DeepSeek-V4-Flash-Vision-Exp
One-line: First multimodal DeepSeek-V4 variant — same 13B-active MoE core as V4-Flash with an added vision tower; text-agent scores hold parity with V4-Flash-0731 while multimodal agent evals jump (ApexBench 36.5 vs 26.2). All agentic numbers lab-reported.
⚠ Vision-language model (pipeline_tag: image-text-to-text) — included because it publishes a full text-agent eval suite. All benchmark numbers below are lab-reported from the model card /
.eval_results/YAMLs; no independent source verified yet.
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | not published for this variant |
| HumanEval+ | — | not published |
| MBPP | — | not published |
| LiveCodeBench | — | not published |
| Aider polyglot | — | not published |
| SWE-bench Verified | — | not published (DeepSWE stand-in below) |
| DeepSWE | 59.3 | lab (model card + .eval_results/deep-swe.yaml) |
| NL2Repo | 57.7 | lab (model card) |
| DSBench-Hard | 63.6 | lab (model card) |
| AutomationBench (Public) | 25.7 | lab (model card) |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | not published |
| τ-Bench | — | not published |
| ToolACE | — | not published |
| GAIA | — | not published |
| Terminal Bench 2.1 | 83.9 | lab (model card + .eval_results/terminal-bench-2.1.yaml) |
| Toolathlon-Verified | 75.9 | lab (model card) |
| Cybergym | 75.3 | lab (model card) |
| ApexBench (Pass@1) | 36.5 | lab (multimodal agent; 0731 text sibling: 26.2†) |
| Agents’ Last Exam | 27.3 | lab (multimodal agent; 0731 text sibling: 25.2†) |
| Chartography | 64.3 | lab (multimodal agent) |
| ZeroBench (Pass@5) | 35.0 | lab (multimodal agent) |
† DeepSeek-V4-Flash-0731 ignored multimodal content in those inputs.
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | not published for this variant (sibling DeepSeek-V4-Flash: 86.2 lab) |
| GPQA Diamond | — | not published for this variant (sibling DeepSeek-V4-Flash: 88.1 lab) |
| MATH | — | not published |
| AIME 2025 | — | not published |
Efficiency
- Active params: ~13B per token (MoE, top-6 of 256 routed + 1 shared) — same core as DeepSeek-V4-Flash
- Total params: ~284B disk footprint + 32-layer vision tower (vision_dim 1024, patch 14, max 384 visual tokens)
- Context length: 1,048,576 tokens (1M) via YaRN rope scaling (base 64K, factor 16)
- VRAM (fp16 rough): ~26 GB active-params equivalent at runtime; full weights ~284 GB on disk — checkpoint ships FP4 experts + FP8 elsewhere (quantization_config fp8, expert_dtype fp4), so stock transformers vLLM support is preliminary; official recipe: vLLM on 4×GB300 or SGLang with DSpark speculative decoding
- Architecture: DeepseekV4ForCausalLM — MoE + Compressed Sparse Attention (CSA) / Heavily Compressed Attention (HCA), 3 nextn-predict layers, DSpark speculative path
What makes it notable
DeepSeek’s first open multimodal V4 release: the text-agent suite is unchanged in spirit from V4-Flash (Terminal Bench 2.1 83.9 vs 82.7, DeepSWE 59.3 vs 54.4, Toolathlon-Verified 75.9 vs 70.3) while the added vision tower unlocks real multimodal agent gains (ApexBench 36.5 vs 26.2, chart/ZeroBench numbers in Opus-4.8 territory). For local agent loops, the caveats from V4-Flash carry over: FP4/FP8 mixed weights need DeepSeek’s custom inference path (vLLM recipe targets a 4×GB300 node), so this is a big-host model — but MIT license and 1M context make it the strongest open-text-agent+vision combo in the Flash size class. Sits between the text-only deepseek-ai—DeepSeek-V4-Flash and the larger deepseek-ai—DeepSeek-V4-Pro siblings.
See also
- welcome
- Sibling: deepseek-ai—DeepSeek-V4-Flash (text-only, same core, more benchmarks)
- Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp