Metrale AI Blueprint Store
Inference deployments, specified down to the kernel.
A blueprint (.kbp) is a serialized specification of one serve, designed to be signed: the model as a circuit graph, the kernel that runs every node, the precision at each node and boundary, the hardware it targets and, where a record backs it, its measured performance. Blueprints replace recipes.
Browse the catalogueWhy blueprints?
- Blueprints
- 4
- Hardware classes
- 2
- Distinct kernels
- 57
- Plan arms
- 72
Preview catalogue. These blueprints are generated from the engine's circuit planner ahead of
met blueprint export. None is signed yet, so every card reads Preview — unsigned, and no .kbp file can be downloaded. Performance appears only where a measured record backs it.4 blueprints
-
GB10 ×3Preview — unsigned
GLM-5.3-Flash-NVFP4
glm-5.3-flash-nvfp4-tp3-512k- Hardware
- NVIDIA GB10 (DGX Spark) sm_121f
- Precision
nvfp4/g16fp8/channelbf16- Kernels
- 34 in 20 families
- Plan arms
- 17 1,411 launches / decode step
50.52 tok/s C=1 decode, 1k context preliminaryFingerprint · preliminary- record
9e928d6097467ec8- git
6d845c0ee- binary
377d5214f286c650- argv
e17fe83b2704457e- env
07768e9f8fd78df2
-
GB10Preview — unsigned
Qwen3.6-35B-A3B-FP8
qwen3.6-35b-a3b-fp8-bf16head- Hardware
- NVIDIA GB10 (DGX Spark) sm_121f
- Precision
fp8/block128x128bf16- Kernels
- 20 in 16 families
- Plan arms
- 16 573 launches / decode step
No measured record attached yet -
H100-SXMPreview — unsigned
Qwen3.6-35B-A3B-FP8
qwen3.6-35b-a3b-fp8-bf16head- Hardware
- NVIDIA H100 SXM5 80GB sm_90a
- Precision
fp8/block128x128bf16- Kernels
- 20 in 16 families
- Plan arms
- 16 573 launches / decode step
No measured record attached yet -
GB10Preview — unsigned
Qwen3.8-27B-NVFP4
qwen3.8-27b-nvfp4-unsloth- Hardware
- NVIDIA GB10 (DGX Spark) sm_121f
- Precision
nvfp4/g16bf16- Kernels
- 19 in 13 families
- Plan arms
- 23 755 launches / decode step
No measured record attached yet
No blueprint matches
Clear a filter or the search to see the whole catalogue.