LocalClaw/LLM/DeepSeek V3.2 (37B/685B MoE)
Open-weight MoE

DeepSeek V3.2 (37B/685B MoE)

Catalogue summary: DeepSeek's server-grade MoE flagship. Hugging Face reports 685B total parameters; the verified Unsloth IQ4_XS GGUF is about 358.3 GB. Active parameters affect compute per token, not the amount of model weights that must be loaded. MIT licensed.

Repository editorial metadata; verify comparative claims in the linked upstream material.

448 GB catalogue minimum448 GB RAMIQ4_XSCoding assistant
Choose an app

Only verified options can be opened. Unsupported apps are clearly marked.

Compare models
LM StudioNot available for this model
Open in UnslothLaunches Unsloth Desktop on this model's download page
OllamaNot available for this model
Open on Hugging FaceFiles, licence and available downloads
llama.cppNot available for this model
Use with LocalClawOptional workspace after the model is installed

Desktop app links require the app to be installed. If nothing opens, LocalClaw will show app-download and model-file fallbacks.

Parameters
685B total (37B active, MoE)
Minimum RAM
448 GB
Model size
358.3 GB
Quantization
IQ4_XS

Can DeepSeek V3.2 (37B/685B MoE) run locally?

DeepSeek V3.2 (37B/685B MoE) has a catalogue minimum of 448 GB RAM with IQ4_XS. Actual memory use and speed vary by context length, runtime, backend and system headroom.

Treat deepseek-v3.2 as a server-grade catalogue target. Use the verified artefact with a compatible multi-GPU or distributed runtime and follow its upstream instructions; this is not a one-click desktop LM Studio recommendation.

Model sourcePublic GGUF repository (verified 2026-09-07): unsloth/DeepSeek-V3.2-GGUFUpstream releaseLicence text
chatcodereasoningpowerqualitygeneralserver-grademoe

Deployment path

01
Check RAM fitServer-grade target. Plan for 448 GB class multi-GPU memory.
02
Load the modelUse deepseek-v3.2 only with a compatible server-grade or distributed runtime; confirm the exact artefact and upstream instructions first.
03
Control locallyUse LocalClaw to manage models, agents, chat, channels and scheduled OpenClaw work.

Catalogue record

  • Family: deepseek-v3
  • Parameters: 685B total (37B active, MoE)
  • Recommended quantization: IQ4_XS
  • Catalogue minimum RAM: 448 GB
  • Catalogue model size: 358.3 GB
  • Tags: chat, code, reasoning, power, quality, general, server-grade, moe

Practical limits

  • Catalogue RAM is a minimum estimate, not a guarantee for every context length or runtime.
  • Speed and memory use vary by quantization, backend, context length and system headroom.
  • Verify architecture, licence and usage restrictions in the linked upstream material before deployment.

Catalogue tags

  • chat
  • code
  • reasoning
  • power
  • quality
  • general

Capability profile

Repository catalogue ratings used by LocalClaw's editorial rubric. They are not a standardized third-party benchmark.

speed
3
quality
10
coding
10
reasoning
10

Technical notes

Developer
DeepSeek AI
License
MIT
Context window
131,072 tokens
Architecture
Mixture of Experts (MoE), 685B total and ~37B active

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.

Very large memoryMac Studio Ultra classCheck model size firstNVIDIA GB10 / server optionsMore practical alternativesCompare smaller models

Related catalogue entries

Linked mechanically by family, shared tags and nearby RAM tier; this is not a quality ranking.

DeepSeek V3 (671B MoE) 671B (37B active)DeepSeek V3.1 (671B MoE) 671B (37B active, MoE)Trinity Large Preview (398B MoE) 398B total (13B active, MoE)DeepSeek V3.2 Exp (671B MoE) 671B (37B active)Hermes 4 (405B) 405BGLM 4.6 (355B MoE) 355B (32B active)DeepSeek R1 0528 (671B MoE) 671B (37B active)Llama 4 Maverick (17B/400B MoE) 400B (17B active, 128 experts)

Where to go next