DEV.co
Open-Source LLM · moonshotai

Kimi-Linear-48B-A3B-Instruct

Kimi-Linear-48B-A3B-Instruct is a 48-billion parameter open-source language model from Moonshot AI that uses a hybrid linear attention architecture (Kimi Delta Attention) to achieve efficient long-context processing. It claims to match or exceed full-attention models while reducing KV cache by 75% and offering up to 6× faster decoding for contexts up to 1M tokens. The model is MIT-licensed, ungated, and trainable on commodity hardware.

Source: HuggingFace — huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct
49.1B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
71.9k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developermoonshotai
Parameters49.1B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads71.9k
Likes565
Last updated2025-12-16
Sourcemoonshotai/Kimi-Linear-48B-A3B-Instruct

What Kimi-Linear-48B-A3B-Instruct is

The model implements Kimi Delta Attention (KDA), a refined gated linear attention mechanism combined with a 3:1 KDA-to-global MLA (Multi-head Latent Attention) hybrid ratio. It is trained on 5.7T tokens. The card reports 49B total parameters with 3B activated parameters per token. Context length is claimed as 1M but not explicitly stated in the HuggingFace metadata. Inference requires Python ≥3.10, PyTorch ≥2.6, and fla-core ≥0.4.0. Supports vLLM deployment with OpenAI-compatible API.

Quickstart

Run Kimi-Linear-48B-A3B-Instruct locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="moonshotai/Kimi-Linear-48B-A3B-Instruct")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Long-document processing and retrieval

Ideal for RAG, document summarization, and long-form analysis where traditional attention becomes prohibitively expensive. 1M token context and reduced KV cache make it cost-effective for large-scale document workloads.

High-throughput inference services

Suited for production APIs where decoding latency (TPOT) and throughput matter. 6× speedup on long contexts and lower memory footprint reduce infrastructure costs for real-time applications.

Fine-tuned domain applications

Good candidate for LoRA/QLoRA adaptation given 48B base size and efficient attention mechanism. Suitable for building custom LLM applications without massive computational overhead.

Running & fine-tuning it

ESTIMATE: ~96 GB VRAM for full-precision (fp32) inference on 48B parameters; ~48 GB for fp16/bfloat16 (typical). KV cache reduction (75% claimed) will be most beneficial at very long sequence lengths. Exact VRAM varies by context length and batch size; test with your workload. Multi-GPU deployment (tensor parallelism) recommended for production; vLLM config shown uses 4 GPUs.

Card does not explicitly detail LoRA, QLoRA, or full fine-tuning support. Given 48B size and custom KDA kernel, LoRA feasibility likely but not guaranteed without testing. Custom attention kernel may complicate gradient computation or optimization. Requires careful adapter placement. Recommend contacting Moonshot AI or consulting FLA repository for verified fine-tuning recipes.

When to avoid it — and what to weigh

  • Strict latency requirements on very short sequences — Linear attention may not be optimal for short prompts where full-attention models and well-optimized kernels excel. Card shows good short-context performance but trade-offs exist.
  • Resource-constrained edge or mobile deployment — At 48B parameters, even with 3B activated params, this requires substantial GPU/TPU resources. Not suitable for consumer devices or very low-power environments.
  • Guaranteed compatibility with existing LLM frameworks — Model requires trust_remote_code=True and fla-core custom kernel support. May face integration friction with frameworks that do not yet support KDA or prohibit custom ops.
  • Use cases requiring proven, long-term stability — Model was released December 2025 (recent). Limited production track record and real-world deployment data compared to established models like GPT-4, Claude, or LLaMA 2/3.

License & commercial use

MIT license. This is a permissive, OSI-approved open-source license that permits unrestricted use, modification, and redistribution for both commercial and private purposes.

MIT license explicitly permits commercial use without further approval or attribution requirement. However, ensure you review the actual license text and any disclaimers in the repository. As is standard for OSI licenses, the software is provided AS-IS with no warranty. If using custom kernels or dependencies (e.g., fla-core), verify their licenses separately.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceMedium
Security considerations

Model requires trust_remote_code=True, enabling execution of custom Python code from the repository. This is standard for transformers with custom ops but poses supply-chain risk if repository is compromised. Verify kernel code before use in sensitive environments. No public security audit or adversarial robustness data provided. Standard LLM risks (prompt injection, jailbreaking) are Unknown.

Alternatives to consider

LLaMA 3.1 or Mistral 8×22B

Larger community adoption, proven production stability, simpler deployment without custom kernels. Trade-off: higher memory footprint and slower long-context inference.

Qwen2.5-72B

Comparable parameter count, strong performance, better supported in standard frameworks. No linear attention innovation; requires more VRAM but easier DevOps.

Claude / GPT-4 (commercial APIs)

If long-context and inference cost are not critical constraints, commercial APIs eliminate infrastructure burden. Kimi-Linear is best when you need to self-host and optimize for throughput.

Software development agency

Ship Kimi-Linear-48B-A3B-Instruct with senior software developers

Clone the repository, run the quick-start example with vLLM, and benchmark against your production constraints. MIT license removes barriers to experimentation. Contact Moonshot AI or the FLA community for production support.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Kimi-Linear-48B-A3B-Instruct FAQ

Can I use this model commercially without paying licensing fees?
Yes. MIT license permits commercial use without fees, royalties, or approval. You must retain the license notice and provide the software AS-IS. Verify any dependencies (e.g., fla-core) have compatible licenses.
How much GPU VRAM do I need?
Estimate ~48–96 GB depending on precision (bfloat16 ~48 GB, fp32 ~96 GB). Use of tensor parallelism (e.g., 4 GPUs × 40 GB) is common. Run a test with your target context length and batch size to confirm.
Does this work with standard tools like vLLM or Ollama out-of-the-box?
vLLM is confirmed via the deployment example. Ollama and llama.cpp are Unknown; they may not support the custom KDA kernel. Test before production use.
What is the actual context length?
Card claims 1M tokens (1048576 in vLLM example). This is not independently verified in the HuggingFace metadata. Confirm with the authors or benchmarking.

Custom software development services

Need help beyond evaluating Kimi-Linear-48B-A3B-Instruct? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Evaluate Kimi Linear for Your Long-Context Workload

Clone the repository, run the quick-start example with vLLM, and benchmark against your production constraints. MIT license removes barriers to experimentation. Contact Moonshot AI or the FLA community for production support.