DeepSeek-V3.2-Speciale

685B parameters, text in, text out.

Verified 2026-10-25 Rank #704 in LLMs · score 72 method v1 Sources 3 Downloads 30d 3.6k · #563 of 619 tracked

Can I run it?411GB at INT4Estimateneeds a multi-GPU nodemethod

Can I use it?MITPermissivelicence file

Do people use it?3.6kdownloads, 30 days#563 of 619 we track · below the median+4 likes, downloads +1% this week743 likesHugging Face

Worth reading

Picked by rule from 19 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 309 HN pts

DeepSeek-v3.2-Exp

DeepSeek-V3.2-Exp introduces DeepSeek Sparse Attention to optimize long-context efficiency while maintaining performance on par with V3.1-Terminus.

github.com 2025-09-29 309 HN pts · 50 comments 537 words about the DeepSeek V3.2 family family

Latest · 3 months ago

DeepSeek-AI/DeepSeek-v3.2

DeepSeek-V3.2 is an open-weights language model for text generation published by deepseek-ai and added to Hugging Bay Jun 29, 2026.

The page explicitly states that file-size fit is an estimate and not a successful run or speed promise.
A recorded SHA-256 checksum fingerprints only the named file bytes and does not prove safety or quality.

huggingbay.xyz 2026-07-11 2 HN pts · 3 comments 766 words about the DeepSeek V3.2 family family

19 best of 19 tracked

The 2 picks above, plus these 17. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleDeepSeek-V3.2 Releaseapi-docs.deepseek.com · 2025-12-01 · 14 HN pts · official
ArticleDeepSeek-v3.2 on GB300: Performance Breakthroughblog.vllm.ai · 2026-02-25 · 2 HN pts
ArticleDeepSeek-v3.2-Expapi-docs.deepseek.com · 2025-10-04 · 3 HN pts · official
ArticleDeepSeek-v3.2simonwillison.net · 2025-12-02 · 7 HN pts
ArticleDeepSeek v3.2 Is Okay and Cheap but Slowthezvi.substack.com · 2025-12-05 · 4 HN pts
CodeDeepSeek-v3.2-Exp: Long-Context Efficiency with DeepSeek Sparse Attention [pdf]github.com · 2025-09-29 · 4 HN pts
ArticleDeepSeek-V3.2: Pushing the Frontier of Open Large Language Modelscas-bridge.xethub.hf.co · 2025-12-08 · 7 HN pts
ArticleGPT-5.2, Grok 4.1, and DeepSeek v3.2 compare as Santa agentsveris.ai · 2025-12-19 · 4 HN pts
ArticleDeepSeek-v3.2: Pushing the Frontier of Open Large Language Modelscas-bridge.xethub.hf.co · 2025-12-01
DiscussionLaunching DeepSeek-v3.2 and DeepSeek-v3.2-Specialetwitter.com · 2025-12-01
ArticleThe DeepSeek v3.2 Breakthrough Simplifiedtripplyons.com · 2025-09-30
ArticleDeepSeek-v3.2: Pushing the Frontier of Open Large Language Modelscas-bridge.xethub.hf.co · 2025-12-01
ArticleDeepSeek-v3.2cas-bridge.xethub.hf.co · 2025-12-01
ArticleDeepSeek-v3.2: Pushing the Frontier of Open Large Language Modelscas-bridge.xethub.hf.co · 2025-12-01
ArticleDeepSeek-v3.2cas-bridge.xethub.hf.co · 2025-12-01
ArticleDeepSeek v3.2 Technical Reportcas-bridge.xethub.hf.co · 2025-12-01
DiscussionDeepSeek-v3.2-Exptwitter.com · 2025-09-29

Show all 17

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 3 quantized builds.

As a service0no API provider listed
On own hardware3quantized builds
Built on1411 fine-tunes · 3 adapters · 0 merges
In apps0Hugging Face Spaces

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Needs an 8×80 GB node even at INT4

About 411 GB at INT4, 823 GB at FP8, 1.65 TB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2025-11-28), amber triangles are 10 dated articles (each links to it).

What people built on it

17 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

17 derivatives on the Hub, most of them fine-tunes

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF ubergarm/DeepSeek-V3.2-Speciale-GGUF for llama.cpp, Ollama, LM Studio 3.9k downloads MLX mlx-community/DeepSeek-V3.2-Speciale-4bit for Apple silicon 5.1k downloads AWQ QuantTrio/DeepSeek-V3.2-Speciale-AWQ for vLLM, SGLang 796 downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Facts

14 fields

Listed in LLMs · MoE

Facts

DeepSeek-V3.2-Speciale has 685B parameters. (safetensors metadata, observed —)
DeepSeek-V3.2-Speciale is a mixture-of-experts model, routing 8 per token. (config.json, observed —)
License: MIT (friction: none). Permissive. (license file, observed —)
Created 2025-11-28, last modified 2025-12-01. (HF API, observed —)

Pipeline tag	text-generation
Library	transformers
Architectures	DeepseekV32ForCausalLM
Model type	deepseek_v32
Params by dtype	BF16: 3,946,184,704 · F32: 42,667,040 · F8_E4M3: 681,408,069,632
Size band	XXL
MoE	yes
Experts per token	8
License	MIT (mit)
Hub snapshot	2026-10-25

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for DeepSeek-V3.2-Speciale: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	1.65 TB · Cluster (≥640 GB)
FP8 (1 byte/param)	823 GB · Cluster (≥640 GB)
INT4 (0.5 bytes/param)	411 GB · Multi-GPU node (<640 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

17 variants

Lineage

Fine-tuned from deepseek-ai/DeepSeek-V3.2-Exp-Base (same organization) Family: deepseek-ai/DeepSeek-V3.2 17 derivatives on the Hub

Base model: deepseek-ai/DeepSeek-V3.2-Exp-Base

finetune: 11quantized: 3adapter: 3

Top variants (17 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
mlx-community/DeepSeek-V3.2-Speciale-4bit	finetune	-4bit	672B	XXL	mit	mlx-community	1	5.1k
ubergarm/DeepSeek-V3.2-Speciale-GGUF	quantized	GGUF	—	—	mit	ubergarm	20	3.9k
QuantTrio/DeepSeek-V3.2-Speciale-AWQ	quantized	AWQ	685B	XXL	mit	QuantTrio	4	796
DevQuasar/deepseek-ai.DeepSeek-V3.2-Speciale-Channel-INT8	finetune	INT8	686B	XXL	mit	DevQuasar	1	883
amd/DeepSeek-V3.2-Speciale-mtp-ptpc	quantized	—	686B	XXL	mit	amd	0	357
gallantpigeon/DeepSeek-V3.2-Speciale-GGUF-preview	finetune	GGUF	—	—	—	gallantpigeon	0	75
DevQuasar/DeepSeek-V3.2-Speciale-BF16	finetune	-BF16	685B	XXL	—	DevQuasar	1	0
OrbitngTwice/LeeSunwooModel	finetune	—	—	—	unknown	OrbitngTwice	0	0
donerci/elektronik_ticaret	finetune	—	—	—	openrail	donerci	0	0
7Mohamed-Gamal7/MOhamed_AI	adapter	—	—	—	apache-2.0	7Mohamed-Gamal7	0	0
Yooghrn/Cincau3	adapter	—	—	—	creativeml-openrail-m	Yooghrn	0	0
cewaw/3D	adapter	—	—	—	mit	cewaw	0	0
Peeacfully/Peace	finetune	—	—	—	—	Peeacfully	0	0
potgieternicojames/moodlequiz	finetune	—	—	—	—	potgieternicojames	0	0
Thierry-Zhang/Mathbot	finetune	—	—	—	mit	Thierry-Zhang	0	0
ajay1706/My_test	finetune	—	—	—	apache-2.0	ajay1706	0	0
VisionStudios/Erebus5	finetune	—	—	—	—	VisionStudios	0	0

Repository changelog

1 change

changed in the last 7 daysNew on the watch list: 2026-10-23all changes

Articles and coverage

19 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official2

DeepSeek-V3.2 Release

api-docs.deepseek.com2025-12-01officialauthorityabout the DeepSeek V3.2 familyHN 14 pts · 0 comments217 words

DeepSeek launches V3.2 and V3.2-Speciale, reasoning-first models built for agents with V3.2-Speciale attaining gold-level results in IMO, CMO, ICPC World Finals & IOI 2025. The new models integrate thinking directly into tool-use after training on 1,800+ environments and 85k+ complex instructions. V3.2-Speciale is API-only until Dec 15th, 2025, 15:59 (UTC Time) and requires higher token usage for complex tasks.written, reviewed, cited

DeepSeek-v3.2-Exp

api-docs.deepseek.com2025-10-04officialauthorityabout the DeepSeek V3.2 familyHN 3 pts · 0 comments137 words

Analysis8

DeepSeek-v3.2-Exp

github.com2025-09-29authorityHN front pageabout the DeepSeek V3.2 familyHN 309 pts · 50 comments537 words

DeepSeek-V3.2-Exp introduces DeepSeek Sparse Attention to optimize long-context efficiency while maintaining performance on par with V3.1-Terminus. The model scores 85.0 on MMLU-Pro and 89.3 on AIME 2025, with a 2025.11.17 fix resolving a Rotary Position Embedding discrepancy in the indexer module. Weights are licensed under the MIT License and supported by vLLM and SGLang.written, reviewed, cited

DeepSeek-AI/DeepSeek-v3.2

huggingbay.xyz2026-07-11about the DeepSeek V3.2 familyHN 2 pts · 3 comments766 words

DeepSeek-V3.2 is an open-weights language model for text generation published by deepseek-ai and added to Hugging Bay Jun 29, 2026. The repository totals 689.5 GB with 178 hosted files available out of 190 listed, distributed in safetensors format. The model is released under the mit license, permitting commercial use, modification, and redistribution.written, reviewed, cited

The page explicitly states that file-size fit is an estimate and not a successful run or speed promise.
A recorded SHA-256 checksum fingerprints only the named file bytes and does not prove safety or quality.
The system cannot verify if the model fits a specific computer because it lacks the user's machine specs.

DeepSeek-v3.2 on GB300: Performance Breakthrough

blog.vllm.ai2026-02-25authorityabout the DeepSeek V3.2 familyHN 2 pts · 0 comments1,908 words

DeepSeek-V3.2 (NVFP4 + TP2) runs on GB300 with 288GB memory per GPU using vLLM v0.14.1 and the FlashInfer FP4 MoE kernel. It achieves 7360 TGS in prefill-only scenarios and 2816 TGS in mixed contexts (ISL=2k, OSL=1k), while DeepSeek R1 reaches ~22476 TGS in prefill. V3.2's lower throughput stems from the Indexer/Sparse MLA overhead, where the kernel execution time for a single DSA layer step is 2.7x that of MLA.written, reviewed, cited

DeepSeek-v3.2

simonwillison.net2025-12-02authorityabout the DeepSeek V3.2 familyHN 7 pts · 0 comments333 words

DeepSeek v3.2 Is Okay and Cheap but Slow

thezvi.substack.com2025-12-05authorityabout the DeepSeek V3.2 familyHN 4 pts · 0 comments2,504 words

DeepSeek-v3.2-Exp: Long-Context Efficiency with DeepSeek Sparse Attention [pdf]

github.com2025-09-29authorityabout the DeepSeek V3.2 familyHN 4 pts · 0 comments

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

cas-bridge.xethub.hf.co2025-12-08about the DeepSeek V3.2 familyHN 7 pts · 1 comments

GPT-5.2, Grok 4.1, and DeepSeek v3.2 compare as Santa agents

veris.ai2025-12-19about the DeepSeek V3.2 familyHN 4 pts · 2 comments1,982 words

All coverage tracked for this model (19) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
DeepSeek-AI/DeepSeek-v3.2	huggingbay.xyz	2026-07-11	2026-10-14	hn
DeepSeek-v3.2 on GB300: Performance Breakthrough	blog.vllm.ai	2026-02-25	2026-10-14	hn
GPT-5.2, Grok 4.1, and DeepSeek v3.2 compare as Santa agents	veris.ai	2025-12-19	2026-10-14	hn
DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models	cas-bridge.xethub.hf.co	2025-12-08	2026-10-14	hn
DeepSeek v3.2 Is Okay and Cheap but Slow	thezvi.substack.com	2025-12-05	2026-10-14	hn
DeepSeek-v3.2	simonwillison.net	2025-12-02	2026-10-14	hn
DeepSeek-v3.2: Pushing the Frontier of Open Large Language Models	cas-bridge.xethub.hf.co	2025-12-01	2026-10-14	hn
DeepSeek-v3.2	cas-bridge.xethub.hf.co	2025-12-01	2026-10-14	hn
DeepSeek-v3.2: Pushing the Frontier of Open Large Language Models	cas-bridge.xethub.hf.co	2025-12-01	2026-10-14	hn
DeepSeek-v3.2	cas-bridge.xethub.hf.co	2025-12-01	2026-10-14	hn
DeepSeek v3.2 Technical Report	cas-bridge.xethub.hf.co	2025-12-01	2026-10-14	hn
Launching DeepSeek-v3.2 and DeepSeek-v3.2-Speciale	twitter.com	2025-12-01	2026-10-14	hn
DeepSeek-v3.2: Pushing the Frontier of Open Large Language Models	cas-bridge.xethub.hf.co	2025-12-01	2026-10-14	hn
DeepSeek-V3.2 Release	api-docs.deepseek.com official	2025-12-01	2026-10-14	hn
DeepSeek-v3.2-Exp	api-docs.deepseek.com official	2025-10-04	2026-10-14	hn
The DeepSeek v3.2 Breakthrough Simplified	tripplyons.com	2025-09-30	2026-10-14	hn
DeepSeek-v3.2-Exp: Long-Context Efficiency with DeepSeek Sparse Attention [pdf]	github.com	2025-09-29	2026-10-14	hn
DeepSeek-v3.2-Exp	twitter.com	2025-09-29	2026-10-14	hn
DeepSeek-v3.2-Exp	github.com	2025-09-29	2026-10-14	hn

Similar models

3 models

Similar models

DeepSeek-R1 · 684B
same size band (XXL), also MoE, same task (text-generation), more providers, permissive license
GLM-5.3 · 753B
same size band (XXL), also MoE, same task (text-generation), more providers
GLM-5.2 · 753B
same size band (XXL), also MoE, same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-14 14:08 UTC · Last verified 2026-10-25

Follow DeepSeek-V3.2-Speciale

Watch this model

/feeds/models/deepseek-ai--DeepSeek-V3.2-Speciale.xml(RSS 2.0) · every change to DeepSeek-V3.2-Speciale
/feeds/families/deepseek-ai--DeepSeek-V3.2.xml(RSS 2.0) · every change across the DeepSeek-V3.2 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.