HUST and Horizon propose Multimodal Flow: one shared Flow objective for language and vision; 1.6B MF-1 claims to beat hybrid and fully discrete architectures
According to a November 5, 2026 report by Jiqizhixin, Wang Xinggang's team at Huazhong University of Science and Technology, along with Horizon and other institutions, proposed Multimodal Flow: language and vision each retain continuous embeddings while sharing a single Flow Matching objective. The 1.6B MF-1 was pretrained from scratch using only 150B tokens. The report says that under controlled comparisons with the same data and parameter budget, it outperforms hybrid and fully discrete architectures on GenEval, GQA, VQAv2, MMBench and other benchmarks.
What changed
According to the November 5, 2026 Jiqizhixin report, Wang Xinggang's team at Huazhong University of Science and Technology, in collaboration with Horizon Robotics and other institutions, proposed Multimodal Flow, which argues for unified modeling of language and vision in a continuous embedding space. The report groups earlier approaches into two categories. Fully discrete routes such as Chameleon and Emu3 quantize images into discrete tokens, concatenate them with language and run autoregression; limited by the visual tokenizer, high-frequency detail is easily lost in quantization. Hybrid discrete-continuous routes such as Transfusion and JanusFlow use discrete prediction for text and continuous diffusion for images, at the cost of maintaining two objective functions and sampling pipelines, making joint optimization harder. Multimodal Flow keeps native continuous embeddings for both but shares one Flow Matching objective. The team instantiated a main model, MF-1: 1.6B parameters, pretrained from scratch and consuming only 150B tokens. Reported figures include a GenEval overall score of 0.821, DPG-Bench long-prompt alignment of 83.44, and a GenEval counting dimension of 0.73; on the understanding side, VQAv2 72.6, MMBench 67.2 and POPE 86.1, averaging 75.3 across the three. In controlled comparisons with the same data, parameter budget and optimization configuration, the report says it surpasses Transfusion-style hybrid architectures and Chameleon-style fully discrete architectures on GenEval, GQA, VQAv2 and MMBench. Architecturally, text is encoded by a frozen T5-small in blocks of eight tokens into text superchunks; images are processed by a frozen SigLIP2 to extract 16×16 spatial grid features into visual superchunks. The two are arranged in task order into a multimodal sequence and fed into a shared Chunk-Causal Flow backbone. Attention uses a Chunk-Causal mask: when predicting the k-th superchunk, earlier superchunks are fully visible, positions within the current chunk are jointly modeled, and future superchunks are masked. Attention layers are shared, while FFNs are separated by modality. During training, multiple target superchunks can be predicted in parallel, each sampling its own Flow timestep; at inference, generation proceeds chunk by chunk and reuses the KV cache of completed superchunks. Changing the superchunk order and target direction switches the same set of parameters among text generation, text-to-image, image-to-text and VQA. The report also relays two sets of ablations: on visual representation, semantic features such as SigLIP2 and DINOv2 were overall better than FLUX.2 VAE, SD-VAE latents and raw pixels on generation and description tasks; the same CFG can simultaneously act on text and image, with the best guidance scale around 3 for image-to-text and around 5 for text-to-image. On transfer, mixed pretraining followed by fine-tuning, compared with fine-tuning directly from random initialization, raised GenEval from 0.527 to 0.821, SEED-Bench from 31.6 to 62.4 and MMBench from 36.0 to 67.2. The team says future work will extend to longer interleaved image-text sequences, video and other structured modalities; the report also includes links to the paper and a GitHub code repository.
What it means for developers
The direct significance of this route for developers is mainly architectural, not that another strong model has appeared. Hybrid approaches must maintain diffusion and autoregressive losses and sampling code in one training stack, increasing engineering complexity and the tuning surface; fully discrete approaches stake part of generation quality on the visual tokenizer. The approach described in the report folds the two into one Flow objective, reducing the number of objective functions to maintain and making task switching a matter of changing superchunk order. If such simplification is carried into your own code, what it mainly saves is complexity in the data pipeline and sampling logic, not compute. The second concrete effect is that the role of the front-end encoders changes. The design in the report freezes T5-small and SigLIP2, so the model only learns the Flow between superchunks. That means swapping the text or visual encoder directly changes the statistical properties of the input distribution, and the ablation results suggest semantic visual features are better suited to a unified generative model than reconstruction-oriented VAE latents. If your scenario requires pixel-level high-fidelity reconstruction, the evidence in the report cannot be directly extrapolated. Third is the inference shape. Chunk-by-chunk generation plus KV cache reuse is not exactly the same as pure autoregressive decoding or diffusion sampling; batching, cache management and streaming output all have to be redesigned. The report gives no throughput, memory or latency figures, and the 1.6B scale is smaller than current mainstream unified models, so judging cost-performance at larger scale from this is premature. The above is an engineering reading of the report, not independent verification of the approach. MF-1's benchmark numbers come from the team's self-report, and the baselines in the controlled comparisons were also run under the team's own training configuration.
Integration and evaluation
The following three are conditional suggestions inferred from the report, not verified conclusions. 1. Whether to rebuild your unified multimodal training stack around this skeleton. The reported change is replacing the dual objectives of hybrid approaches with one shared Flow Matching objective, and surpassing hybrid and fully discrete baselines under a controlled budget. The choice is whether a new project should directly adopt the Chunk-Causal Flow structure with shared attention and modality-specific FFNs. The premise is that you already have a freezeable semantic encoder on the input side and that downstream metrics cover both image-text generation and understanding; if you actually care only about pure text quality, the report does not cover that. A measurable validation is to reproduce the same scale ladder: train 0.6B, 1.2B and 1.6B variants, and see whether Flow objective loss and language PPL fall together while GenEval and image-to-text CLIPScore rise with scale; if the four curves do not move in the same direction, that indicates the objective does not hold on your data. 2. Whether to choose semantic features or reconstruction-oriented latents for the visual front end. The report's ablation conclusion is that under the same training budget, SigLIP2 and DINOv2 outperform FLUX.2 VAE, SD-VAE latents and raw pixels. The choice is whether to first freeze a semantic encoder as the visual superchunk source or continue using VAE latents with a diffusion head. The premise is that the target task is mainly generation and description; if reconstruction fidelity comes first, this conclusion cannot be copied directly. A measurable approach is an A/B test under the same budget: use GenEval overall score and description-type CLIPScore as primary metrics, add a reconstruction error as a cross-check, and reweigh if the latter degrades markedly. 3. How to configure guidance. The report says the same CFG can act in both directions on text and image, with the best guidance scale around 3 for image-to-text and around 5 for text-to-image. The choice is to stop maintaining separate guidance implementations for the two directions and use one CFG with task-specific scales. The premise is that your training objective is indeed a unified Flow, not two independent heads. A measurable validation is to sweep the scale: look for peaks in alignment metrics such as DPG-Bench around 3 for image-to-text and around 5 for text-to-image, while recording the cost to diversity or consistency as the scale rises, and confirm the peak is not bought by sacrificing the other. One more basic point: clarify weights and licensing before investing. The report includes a paper and a code repository, but does not say whether weights are released or under what license they would be released.
Limits and open questions
The MF-1 and Multimodal Flow described here have no corresponding entry in KimiSeek's current model catalog, and this model version is not currently listed on KimiSeek. The report also mentions no hosted API, so there is no vendor API availability claim to point to; if a hosted service appears later, its availability scope and billing will be subject to the vendor's documentation and the site's pricing page. A few other limitations: the report uses the phrase 'open-source code' and attaches a code repository, but does not say whether model weights are released or provide license terms; code availability does not equal open weights, let alone an open-source license. All benchmark numbers, ablation conclusions and controlled comparison results in this article come from the team's self-report and the report's relay; this article has not independently reproduced them, and the baseline numbers were also obtained under the team's own training configuration. At 1.6B parameters and 150B tokens, the scale is clearly smaller than current mainstream unified models; the report itself calls it only a starting point, and whether the conclusions hold as scale increases is unknown. The report gives no inference throughput, memory usage or cost data, so the actual speedup from KV cache reuse cannot be judged. Modalities such as video and long interleaved sequences appear only in future plans and have no results yet.