Open Source
MInference
[NeurIPS'24 Spotlight, ICLR'25, ICML'25] To speed up Long-context LLMs' inference, approximate and dynamic sparse calculate the attention, which reduces inference latency by up to
Our verdict
We have not reviewed MInference yet
No one at Enterprise DNA has checked this listing against its sources. The description comes from public listings gathered automatically. Any fact below shows where it came from and when it was last checked.
From the listing
About MInference
MInference is a framework that speeds up long-context large language model inference by approximating the attention computation with dynamic sparse patterns. It targets the pre-filling phase and can reduce latency by up to 10x on an A100 GPU while preserving accuracy. The tool is implemented in Python and open source under the Microsoft organization.
Best for
Developers optimizing long-context LLM inference on NVIDIA GPUs
Use cases
- Accelerating pre-filling for long-context LLM inference pipelines
- Reducing latency in applications that process large input sequences
- Evaluating sparse attention as a drop-in replacement for full attention
Notes
1,217 stars on GitHub. Last updated 2026-04-08. Licensed MIT.
Pros
- Up to 10x latency reduction for pre-filling on A100 hardware
- Maintains accuracy despite sparse approximation
- Open source with a published NeurIPS spotlight paper
Cons
- Optimizations are limited to the pre-filling phase, not generation
- Requires integration into existing inference codebases
- Performance gains depend on specific model architectures and hardware
Key facts
What we have checked
- Latest version
- 0.1.6 (released 17 June 2025)
- GitHub stars
- 1,230
Each value shows when we last confirmed it and where it came from. Values we have not confirmed are left out, never guessed.
Stars and usage
How MInference is growing
1,217 to 1,230
Alternatives
Alternatives to MInference
| Record | Type | How we know | Checked |
|---|---|---|---|
| DeepSpeedThis is an alternative to it | Open source | Named in the listing | Not dated |
| SGLangThis is an alternative to it | Open source | Named in the listing | Not dated |
| TensorRT-LLMThis is an alternative to it | Open source | Named in the listing | Not dated |
| vLLMThis is an alternative to it | Open source | Named in the listing | Not dated |
Compare
Compare MInference side by side
Sources
Where this comes from
- GitHubOfficial
- github api via nightly crawl, recovered from git history 20e874270887c1008f860ee0198e461a23c9e438Registry
A source shows what was published when we checked it. It does not certify security, performance or a working installation. Read our method.
Changelog
What changed
Released 0.1.6
Released 0.1.5.post1
Released 0.1.5
Released 0.1.4.post4
Released 0.1.4.post3
Released 0.1.4.post2
Released 0.1.4.post1
Released 0.1.4
Released 0.1.3
Released 0.1.2
Open-source & AI alternatives
Swap-in tools that solve the same job. Weigh the trade-offs before you commit.
DeepSpeed
Community
DeepSpeed is a deep learning optimization library that makes distributed training and inference easy, efficient, and effective.
SGLang
Community
SGLang is a high-performance serving framework for large language models and multimodal models.
TensorRT-LLM
Community
TensorRT LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and supports state-of-the-art optimizations to perform inference efficiently on NV
vLLM
Community
A high-throughput and memory-efficient inference and serving engine for LLMs
Pairs with
Other entries in the index that connect to this one. Click through to see the chain.
Get the free Developer’s Field Guide
A 27-page field guide to the AI coding workflow with Claude. Claude Code, MCP servers, the prompt patterns that work, and what to delegate. Free.
Enter your work email. We send it straight over, plus a few short notes worth knowing. Unsubscribe any time.