# lightseekorg/tokenspeed

TokenSpeed is a speed-of-light LLM inference engine.

Repository: https://github.com/lightseekorg/tokenspeed
Canonical: https://ross.abutalabs.com/products/tokenspeed
Homepage: https://lightseek.org/blog/lightseek-tokenspeed.html
Language: Python
License: MIT
License Family: permissive
Topics: deepseek, gpt-oss, kimi, lightseek, llm, minimax, qwen, speed-of-light, tokenspeed, blackwell, glm, vlm, inkling, b200, b300, gb200, gb300, mi355x, mi455x
Last push: 2026-08-26T20:09:14+00:00

## Health v2 (maintenance only)
Score: 68/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 62, longevity 8
- inputs: {"age_days": 119, "days_push": 7, "days_rel": 40, "gap_med": null, "n_releases_24m": 1}
- flags: young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1989, forks 256 (observed 2026-08-28T04:06:01.703616+00:00)

## What it is
TokenSpeed is a high-performance LLM inference engine designed for agentic workloads, aiming for TensorRT-LLM-level performance with vLLM-level usability. It features a compiler-backed SPMD modeling layer, a C++/Python scheduler with type-safe KV cache reuse, pluggable kernels, and an AsyncLLM entrypoint.

## Use cases
- serve large language models at maximum throughput
- run inference for agentic coding workloads
- deploy frontier models like Qwen, Kimi, DeepSeek, and GLM on day 0
- run FP4 inference on NVIDIA Blackwell and AMD accelerators
- avoid hand-writing tensor parallelism and collective communication code
- serve multi-hundred-billion parameter MoE models efficiently

## When to choose
- you need maximum tokens-per-second for production agentic LLM serving
- you run modern datacenter GPUs like NVIDIA B200/GB300 or AMD MI355X/MI455X
- you want vLLM-like usability with compiler-generated parallelism
- you need day-0 support for the latest frontier open models

## When to avoid
- you only need lightweight local inference on consumer hardware
- you need broad CPU-only or edge-device support
- you require a long-stable, battle-tested engine with extensive ecosystem tooling
- your stack depends on features specific to vLLM or TensorRT-LLM

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, gpu-computing, machine-learning, compiler, concurrency
- domain: large-language-models, machine-learning, gpu-computing, performance
- platform: python
- tags: inference-engine, tensor-parallelism, spmd, kv-cache, blackwell, agentic-workloads, vllm-alternative, tensorrt-llm, ai-agents, gpu, linux, docker

## Member repositories
- lightseekorg/tokenspeed (main) score 68

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:01.703616+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:02:56.214781+00:00, confidence not recorded.
  - readme: https://github.com/lightseekorg/tokenspeed (fetched 2026-08-28T04:06:01.703616+00:00, sha a0f354b1464b)
  - homepage: https://lightseek.org/blog/lightseek-tokenspeed.html (fetched 2026-08-29T10:42:38.210963+00:00, sha 906baef7cc38)
  - registry_pypi: https://pypi.org/pypi/tokenspeed/json (fetched 2026-08-29T10:42:38.220527+00:00, sha ab516c07b37f)
- Data as of 2026-08-30T08:39:29.467469+00:00.