Skip to content
Tool

Tensward

Tunes a local vLLM or llama.cpp server by measuring one flag change at a time

Pricing
Open source
Platforms
CLI, Linux, macOS
Listing updated
today

Similar tools

About Tensward

Tensward is an open-source AI tool for developer tools. It is available on CLI, Linux, macOS and self-hosted.

The README says Tensward starts its own copy of vLLM or llama.cpp on a local port, replays a workload built from your prompts or from chat, agent, code and RAG presets, and reads engine metrics and GPU state to name the resource that limits throughput. It then proposes a single flag change, measures it, and compares the answers before and after.

A change is kept in the project only when it is faster and the answers pass a check. Reworded answers are held back for manual review, and empty, repeating or invalid tool-call output is refused. The tool never stops or reconfigures a server you already run, and it needs the GPU memory free while measuring.

The project targets people who self-host LLMs on their own hardware and want a measured configuration rather than guesses. Support is listed for Apple Silicon and NVIDIA hardware, and the README describes the tool as free under Apache-2.0.

What it does

Single-change tuning loop
Each try offers one flag change, measures it against the previous setup, and keeps it only if it improves results.
Answer consistency check
Compares before and after answers and tool-call validity so faster but altered output is not kept silently.
Bottleneck analysis
Reports which resource limits the run, such as CPU-offloaded experts or a saturated GPU, using engine metrics.
Workload presets
Uses your own prompts or built-in chat, agent, code and RAG workloads with tool calls.
Local serving and export
Serves the kept configuration on a local endpoint and can export it as a Docker Compose, systemd or Helm setup.
Memory fit refusal
Refuses to start a configuration that does not fit GPU memory and names the process holding it.

Good for

  • Finding a faster llama.cpp flag set for a large MoE model that does not fit the GPU as-is
  • Lowering time to first token on a vLLM server that is saturated under concurrent requests
  • Checking whether a speed-up changes model answers before rolling it out
  • Producing a reproducible Docker Compose service for a locally tuned model

Pricing

Open source

Free under Apache-2.0; no paid plan mentioned in the README.

Tensward FAQ

Does Tensward modify my running server?

The README says it never stops or reconfigures your running server; it starts its own engine copy on a local port.

Which engines and hardware are supported?

vLLM and llama.cpp are supported. The README's table lists vLLM for NVIDIA GPUs only, and llama.cpp for NVIDIA, Apple Silicon and CPU.

Does it download models?

No. The README says Tensward never downloads models and tunes the one you already serve.

What is the license?

The README states Apache License 2.0.

Is Tensward free?

Yes. Tensward is open source, so you can use and self-host it for free.

What are the best alternatives to Tensward?

The closest alternatives to Tensward are laplacesscissors, Werkbank, Llmtee, PromptJury and gateway.

Where this comes from

  • Description: written from the project's README.