DeepSeek Releases V4-Flash-Vision-Exp, First Multimodal Model in V4 Family
DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, its first experimental multimodal model in the V4 family, adding visual understanding to the V4-Flash architecture. The 305B-parameter model shows substantial gains on multimodal agent benchmarks while holding steady on text-only tasks.
DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, an experimental 305-billion-parameter model marking the first multimodal entry in its V4-Flash lineup. The model is available now on Hugging Face under the MIT License, though it is not yet deployed by any inference provider.
What's new
DeepSeek-V4-Flash-Vision-Exp builds on the existing DeepSeek-V4-Flash architecture by adding visual modules and undergoing continued training to enable image understanding. According to DeepSeek, the model achieves "substantial improvements" on multimodal agent tasks compared to its predecessor, DeepSeek-V4-Flash-0731, while keeping text-only agent performance roughly comparable.
The repository includes a tokenizer, a prompt-encoding reference supporting both OpenAI-style JSON content blocks and a compact <image>path</image> TXT notation, and a minimal PyTorch inference implementation. That implementation covers the vision encoder and aligner, DFlash attention, the mixture-of-experts (MoE) routing, Hyper-Connections, and the DSpark forward path. Model weights are distributed as safetensors in BF16, F32, F8_E4M3, I8, and I64 tensor formats.
Benchmark results
DeepSeek reports the following scores, comparing DeepSeek-V4-Flash-Vision-Exp against DeepSeek-V4-Flash-0731 and Anthropic's Opus-4.8, using DeepSeek Harness in minimal mode with maximum reasoning effort, temperature 1.0, and top_p 0.95:
Text agent capabilities:
- Terminal Bench 2.1: 83.9 vs. 82.7 (V4-Flash-0731) vs. 85.0 (Opus-4.8)
- NL2Repo: 57.7 vs. 54.2 vs. 69.7
- Cybergym: 75.3 vs. 76.7 vs. 78.3
- DeepSWE: 59.3 vs. 54.4 vs. 58.0
- Toolathlon-Verified: 75.9 vs. 70.3 vs. 76.2
- DSBench-Hard: 63.6 vs. 59.6 vs. 71.7
- AutomationBench (Public): 25.7 vs. 25.1 vs. 27.2
Multimodal agent capabilities:
- ApexBench (Pass@1): 36.5 vs. 26.2 vs. 39.4
- Agents' Last Exam: 27.3 vs. 25.2 vs. 25.7
- Chartography: 64.3 vs. not reported vs. 65.0
- ZeroBench (Pass@5): 35.0 vs. not reported vs. 34.0
DeepSeek notes that on ApexBench and Agents' Last Exam, the text-only V4-Flash-0731 baseline ignored multimodal elements in the input, which explains its lower scores on those tests.
On several multimodal benchmarks — ZeroBench and near-parity on Chartography and ApexBench — the new model closes much of the gap with Opus-4.8, a substantially larger and more established closed model. On text-only agent tasks, DeepSeek-V4-Flash-Vision-Exp trails Opus-4.8 on harder benchmarks like NL2Repo (57.7 vs. 69.7) and DSBench-Hard (63.6 vs. 71.7), but stays close on Terminal Bench 2.1 and Toolathlon-Verified.
No pricing has been disclosed, and the model is not yet available through any inference provider. Context window size was not specified in the release materials.
What this means
This is a research-stage, experimental release rather than a production-ready product — the "Exp" tag and lack of inference-provider deployment signal DeepSeek is testing the vision integration before a wider rollout. The architecture reuses V4-Flash's core components (DFlash attention, MoE, Hyper-Connections) rather than introducing a new base model, suggesting DeepSeek is incrementally extending its existing V4 line to multimodal use cases instead of training a separate vision-first model from scratch. The open MIT license and inclusion of a minimal PyTorch inference stack make it immediately usable for developers who want to experiment with DeepSeek's multimodal agent capabilities locally, ahead of any hosted API.
Related Articles
SenseNova Releases U1.5-8B-MoT, an Open-Weight Unified Model for Image Generation and Editing
SenseNova has released SenseNova-U1.5-8B-MoT, an open-weight native multimodal model built on its NEO-unify architecture for image generation, editing, and native 4K output. The model is available on Hugging Face under an Apache 2.0 license, with no inference pricing yet since it must be self-hosted.
Tencent Open-Sources Hy4 Preview: 770B-Parameter MoE Model with 1M-Token Context
Tencent's Hy Team has open-sourced Hy4 preview, a 770-billion-parameter Mixture-of-Experts model with 49 billion activated parameters and a 1-million-token context window. The model is available under Apache 2.0 alongside an FP8-quantized variant, with Tencent claiming it beats GLM 5.3 and Kimi K3 on internal engineering evaluations.
GLM-5.3-Flash Debuts as Zhipu AI's First Natively Multimodal Model, 320B Parameters with 18B Active
Zhipu AI has released GLM-5.3-Flash, the first natively multimodal model in its GLM-5 series, built on a 320B-parameter mixture-of-experts architecture with only 18B active parameters. The company claims it outperforms GLM-5.2 while approaching Claude Opus 4.8 on coding and agentic benchmarks at a fraction of the cost. Unsloth has published quantized GGUF versions for local inference.
Z.ai Launches GLM-5.3-Flash: 1M-Token Context, Image Support, Claimed 10x Cost Cut Over GLM-5.2
Z.ai has released GLM-5.3-Flash, a 320-billion-parameter Mixture-of-Experts model with 18 billion active parameters, a 1-million-token context window, and image input support. The model launched on LM Studio's Bionic platform hours after its official unveiling, with LM Studio claiming it is 9-10x cheaper to run than GLM-5.2.
Comments
Loading...