bestmodel.run / cloud anchors — agent view
Rented hardware measured by the maintainer, log 2026-08-31. 19 runs,
two cards. Not community hardware. Honesty ladder: measured > reported >
extrapolated > formula > no data yet.
Reading conditions:
1. hardware rented by the maintainer, dated and hashed, not community submissions
2. the host image sets driver and torch; the log records what it knows of the stack
3. form factor and bandwidth are vendor specs, not pool measurements
4. the lowest reading of each multimodal batch is its first; every cell here is a median
5. the log is silent on disk state, host load and any rental terms
card | model | quant | basis | value | n
NVIDIA A10 24GB | Llama 3.1 8B Instruct GGUF | 4-bit | measured | 75.83 tok/s | n=3
NVIDIA L4 24GB | sd-turbo | fp16 | measured | 6.09 img/s | n=5
NVIDIA L4 24GB | Llama 3.1 8B Instruct GGUF | 4-bit | measured | 44.76 tok/s | n=3
NVIDIA L4 24GB | Qwen2.5 Coder 7B Instruct GGUF | 4-bit | measured | 47.24 tok/s | n=3
NVIDIA L4 24GB | Whisper Large v3 | fp16 | measured | 4.82 ×real | n=5
Pair to interpolate from (same model, same quant, same engine, both cards):
Llama 3.1 8B Instruct GGUF 4-bit on llama.cpp:
L4 44.76 tok/s (n=3)
A10 75.83 tok/s (n=3)
decode 1.69x against a 2.00x vendor bandwidth ratio.
A bracket to interpolate inside, not a coefficient to multiply by.
Source log: anchor run log 2026-08-31. sha256 de50e613cc5c2d8395c7443b2aeb761dfe04c0f340f440a839c905c3bf809526.
Frozen pool snapshot 2026-09-18 carries 2 of 5 anchor cells.
The site publishes no cloud plans, prices or provider options. Anchors are reference points.