An experimental multimodal model for text reasoning, coding, and image understanding. Two-node validation covers text and image input, concurrent requests, and long-context processing.
Multimodal understandingText, images → text
2 MegaCubesTest report available
01 / Model parameters
Model architecture & capabilities
DeepSeek-V4-Flash-Vision-Exp adds vision training to the DeepSeek-V4-Flash language backbone for image, chart, and document understanding, combined with text reasoning and tool use. It produces text or tool calls. Compatibility of this experimental release depends on the specific weights and inference engine.
Q&A and coding tasks
Image, chart, and document understanding
Agent tasks using visual information
Publisher / model
DeepSeek · DeepSeek-V4-Flash-Vision-Exp
Parameter count
284B language backbone / 13B active; vision module separate, total parameters of the full vision model not specified in the official model card
Architecture
43-layer MoE; 256 routed experts, 6 selected per step, plus 1 shared expert; includes a vision encoder
Model context
Configured limit of 1,048,576 tokens; see below for the range validated on MegaCube
Input / output
Text and image input; text and tool-call output
Precision & version
Official configuration includes FP4 experts and FP8 quantization; experimental Vision-Exp version
License
MIT
The model analyzes images and returns text or tool calls; it does not generate images. Two-device tests and official agent benchmarks use different tasks and environments, so their scores are kept separate.
Context 262,144; maximum concurrency 8; BatchTokens=16384
Acceleration & caching
MTP=6; KV=nvfp4_ds_mla; prefix caching and chunked prefill
Connect two Cubes
Verify the high-speed direct link and management network, using identical weights and images on both devices.
Configure the head node
Enter the node addresses, model path, and service parameters in the launcher configuration.
Launch both ranks
Start both ranks from the head node; the launcher synchronizes dependencies and completes warm-up.
Validate text and images
Check API health, then validate text, images, and the target concurrent workload.
03 / Performance tests
MegaCube performance test results
2026-09-01; 16 performance test configurations. Each configuration has 10 warm-up generations at the target output length, followed by at least 180 seconds of measurement. Memory figures are totals across both nodes.
102.19 tok/s16K/1K tokens · C8 aggregate output
0.453 s2K/256 tokens · C1 TTFT P95
250,004Long input validated without a warm cache
Full performance matrix
Input/output
Concurrency
RPS
Input throughput (tok/s)
Output throughput (tok/s)
TTFT P95 (s)
TPOT P95 (ms)
E2E P95 (s)
MTP acceptance rate
Prefix hit rate
Memory (GiB)
Success rate
512/128
1
.201
103.04
25.76
.453
56.57
7.58
15.16%
50.00%
224.44
100%
512/128
2
.275
141.02
35.07
.630
85.64
11.13
13.49%
50.00%
224.45
100%
512/128
4
.394
201.85
50.46
.910
118.67
15.71
15.23%
50.00%
224.49
100%
512/128
8
.497
254.57
63.37
2.100
194.19
25.08
14.68%
50.00%
224.43
100%
2048/256
1
.182
373.34
46.67
.453
30.20
8.15
44.86%
87.50%
224.50
100%
2048/256
2
.223
457.25
57.16
.790
76.71
20.11
34.22%
87.50%
224.44
100%
2048/256
4
.320
655.61
81.95
1.070
93.22
24.39
37.04%
87.50%
224.44
100%
2048/256
8
.364
745.47
93.18
1.940
178.77
46.61
30.47%
87.50%
224.47
100%
16384/1024
1
.039
635.74
39.73
.440
46.17
47.67
31.13%
98.44%
224.46
100%
16384/1024
2
.055
900.40
52.02
.670
80.40
70.23
29.03%
98.44%
224.48
100%
16384/1024
4
.083
1357.77
72.09
1.380
115.85
80.04
28.51%
98.44%
224.48
100%
16384/1024
8
.109
1782.75
102.19
2.212
174.99
109.56
32.52%
98.44%
224.52
100%
32768/1024
1
.023
756.65
23.65
.500
57.02
58.80
11.85%
99.22%
224.56
100%
32768/1024
2
.032
1052.09
32.88
.850
76.64
78.99
10.74%
99.22%
224.58
100%
32768/1024
4
.045
1478.21
46.19
1.410
109.53
112.84
11.07%
99.22%
224.57
100%
32768/1024
8
.058
1894.15
59.19
2.190
168.52
174.50
12.10%
99.22%
224.61
100%
The matrix includes prefix-cache hits. TTFT measured with a warm cache does not represent input processing time for an uncached long prompt.
Functional and compatibility validation of this experimental version covers only the scenarios in this report.
C1 / C4 / C8 denotes 1 / 4 / 8 concurrent requests. TTFT (time to first token) measures elapsed time from submitting a request to receiving its first output token. TPOT (time per output token) measures average time per generated token after the first token. Aggregate throughput is the combined output rate across concurrent requests; a single-request token generation rate is measured separately in tok/s.
Public benchmark references
Model capability & platform performance benchmarks
The following results are published by model developers or platform providers and are separate from the MegaCube tests above. Capability scores measure task performance; tok/s measures token throughput. Results obtained with different models, numerical precision, hardware, or test protocols are not directly comparable.
DeepSeek experimental vision model official benchmarks
DeepSeek · Checked on 2026-09-20
Evaluated model
DeepSeek-V4-Flash-Vision-Exp
Test platform
Inference hardware not disclosed by the source
DeepSeek experimental vision model official benchmarks · Data
Benchmark
Score
What it measures
Terminal-Bench 2.1
83.9
Terminal operation tasks
DeepSWE
59.3
Software engineering tasks
Toolathlon Verified
75.9
Tool-calling tasks
Test conditionsText-agent benchmarks use DeepSeek Harness Minimal, maximum reasoning effort, temperature=1, and top_p=0.95. These scores describe task-completion ability, not two-device response speed.
