DeepSeek V4 Flash 0423

deepseek-ai/DeepSeek-V4-Flash

13B active / 284B totalMoEText generationText input

DeepSeek · 1 recorded configuration

Measured setup

Hardware

  • 2× RTX 3090 · CPU offload

Runtime and tool

  • live llama-server

Quantisation

  • UD-IQ3_XXS

MTP draft depth

  • MTP off

Every option opens a recorded setup. Where the corpus recorded nothing at this page's other settings, the option names the fields it switches with it. Allocation and runtime detail stay in the technical record below.

2× RTX 3090 · llama.cpp

UD-IQ3_XXS · MTP off

PP · left axis · not measuredTG · right axis
View exact measurements
Exact context depth and measured fresh prefill, bulk prefix, and decode rates for DeepSeek V4 Flash 0423
Context depth (tokens)Fresh prefill PP (tok/s)TG (tok/s)
904—17.00
Technical record and provenance

Run details

Source run ID
rtx3090x2-deepseek-v4-flash-0731-ud-iq3-xxs-mmap
Configuration label
mmap · 904-token prompt · cache reused · median of 3
Hardware
2× RTX 3090
Runtime
llamaCpp
Weight placement
partially-offloaded
CPU expert layers
26
Placement detail
26 MoE expert layers on CPU; remaining tensors on two GPUs; exact GPU split unverified
KV-cache precision
Q8_0
Flash Attention
Enabled
Batch size
4,096
Microbatch size
512
Prompt tokens
904
Generated tokens
640
Speculative decoding
None (standard decoding)
Measured repetitions
3
Notes
Recovered from trpro:/home/mrx/perf/final-benchmark-mmap.json. Median of three measured server decode rates (17.027, 16.998, 16.976 tok/s), excluding one warm-up. One model load with a 65,536-token slot; this is not a context sweep. The chat prompt occupies 904 tokens; repeated requests reuse 900 tokens and evaluate four fresh tokens before generating 640 tokens. No comparable uncached prefill measurement is published. The exact benchmark date and runtime build were not recorded. Startup warns that repeated --override-tensor flags retain only the last value: the CPU expert override is retained, but the intended CUDA0/CUDA1 overrides are not verified. The saved final_bench.py is archival evidence only and must not be executed: its cleanup uses an unscoped pkill.
Source article
/benchmarks/
Raw log path
research/benchmarks/logs/trpro/deepseek-v4-flash-0731-ud-iq3-xxs/

Compare measured entries