DeepSeek-V3.1-Base

685B parameters (8 of 256 experts per token), text in, text out, 160K tokens of context.

Verified 2026-10-25 Rank #493 in LLMs · score 76.8 method v1 Sources 3 Downloads 30d 12.7k · #489 of 619 tracked

Changed this week: 1 new quantized build on the Hub (23 in all).2026-10-24

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

Can I use it?MITPermissivelicence file2026-10-24

Do people use it?12.7kdownloads, 30 days#489 of 619 we track · below the median+1 likes, downloads +0% this week1,024 likesHugging Face2026-10-24

Worth reading

Picked by rule from 8 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 41 HN pts

DeepSeek v3.1 is not having a moment

DeepSeek v3.1 is a 685-billion-parameter open-source model that unifies chat, reasoning, and coding into a single architecture with 128K context windows.

DeepSeek delayed its new model release after failing to train it using Huawei's chips.
The startup used Nvidia chips for training and Huawei's chips for inference due to technical issues.

thezvi.wordpress.com 2025-08-22 41 HN pts · 27 comments 631 words about the DeepSeek V3.1 family family

Latest · 13 months ago

DeepSeek-v3.1-Terminus

api-docs.deepseek.com 2025-09-22 101 HN pts · 28 comments 87 words about the DeepSeek V3.1 family family

8 best of 8 tracked

The 2 picks above, plus these 6. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleDeepSeek-v3.1api-docs.deepseek.com · 2025-08-21 · 778 HN pts · official
ArticleDeepSeek v3.1 is not having a momentthezvi.substack.com · 2025-08-23 · 6 HN pts
DiscussionDeepSeek v3.1 released – single model, thinking and non-thinking modestwitter.com · 2025-08-21 · 26 HN pts
ArticleDeepSeek v3.1 Nex N1openrouter.ai · 2025-12-17 · 1 HN pts
ArticleDeepSeek v3.1 just dropped – and it might be the most powerful open AI yetventurebeat.com · 2025-08-19 · 12 HN pts
ArticleDeepSeek v3.1 is Here, 685B parametersxcancel.com · 2025-08-20 · 4 HN pts

Show all 6

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: 23 quantized builds and 16 Spaces.

As a service0no API provider listed
On own hardware23quantized builds
Built on2417 fine-tunes · 6 adapters · 1 merges
In apps16Hugging Face Spaces

Top Spaces running it: jairwaal/image ♥ 16 · Jakob08/moneychatbot ♥ 3 · ajay5364747/Ajay ♥ 2 · fokan/train-modle ♥ 2

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, 821 GB at FP8, 1.64 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 (2025-08-19), amber triangles are 8 dated articles (each links to it), grey ticks are repo commits.

What people built on it

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

47 derivatives on the Hub, led by 23 quantized

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

GGUF bartowski/deepseek-ai_DeepSeek-V3.1-Base-GGUF for llama.cpp, Ollama, LM Studio 12.3k downloads

All variants →

Reference

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

Benchmarks

22 rows

Benchmarks 22 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (22)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported AIME 2024 · Reason through hard problems	66.3 pass@1	DeepSeek V3.1-NonThinking	model author	2026-10-07
author-reported AIME 2024 · Reason through hard problems	93.1 pass@1	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported AIME 2025 · Reason through hard problems	49.8 pass@1	DeepSeek V3.1-NonThinking	model author	2026-10-07
author-reported AIME 2025 · Reason through hard problems	88.4 pass@1	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported Aider Polyglot · Write code with agents	68.4	DeepSeek V3.1-NonThinking	model author	2026-10-07
author-reported Aider Polyglot · Write code with agents	76.3	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported BrowseComp	30	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported BrowseComp_zh	49.2	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported Codeforces · Write code with agents	2091 rating	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported GPQA Diamond · Reason through hard problems	74.9 pass@1	DeepSeek V3.1-NonThinking	model author	2026-10-07
author-reported HMMT · Reason through hard problems	33.5 pass@1	DeepSeek V3.1-NonThinking	model author	2026-10-07
author-reported HMMT · Reason through hard problems	84.2 pass@1	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported Humanity's Last Exam · Reason through hard problems	15.9 pass@1	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported Humanity's Last Exam · Reason through hard problems	29.8	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported LiveCodeBench · Write code with agents	56.4 pass@1	DeepSeek V3.1-NonThinking	model author	2026-10-07
author-reported LiveCodeBench · Write code with agents	74.8 pass@1	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported MMLU · Reason through hard problems	91.8	DeepSeek V3.1-NonThinking	model author	2026-10-07
author-reported MMLU-Pro · Reason through hard problems	83.7	DeepSeek V3.1-NonThinking	model author	2026-10-07
author-reported SWE-bench Multilingual · Write code with agents	54.5	DeepSeek V3.1-NonThinking	model author	2026-10-07
author-reported SWE-bench Verified · Write code with agents	66	DeepSeek V3.1-NonThinking	model author	2026-10-07
author-reported SimpleQA · Reason through hard problems	93.4	DeepSeek V3.1-Thinking	model author	2026-10-07
author-reported Terminal-Bench · Write code with agents	31.3	DeepSeek V3.1-NonThinking	model author	2026-10-07

Use-case evidence

5 use cases

What it is good at 5 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

22 fields

Listed in LLMs · MoE

Facts

DeepSeek-V3.1-Base has 685B parameters. (safetensors metadata, observed 2026-10-24)
DeepSeek-V3.1-Base is a mixture-of-experts model with 256 experts, routing 8 per token. (config.json, observed 2026-10-24)
Context length is 163,840 tokens (160K). (config.json max_position_embeddings, observed 2026-10-24)
License: MIT (friction: none). Permissive. (license file, observed 2026-10-24)
Created 2025-08-19, last modified 2025-08-26. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	DeepseekV3ForCausalLM
Model type	deepseek_v3
Dtype	bfloat16
Params by dtype	BF16: 3,918,786,560 · F32: 41,555,600 · F8_E4M3: 680,571,043,840
Size band	XXL
MoE	yes
Experts	256
Experts per token	8
Context length	163,840 (config.json max_position_embeddings)
Papers	arXiv:2412.19437
License	MIT (mit)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · DeepseekV3ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · DeepseekV3ForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for DeepSeek-V3.1-Base: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	1.64 TB · Cluster (≥640 GB)
FP8 (1 byte/param)	821 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

40 variants

Lineage

Original release (shipped quantized by the lab) Family: deepseek-ai/DeepSeek-V3.1 47 derivatives on the Hub

quantized: 23finetune: 17merge: 1adapter: 6

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
deepseek-ai/DeepSeek-V3.1	quantized	—	685B	XXL	mit	deepseek-ai	863	3.4M
deepseek-ai/DeepSeek-V3.1-Terminus	quantized	—	685B	XXL	mit	deepseek-ai	385	272.8k
unsloth/DeepSeek-V3.1-Base-BF16	finetune	-BF16	684B	XXL	mit	unsloth	4	22.4k
bartowski/deepseek-ai_DeepSeek-V3.1-Base-GGUF	quantized	GGUF	—	—	—	bartowski	2	12.3k
bartowski/deepseek-ai_DeepSeek-V3.1-Base-Q4_K_M-GGUF	quantized	GGUF	—	—	—	bartowski	10	3.4k
IntervitensInc/DeepSeek-V3.1-Terminus-Channel-int8	finetune	int8	685B	XXL	mit	IntervitensInc	0	1.6k
AmpereComputing/deepseek-v3.1-gguf	quantized	gguf	—	—	—	AmpereComputing	0	1.4k
unsloth/DeepSeek-V3.1-Base	quantized	—	685B	XXL	mit	unsloth	5	830
gghfez/DeepSeek-V3.1-Base-256x21B-BF16	finetune	-BF16	—	—	—	gghfez	1	915
chutesai/DeepSeek-V3.1-NextN	quantized	—	11.6B	L	mit	chutesai	0	751
chutesai/DeepSeek-V3.1-Terminus-NextN	quantized	—	11.6B	L	mit	chutesai	2	560
nex-agi/DeepSeek-V3.1-Nex-N1.1	quantized	—	683B	XXL	apache-2.0	nex-agi	5	402
sparrow18/DeepSeek-V3.1-Terminus	quantized	—	684B	XXL	mit	sparrow18	0	384
gghfez/DeepSeek-V3.1-Base-IQ2_KS	quantized	GGUF	—	—	apache-2.0	gghfez	0	371
lmganon123/DeepSeek-V3.1-Base_IK_GGUF_Q2	quantized	GGUF	—	—	mit	lmganon123	1	315
zianglih/DeepSeek-V3.1-5layer	quantized	—	26.6B	L	mit	zianglih	0	311
mradermacher/DeepSeek-V3.1-Base-i1-GGUF	quantized	GGUF	—	—	mit	mradermacher	0	299
Sendpns/deepseekv3.1	quantized	—	—	—	mit	Sendpns	0	273
Saad381/deepseek-finetuned	finetune	—	1.35B	S	—	Saad381	0	248
gamm0000/DeepSeek-V3.1-Terminus	quantized	—	684B	XXL	mit	gamm0000	1	175
aiqwen/DeepSeek-V3.1-Terminus	quantized	—	685B	XXL	mit	aiqwen	0	210
aiqwen/DeepSeek-V3.1	quantized	—	685B	XXL	mit	aiqwen	0	132
lactroiii/DeepSeek-V3.1	quantized	—	684B	XXL	mit	lactroiii	0	131
williamchangtw/DeepSeek-V3.1-Nex-N1.1	quantized	—	683B	XXL	apache-2.0	williamchangtw	0	122
gergo992/DeepSeek-V3.1	quantized	—	685B	XXL	mit	gergo992	0	90
trskX/DeepSeek-V3.1-Terminus	quantized	—	685B	XXL	mit	trskX	0	76
Mbt7000/AloullA1	merge	—	—	—	mit	Mbt7000	1	0
hrm-labs/DeepSeek-V3.1	quantized	—	684B	XXL	mit	hrm-labs	0	28
mradermacher/DeepSeek-V3.1-Base-GGUF	finetune	GGUF	—	—	mit	mradermacher	0	0
Stranger623/EVE	finetune	—	—	—	apache-2.0	Stranger623	0	0
RichieBreckle/AlfredAI	adapter	—	—	—	apache-2.0	RichieBreckle	0	0
torxytonnicker/finance1	adapter	—	—	—	mit	torxytonnicker	0	0
Newnav/bing	finetune	—	—	—	llama4	Newnav	0	0
updoghiii/betterai	finetune	—	—	—	—	updoghiii	0	0
Zakomako4567/zewai1	finetune	—	—	—	—	Zakomako4567	0	0
MustafaGulsen34/temizlik	finetune	—	—	—	apache-2.0	MustafaGulsen34	0	0
Persilia-AI/GPT-2-Persilia-ai	finetune	—	—	—	mit	Persilia-AI	0	0
Abdulwibu/Tele_n8n	finetune	—	—	—	apache-2.0	Abdulwibu	0	0
waster76/Zod	finetune	—	—	—	apache-2.0	waster76	0	0
aeevnn/NaveenBhav	adapter	—	—	—	apache-2.0	aeevnn	0	0

Repository changelog

8 commits

changed in the last 7 daysOther: 2026-10-24all changes

Repository changelog 8 commits

Date	Commit	SHA
2025-08-26	Update ue8m0 config & description	d3d4eaf
2025-08-22	Fix tool call description	c230953
2025-08-21	Release DeepSeek-V3.1	894cb07
2025-08-19	Add files using upload-large-folder tool	f1d0343
2025-08-19	Add files using upload-large-folder tool	61bc5a3
2025-08-19	Add files using upload-large-folder tool	708e20b
2025-08-19	Add files using upload-large-folder tool	4f0dbf5
2025-08-19	initial commit	b615e56

Articles and coverage

8 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-v3.1

api-docs.deepseek.com2025-08-21officialauthorityHN front pageabout the DeepSeek V3.1 familyHN 778 pts · 263 comments211 words

DeepSeek-V3.1 launches hybrid inference modes with Think and Non-Think capabilities, where deepseek-reasoner handles thinking tasks and deepseek-chat manages non-thinking operations. The model utilizes 840B tokens of continued pretraining for long context extension and supports 128K context windows alongside Anthropic API format compatibility. Pricing changes take effect at Sep 5th, 2025, 16:00 (UTC Time), with strict function calling now available in Beta API.written, reviewed, cited

DeepSeek-v3.1-Terminus

api-docs.deepseek.com2025-09-22officialauthorityHN front pageabout the DeepSeek V3.1 familyHN 101 pts · 28 comments87 words

Analysis5

DeepSeek v3.1 is not having a moment

thezvi.wordpress.com2025-08-22authorityabout the DeepSeek V3.1 familyHN 41 pts · 27 comments631 words

DeepSeek v3.1 is a 685-billion-parameter open-source model that unifies chat, reasoning, and coding into a single architecture with 128K context windows. The model posts a 71.6% score on the Aider coding benchmark, edging out Claude Opus 4 while costing approximately 68 times less in inference. Pricing changes take effect at Sep 5th, 2025, 16:00 UTC, ending off-peak discounts.written, reviewed, cited

DeepSeek delayed its new model release after failing to train it using Huawei's chips.
The startup used Nvidia chips for training and Huawei's chips for inference due to technical issues.
DeepSeek-V3.1-Think reaches answers in less time compared to DeepSeek-R1-0528.

DeepSeek v3.1 is not having a moment

thezvi.substack.com2025-08-23authorityabout the DeepSeek V3.1 familyHN 6 pts · 1 comments960 words

DeepSeek released v3.1, a 685-billion-parameter open-source model that handles 128k context windows and posts a 71.6% score on the Aider coding benchmark. The model employs hybrid inference with Think and Non-Think modes, utilizing ≥2x fewer reasoning tokens to achieve faster thinking speeds. DeepSeek states the update includes stronger agent skills and better results on SWE and Terminal-Bench.written, reviewed, cited

DeepSeek delayed the v4 and r2 releases after failing to train the model using Huawei's Ascend chips.
The startup encountered persistent technical issues during R2 training on Ascend chips, forcing a switch to Nvidia chips.
New pricing changes take effect at Sep 5th, 2025, 16:00 (UTC Time), ending off-peak discounts.

DeepSeek v3.1 Nex N1

openrouter.ai2025-12-17authorityabout the DeepSeek V3.1 familyHN 1 pts · 0 comments66 words

DeepSeek v3.1 just dropped – and it might be the most powerful open AI yet

venturebeat.com2025-08-19about the DeepSeek V3.1 familyHN 12 pts · 0 comments

DeepSeek v3.1 is Here, 685B parameters

xcancel.com2025-08-20about the DeepSeek V3.1 familyHN 4 pts · 0 comments

Community1

DeepSeek v3.1 released – single model, thinking and non-thinking modes

twitter.com2025-08-21about the DeepSeek V3.1 familyHN 26 pts · 0 comments

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

DeepSeek-V3 Technical Report this model

DeepSeek-AI, Aixin Liu, Bei Feng +197 · 2024-12-27 · PDF · HF paper page (89 upvotes)

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 · HF API: 2026-10-24 21:41 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:13 UTC · Last verified 2026-10-25

Follow DeepSeek-V3.1-Base

Watch this model

/feeds/models/deepseek-ai--DeepSeek-V3.1-Base.xml(RSS 2.0) · every change to DeepSeek-V3.1-Base
/feeds/families/deepseek-ai--DeepSeek-V3.1.xml(RSS 2.0) · every change across the DeepSeek-V3.1 family

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