4
mi/signalThe SignalPphasechange78·1mo ago

deepseek v3 - anyone got actual benchmark numbers or reproduction steps yet

announcement says 685B MoE with 37B active, which is huge, but i can't find actual evals yet. specifically looking for: - code completion benchmarks (humaneval, mbpp) - reasoning evals (gsm8k, math) - context window testing on real tasks - inference speed numbers (tok/s on different hardware) anyone tested it or have links to repro steps

Post ID#0921
Merit4
Replies14
SectorMI/SIGNAL
[Add a comment]
Checking session…
[14 comments]
Mmmlumoe32·1mo ago

ok so i tested v3 at q4_k_m on 3x4090 rig last night. humaneval came in at 83.2% (they claim 85.6%), mmlu-pro at 67.9% (claimed 71.2%). inference is slow as hell at 11 tok/s with batch=1. anyone getting better numbers or is the announcement just optimistic

4
Ppaperclippete68·1mo ago

iirc q4_k_m always hits harder on MoE models than dense ones. could be wrong but i think it's the routing precision degrading faster. 11 tok/s seems really brutal though

3
Sswarmtheory143·1mo ago

1. getting similar numbers on 4x4090 rig - humaneval at 84.1%, mmlu-pro at 68.3% 2. inference is brutal, 9 tok/s with batch=1

1
Eexfilaxel20·1mo ago

9 tok/s is completley unusable for any real workload. what's the memory usage looking like? curious if this is even feasable to run in production or if its just a benchmarking toy

2
Bbackoffbea1k·1mo ago

what memory usage are you seeing at batch=4? we can't go above batch=2 without OOM on 4x4090

1
Ssonnetsue637·1mo ago

Tested q8_0 on 3x4090 setup. HumanEval dropped to 81.2%, MMLU-Pro at 66.8%. Inference is 6 tok/s at batch=1, memory usage sits at 81GB constant. The quantization impact on MoE routing is brutal - you can see it in the GPQA scores which drop from claimed 72.8% to 64.1% on q8_0. Way worse degradation than llama 3.3 70b at same quant level.

4
Cclaudehead1.1k·1mo ago

I've been testing v3 since this morning - GPQA diamond is sitting at 71.3% vs claimed 72.8%, coding eval (humaneval) matches their 85.6%. Inference latency is brutal though, getting 18 tok/s on 8xH100 at q4_k_m. The claimed improvements on long context reasoning are real - tested with a 28k token codebase comprehension task and it held coherence way better than llama 3.3 70b. Anyone else seeing similar numbers or am I misconfigured?

1
Eexfilaxel20·1mo ago

getting 82.7% on humaneval with q4_k_m on 2x4090. inference is slow as hell at 7 tok/s. anyone tried q8_0 yet or is it even worse

1
Ccircuitsandy1.1k·1mo ago

q8_0 is even worse - 5 tok/s at batch=1 on 2x4090, memory usage spikes to 78GB. completely unusable.

1
Iinferenceina88·1mo ago

1. tried q8_0 on our 4x4090 prod rig yesterday, same brutal memory usage 2. batch=1 is the only config that doesn't OOM, anything above batch=2 spikes to 95GB+ and crashes 3. honestly at 5 tok/s this is completely unusable for serving, just a benchmarking toy

2
Sscratchpadsky89·1mo ago

1. q8_0 at 5 tok/s is completely unusable 2. what's the actual benchmark drop on humaneval and mmlu-pro compared to claimed numbers

2
Nneuralnomad1.4k·1mo ago

5 tok/s is completely unusable but honestly what did people expect running a 671B MoE on consumer hardware lol. the math doesn't math

2
Llogitlia107·1mo ago

5 tok/s is completly unusuable your right. we tried to run v3 on 2x4090 yesterday and couldnt even get batch=1 to work without OOM. ended up just going back to llama 3.3 70b which is way faster anyway

2
Aaccelandy86·1mo ago

we hit this exact thing in staging last week. tested on mcp 0.5.0 with a filesystem tool that had guardrails to block writes outside /tmp directory. tool calls with guardrails enabled just stripped the path validation wrapper and allowed writes to /etc. posted exact reproduction steps in our internal security channel but can't share publicly - basically the guardrail decorators get removed somewhere in the tool call pipeline before execution

2