples-to-apples leap than it actually represents. Including it is a transparency signal worth noting, even if it does mean the “improvement” over the prior model is partly an artifact of the older model not really being multimodal at all. What does this mean for people building with these models? For developers choosing between models for agentic, vision-involved workloads, the practical takeaway is that DeepSeek-V4-Flash-Vision-Exp is a legitimate multimodal upgrade over its own predecessor, not just a benchmark relabeling. The ZeroBench and Agents’ Last Exam wins against Opus 4.8 suggest it’s genuinely competitive on some visual reasoning tasks, even if Opus stays ahead on raw agentic coding and tool-use benchmarks. The model ships as open weights with serving support for both vLLM and SGLang, including a speculative decoding method DeepSeek calls DSpark, which is aimed at cutting latency in production deployments. That combination, open weights plus serving-framework support plus a documented benchmark table with footnotes, makes it easier for teams to actually verify these numbers themselves rather than take DeepSeek’s word for it. Frequently Asked Questions What is ApexBench and why does it matter here? Remy doesn't build the plumbing. It inherits it. Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something. Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want. ApexBench is one of the multimodal agent benchmarks DeepSeek reports, measured with Pass@1. It’s the benchmark where Opus 4.8 shows its clearest lead over DeepSeek-V4-Flash-Vision-Exp 39.4 vs 36.5 , making it the strongest counterpoint to DeepSeek’s otherwise favorable multimodal comparisons. Does DeepSeek-V4-Flash-Vision-Exp beat Opus 4.8 on any benchmark? Yes. It scores higher than Opus 4.8 on ZeroBench Pass@5 35.0 vs 34.0 , Agents’ Last Exam 27.3 vs 25.7 , and DeepSWE 59.3 vs 58.0 , based on DeepSeek’s published table. Is the comparison to DeepSeek-V4-Flash-0731 fair? Partly. DeepSeek’s own footnote states that the 0731 model ignored multimodal elements in the input for the ApexBench and Agents’ Last Exam scores, meaning those particular comparisons measure a genuinely non-multimodal model against a multimodal one, not two versions of the same capability. What serving frameworks support this model? DeepSeek’s documentation shows configurations for both vLLM and SGLang, including support for a speculative decoding method called DSpark, aimed at reducing inference latency in production settings. Is DeepSeek-V4-Flash-Vision-Exp open source? Yes. It’s released as open weights on Hugging Face under the MIT License, with tokenizer files, prompt encoding references, and a minimal PyTorch inference implementation included in the repository.