Celva
Inference research

Open models on frontier chips.

We optimize and run open models on frontier AMD and NVIDIA chips, starting with GLM 5.2. We test builds, runtimes, kernels, batching, and placement against traffic.

Method

We tune the serving stack.

We test the whole path: model, runtime, kernels, chips, and traffic.

01

Profile

Traffic, context length, concurrency, and the tail-latency budget.

ARRIVAL · CONTEXT · CONCURRENCY · TAIL LATENCY
02

Search

Model build, runtime, kernels, batching, parallelism, and placement.

BUILD · RUNTIME · KERNELS · BATCHING · CHIPS
03

Validate

Replay on target hardware. Record throughput, p99 ITL, and cost.

REPLAY · THROUGHPUT · P99 ITL · COST
Systems

Model. Runtime. Kernels. Chips.

One system, measured against traffic.

Model
open weights / quality constraints

Quality and context remain part of the objective.

Runtime
engine / scheduler

Engine, scheduler, KV cache, attention kernels, and batching change the result.

Chips
target hardware

Hardware is part of the search, not a final handoff.

Workload
traffic / context / concurrency

Traffic shape sets the conditions every candidate has to meet.

Deliverable
pinned config / replay

Model build, runtime, kernel, scheduler, and placement with replay evidence.