The lowest memory cap at which the serving unit starts and answers a request, with the unit's own settings (262,144-token context, fp8 KV cache, speculative decoding on), on one 96 GB card.
License
Apache-2.0
Tileward PersonaPlex 7B
Compressed from NVIDIA PersonaPlex.
Parameters7BDense
Context—
EncodingTWX4Tileward
Size—
VRAM18.1 GBone conversation
Price$0.08per minute
Tileward-PersonaPlex-7b
Capabilities
Voice
VRAM
The peak GPU memory with one conversation, measured with NVIDIA's own per-frame calls on an L40S and on an A10G.
Size1,007.9 GB35% smaller than Kimi K3 as published (1.5×)
VRAM136.7 GBweights on each of 8 cards
PriceOn request
Tileward-Kimi-K3-TWC25
VRAM
What the weights took on each of 8 NVIDIA B300 cards when the serving engine loaded them, as it reported it, with the model split across the cards. The cache and the working memory come on top.
Size1,178.1 GB25% smaller than Kimi K3 as published (1.3×)
VRAM157.9 GBweights on each of 8 cards
PriceOn request
Tileward-Kimi-K3-TWMX34
VRAM
What the weights took on each of 8 NVIDIA B300 cards when the serving engine loaded them, as it reported it, with the model split across the cards. The cache and the working memory come on top.
Rates are list prices, before a plan’s discount. Sizes are weights on disk, in decimal gigabytes. VRAM is the memory a model needs to serve on one card, in decimal gigabytes: the line under each figure is what it was measured at, and a row’s details say how. A served model is called by its id at https://api.tileward.com/v1, with any OpenAI-compatible client: the docs show how.
Need a model that is not served today? We can serve it for you, hosted or on your hardware.