DeepSeek-V3
No benchmarks for DeepSeek-V3 yet.
No DeepSeek-V3 benchmarks yet.
Run on your hardware to populate this page: pipx install llm-speed && llm-speed bench
$ pipx install llm-speed && llm-speed bench
Community folklore
14 unverified claims extracted from Reddit/HN comments. Lower trust than signed runs above; every row links to the source.
- communityconfidence 75%
5.00tok/s — deepseek-V3 on Together AI via hosted-api FP4
our signed data: Together AI · deepseek-V3
“VRAM, higher vRAM might helpful for long-context, PCIe5 might helpful for prefill. DDR4 Epyc or Icelake-SP Xeons with GPU may only reach \~5t/s(from my deepseek-V3 experience), which is pretty slow. - Serious one: 4x RTX Pro 6000 Blackwell together with modern Xeon or Epyc shoul…”
- communityconfidence 65%
10.00tok/s — DeepSeek V3 on RTX 4090 FP8
our signed data: RTX 4090 · DeepSeek V3
“g fast. Expert weights on CPU (4-bit GGUF): All the huge MoE banks live in system RAM and load as needed. End result: I’m seeing about \~10 tokens/sec with a 32K context window—pretty smooth for local tinkering. KTransformers made it so easy with its Docker image. It handles …”
- communityconfidence 65%
56.00tok/s — deepseek v3.2 via vllm FP16
our signed data: deepseek v3.2
“: With the vllm-gfx906-mobydick fork, you can also run smaller recent models (as the base is vllm v0.17.1) like **Qwen3.5 27B** (reaching **56 tok/s** at MTP5 and TP4 but it fits also on 1 MI50 32GB with 65k context; maybe later, if you are interested in, I can also make another …”
- communityconfidence 60%
25.00tok/s — DeepSeek V3 on H100 via vllm
our signed data: H100 · DeepSeek V3
“other developers, unlimited tokens Running DeepSeek V3 (685B) requires 8×H100 GPUs which is about $14k/month. Most developers only need 15-25 tok/s. sllm lets you join a cohort of developers sharing a dedicated node. You reserve a spot with your card, and nobody is charged until…”
- communityconfidence 60%
10.85tok/s — DeepSeek-V3.1 on RTX 5090 via llama.cpp
our signed data: RTX 5090 · DeepSeek-V3.1
“gh the system popping out through a small gap left by RTX 5090). DeepSeek-V3.1-Terminus with context = 37279 tokens: PP = 151.76 tps, TG = 10.85 tps Some things I discovered running local LLMs: * For water-cooled CPU systems, there is not enough air circulation to cool the RAM…”
- communityconfidence 60%
151.8tok/s — DeepSeek-V3.1 on RTX 5090 via llama.cpp
our signed data: RTX 5090 · DeepSeek-V3.1
“is threaded through the system popping out through a small gap left by RTX 5090). DeepSeek-V3.1-Terminus with context = 37279 tokens: PP = 151.76 tps, TG = 10.85 tps Some things I discovered running local LLMs: * For water-cooled CPU systems, there is not enough air circulatio…”
- communityconfidence 60%
10.00tok/s — DeepSeek V3 on RTX 4090 via llama.cpp
our signed data: RTX 4090 · DeepSeek V3
“nt=share_button) \- DeepSeek V3-0324 Q4\_K\_M with 512GB RAM DDR5 RAM 4800MHz and one RTX 4090; Prompt eval speed: 40 t/s; generation: \~10 t/s; DDR5 with 8 channels certainly improves the speed but it is not blazing fast. 3. If you have more VRAM, you can try --ubatch-size …”
- communityconfidence 60%
40.00tok/s — DeepSeek V3 on RTX 4090 via llama.cpp
our signed data: RTX 4090 · DeepSeek V3
“css&utm_term=1&utm_content=share_button) \- DeepSeek V3-0324 Q4\_K\_M with 512GB RAM DDR5 RAM 4800MHz and one RTX 4090; Prompt eval speed: 40 t/s; generation: \~10 t/s; DDR5 with 8 channels certainly improves the speed but it is not blazing fast. 3. If you have more VRAM, yo…”
- communityconfidence 55%
10.00tok/s — Deepseek v3.2 via vllm AWQ
our signed data: Deepseek v3.2
“16x AMD MI50 32GB at 10 t/s (tg) & 2k t/s (pp) with Deepseek v3.2 (vllm-gfx906) Deepseek 3.2 AWQ 4bit @ 10 tok/s (output) // 2000 tok/s (input of 23k tok) on vllm-gfx906-deepseek with 69000 context length **Power draw**: 550W (idle) / 2400W (peak inf”
- communityconfidence 55%
27.64tok/s — DeepSeek V3 on RTX 5090 IQ4_XS
our signed data: RTX 5090 · DeepSeek V3
“at X16 5.0, 5090(2) at X4 5.0, all the rest at X4 4.0: 940 t/s PP, 26.75 t/s TG. * 5090s at X16 5.0, all the rest at X16 4.0: 1170 t/s PP, 27.64 t/s TG. **DeepSeek V3 0324, IQ4\_XS, offloading about 120GB to CPU:** Loading with: ./llama-server -m '/run/media/pancho/MyDrive…”
Request a DeepSeek-V3 run
No measured run yet for these pairs. Vote for the one you want; the most-requested runs go first. See all requests.
We never sell or share your email. Privacy