LocalClaw/LLM/Ling-2.6-flash (104B MoE)
Open-weight local LLM

Ling-2.6-flash (104B MoE)

InclusionAI's MIT-licensed instruct MoE optimized for fast agent workloads. 104B total parameters, only 7.4B active, hybrid linear attention, 262K context and strong tool-use / multi-step execution with high token efficiency.

Large-memory workstation 80 GB RAM Q4_K_M Agentic coding and codebase automation
Run with LocalClaw Compare all models
Parameters
104B (7.4B active)
Minimum RAM
80 GB
Model size
65 GB
Quantization
Q4_K_M

Can Ling-2.6-flash (104B MoE) run locally?

Ling-2.6-flash (104B MoE) needs a serious workstation with large unified memory or high VRAM.

Search for Ling-2.6-flash in LM Studio or another GGUF-compatible runtime.

inclusionAI/Ling-2.6-flash
chatcodereasoningspeedquality

Install path

01
Check RAM fitMinimum 80 GB RAM. Start with the Q4_K_M quant.
02
Load the modelSearch Ling-2.6-flash in LM Studio.
03
Control locallyUse LocalClaw to manage models, agents, chat, channels and scheduled OpenClaw work.

Strengths

  • ⚡ Very efficient MoE: 104B total but only 7.4B active parameters
  • 262K context window for large prompts, repositories and agent traces
  • Hybrid linear attention improves prefill/decode throughput on long workloads
  • Strong tool-use and agent execution focus: BFCL-V4, TAU2-bench, SWE-bench Verified, Claw-Eval and PinchBench are highlighted by InclusionAI
  • Designed to reduce unnecessary reasoning tokens for high-frequency production agent tasks
  • MIT license — permissive open-source usage

Limitations

  • Large 104B MoE files — server-grade hardware is still recommended despite only 7.4B active parameters
  • Best support currently targets SGLang/vLLM with trust_remote_code rather than simple consumer one-click runtimes
  • Official notes mention possible tool hallucinations in highly complex scenarios
  • Natural Chinese/English switching and highly complex instruction following still have room to improve

Best use cases

  • Agentic coding and codebase automation
  • Tool-calling assistants and multi-step workflow execution
  • Long-context analysis up to 262K tokens
  • Production agents where token efficiency and latency matter
  • Open-source research comparing efficient MoE architectures
  • OpenClaw / Claude Code / Kilo Code style agent workflows

Capability profile

speed
8
quality
9
coding
9
reasoning
8

Technical notes

Developer
InclusionAI / Ant Group
License
MIT
Context window
262,144 tokens
Architecture
Sparse MoE instruct model — 104B total parameters, 7.4B active. Hybrid linear attention using a 1:7 MLA + Lightning Linear architecture, optimized for high-throughput agent inference.

This model fits these next steps

Hardware fit is based on LocalClaw's RAM tier, model size and quantization metadata. Always leave memory headroom for your OS and runtime.

Large unified memoryMac Studio M4 Max 128GB128GB AI PC classNVIDIA GB10 / DGX SparkHigh-memory picks128GB RAM guide

Similar models to compare

Ling 1T (1T MoE) 1T (50B active)MiniMax M2 (230B MoE) 230B (10B active)Kimi K2.5 (32B/1T MoE) 32B active (1T total MoE)DeepSeek V3.2 (37B/671B MoE) 37B (671B MoE)Qwen 3.6 (27B) 27B

Where to go next