DeepSeek-V2-Lite

deepseek-ai DeepSeek License (deepseek) out 2 years ago

15.7B parameters (6 of 64 experts per token), text in, text out, 160K tokens of context.

Verified 2026-10-26 07:27 UTC Rank #329 in LLMs · score 82 method v1 Sources 3 Downloads 30d 141.6k · #274 of 619 tracked

Changed this week: 1 new quantized build on the Hub (27 in all).2026-10-24

Can I run it?9.4GB at INT4Estimatefits a 16 GB cardmethodCommunity: DeepSeek-V2-Lite — Parameters, Context, Benchmarks & Pricing | AI Atlas →

Can I use it?DeepSeek License (deepseek)Attribution requiredlicence file2026-10-26

Do people use it?141.6kdownloads, 30 days#274 of 619 we track · top 44%+4 likes, downloads −7% this week209 likesHugging Face2026-10-26

Worth reading and watching

Picked by rule from 16 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 21 HN pts

Deepseek-V2 censors prompt 'What happened at Tienanmen square?'

old.reddit.com 2024-05-16 21 HN pts about the DeepSeek V2 family family

Latest · 38 days ago

DeepSeek v2: 437x less memory footprint vs. v1 [video]

youtube.com 2026-09-18 1 HN pts about the DeepSeek V2 family family

16 best of 16 tracked

The 3 picks above, plus these 13. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleDeepSeek V2, near GPT4 performance with 30% GPT3.5 cost ($0.14M input tokens)deepseek.com · 2024-05-08 · 16 HN pts · official
CodeDeepSeek-V2: A Strong, Economical, and Efficient Moe Language Modelgithub.com · 2024-05-07 · 14 HN pts
CodeDEEPSEEK_V2_LITE_RESEARCH.mdgithub.com · undated
PaperDeepSeek-V2: A Strong, Economical, and Efficient MOE Language Modelarxiv.org · 2025-01-01 · 3 HN pts
CodeDeepSeek-V2 integrated, RAGFlow v0.5.0 is releasedgithub.com · 2024-05-08 · 7 HN pts
DiscussionLocal DeepSeeK-V2 Inference: 14T/S for Decode W Only 21GB 4090 and 136GB DRAMold.reddit.com · 2024-07-30 · 8 HN pts
Articlehttps://cdn.jsdelivr.net/gh/deepseek-ai/deepseek-v2@main/README.mdcdn.jsdelivr.net · undated
ArticleDeepSeek-V2-Lite — Parameters, Context, Benchmarks & Pricing | AI Atlasai-atlas.co · undated
ArticleDeepSeek v2 Lite — Hardware Requirements & Compatibility | llmrunllmrun.dev · undated
ArticleDeepSeek V2: Architecture, Benchmarks & Legacydeepseekai.guide · 2026-04-25
ArticleBest Open Source LLM:Tencent Hunyuan-Large Outperforming Llama 3 and DeepSeek V2techwavearena.com · 2024-11-08
ArticleModel Variants | deepseek-ai/DeepSeek-V2 | DeepWikideepwiki.com · 2025-04-18
ArticleDeepSeek-V2: A Huge LLM with Efficient Inferencethesalt.substack.com · 2024-05-22

Show all 13

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: 27 quantized builds and 16 Spaces.

As a service0no API provider listed
On own hardware27quantized builds
Built on3533 fine-tunes · 2 adapters · 0 merges
In apps16Hugging Face Spaces

Top Spaces running it: Justinrune/LLaMA-Factory ♥ 12 · mlanvvs/fitcheck ♥ 2 · luancy1208/CHIP ♥ 1

Will it fit?

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

Fits a 16 GB card at INT4

About 9.4 GB at INT4, 19 GB at FP8, 38 GB 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 (2024-05-15), amber triangles are 7 dated articles (each links to it), grey ticks are repo commits.

What people built on it

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

62 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 legraphista/DeepSeek-V2-Lite-IMat-GGUF for llama.cpp, Ollama, LM Studio 34.8k downloads NVFP4 carlyou/DeepSeek-V2-Lite-NVFP4 for vLLM, SGLang 2.8k downloads

All variants →

Reference

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

Benchmarks

16 rows

Benchmarks 16 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (16)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 BBH · Reason through hard problems	44.1	Base Model	model author	2026-10-07
author-reported BBH · Reason through hard problems	48.1	Chat Model (SFT)	model author	2026-10-07
author-reported C-Eval · Work across languages	60.3	Base Model	model author	2026-10-07
author-reported C-Eval · Work across languages	60.1	Chat Model (SFT)	model author	2026-10-07
author-reported CMMLU · Work across languages	64.3	Base Model	model author	2026-10-07
author-reported CMMLU · Work across languages	62.5	Chat Model (SFT)	model author	2026-10-07
author-reported GSM8K · Reason through hard problems	41.1	Base Model	model author	2026-10-07
author-reported GSM8K · Reason through hard problems	72	Chat Model (SFT)	model author	2026-10-07
author-reported HumanEval · Write code with agents	29.9	Base Model	model author	2026-10-07
author-reported HumanEval · Write code with agents	57.3	Chat Model (SFT)	model author	2026-10-07
author-reported MATH-500 · Reason through hard problems	17.1	Base Model	model author	2026-10-07
author-reported MATH-500 · Reason through hard problems	27.9	Chat Model (SFT)	model author	2026-10-07
author-reported MBPP · Write code with agents	43.2	Base Model	model author	2026-10-07
author-reported MBPP · Write code with agents	45.8	Chat Model (SFT)	model author	2026-10-07
author-reported MMLU · Reason through hard problems	58.3	Base Model	model author	2026-10-07
author-reported MMLU · Reason through hard problems	55.7	Chat Model (SFT)	model author	2026-10-07

Use-case evidence

6 use cases

What it is good at 6 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.

Reason through hard problemsauthor-reportedAuthor reports MMLU 58.3 Author reports BBH 48.1 Author reports GSM8K 72.0 +1 more
Work across languagesauthor-reportedAuthor reports C-Eval 60.3 Author reports CMMLU 64.3
Write code with agentsauthor-reportedAuthor reports HumanEval 57.3 Author reports MBPP 45.8
Chat and follow instructionsstructuraldecision-model estimate p=0.76 tag: conversational chat template present
Read long documentsstructuralconfig: 163,840 context
Run on a laptop or phonestructuralname: Lite

Facts

22 fields

Listed in LLMs · MoE

Facts

DeepSeek-V2-Lite has 15.7B parameters. (safetensors metadata, observed 2026-10-26)
DeepSeek-V2-Lite is a mixture-of-experts model with 64 experts, routing 6 per token. (config.json, observed 2026-10-26)
Context length is 163,840 tokens (160K). (config.json max_position_embeddings, observed 2026-10-26)
License: DeepSeek License (deepseek) (friction: attribution). (license file, observed 2026-10-26)
Created 2024-05-15, last modified 2024-06-25. (HF API, observed 2026-10-26)

Pipeline tag	text-generation
Library	transformers
Architectures	DeepseekV2ForCausalLM
Model type	deepseek_v2
Dtype	bfloat16
Params by dtype	BF16: 15,706,484,224
Size band	L
MoE	yes
Experts	64
Experts per token	6
Context length	163,840 (config.json max_position_embeddings)
Papers	arXiv:2405.04434
License	DeepSeek License (deepseek) (other)
Hub snapshot	2026-10-26

Servability

vLLM · SGLang

Servability

vLLM: supported · DeepseekV2ForCausalLM · 4b1c04e76fe9ac3ca8b7e541dd423e7d5f38dc60SGLang: supported · DeepseekV2ForCausalLM · 60d941293e3af27a12c6eb64e2c3e2e1a17f100f

vLLM registry (4b1c04e76fe9ac3ca8b7e541dd423e7d5f38dc60) · SGLang registry (60d941293e3af27a12c6eb64e2c3e2e1a17f100f) · 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-V2-Lite: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	38 GB · Single high-end GPU (<80 GB)
FP8 (1 byte/param)	19 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)	9.4 GB · Single GPU (<24 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 model Family: deepseek-ai/DeepSeek-V2 62 derivatives on the Hub

quantized: 27finetune: 33adapter: 2

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
legraphista/DeepSeek-V2-Lite-IMat-GGUF	quantized	GGUF	—	—	—	legraphista	5	34.8k
BBQGOD/DeepSeek-GRM-16B	finetune	—	15.7B	L	other	BBQGOD	7	3.5k
slowfastai/DeepSeek-V2-Lite-bnb-4bit	quantized	-bnb	16.2B	L	—	slowfastai	0	3k
carlyou/DeepSeek-V2-Lite-NVFP4	quantized	NVFP4	9.02B	L	other	carlyou	0	2.8k
itlwas/DeepSeek-V2-Lite-Q4_K_M-GGUF	quantized	GGUF	—	—	other	itlwas	0	2.8k
yschoi31/DeepSeek-V2-Lite-Eagle3	finetune	—	175M	S	apache-2.0	yschoi31	0	1.1k
zjiayu064/DeepSeek-V2-Lite-BitsMoE-2bit	quantized	—	2.36B	M	other	zjiayu064	0	1.1k
Lucy-in-the-Sky/DeepSeek-V2-Lite-Q2_K-GGUF	quantized	Q2_K	—	—	other	Lucy-in-the-Sky	0	876
carlyou/DeepSeek-V2-Lite-FP8-Group	quantized	-FP8	15.9B	L	other	carlyou	0	866
xd2010/DeepSeek-V2-Lite-densemixer-math7k-2epoch	finetune	—	15.7B	L	—	xd2010	0	820
OPEA/DeepSeek-V2-Lite-int4-sym-inc	quantized	int4	2.62B	M	—	OPEA	1	764
wencan-lab/DeepSeek-V2-Lite-Q4_K_M-GGUF	quantized	GGUF	—	—	other	wencan-lab	0	812
phi0112358/DeepSeek-V2-Lite-Q5_K_M-GGUF	quantized	GGUF	—	—	other	phi0112358	0	742
HectorHe/DeepSeek-V2-Lite-aux-free-sft-math7k-1epoch-1e-4-gamma	finetune	—	15.7B	L	—	HectorHe	1	599
Beambutbetter/Deepseek-V2-Lite-16B-NVFP4	quantized	NVFP4	8.06B	M	apache-2.0	Beambutbetter	3	498
jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6	finetune	—	2.66B	M	other	jhyuckkim	0	648
jhyuckkim/DeepSeek-V2-Lite-Dense-3B-SF-K6	finetune	—	2.66B	M	other	jhyuckkim	0	640
xd2010/DeepSeek-V2-Lite-esft-math7k-2epoch	finetune	—	15.7B	L	—	xd2010	0	620
jhyuckkim/DeepSeek-V2-Lite-Dense-3B-ACP-K6	finetune	—	2.66B	M	other	jhyuckkim	0	430
jhyuckkim/DeepSeek-V2-Lite-Dense-3B-DO-ACP-K6	finetune	—	2.66B	M	other	jhyuckkim	0	417
xd2010/DeepSeek-V2-Lite-aux-free-sft-math7k-2epoch-frozen-router	finetune	—	15.7B	L	—	xd2010	0	397
Theta-Lev/DeepSeek-V2-Lite-Q8_0-GGUF	quantized	Q8_0	—	—	other	Theta-Lev	0	389
deltanym/DeepSeek-V2-Lite-Q6_K-GGUF	quantized	Q6_K	—	—	other	deltanym	0	371
Lucy-in-the-Sky/DeepSeek-V2-Lite-Q4_K_M-GGUF	quantized	GGUF	—	—	other	Lucy-in-the-Sky	0	367
fxmeng/DSV2-Lite-LatentIndex-Sparse-1250	finetune	—	—	—	other	fxmeng	0	356
fxmeng/DSV2-Lite-IndexCache-Sparse-1250	finetune	—	—	—	other	fxmeng	0	342
ijohn07/DeepSeek-V2-Lite-Q8_0-GGUF	quantized	Q8_0	—	—	other	ijohn07	0	339
Lucy-in-the-Sky/DeepSeek-V2-Lite-Q8_0-GGUF	quantized	Q8_0	—	—	other	Lucy-in-the-Sky	0	334
ijohn07/DeepSeek-V2-Lite-Q6_K-GGUF	quantized	Q6_K	—	—	other	ijohn07	0	324
HectorHe/DeepSeek-V2-Lite-aux-free-sft-math7k-1epoch-bs4	finetune	—	15.7B	L	—	HectorHe	0	262
fxmeng/DSV2-Lite-DSA-Sparse-1250	finetune	—	—	—	other	fxmeng	0	259
HectorHe/DeepSeek-V2-Lite-aux-free-sft-math7k-1epoch-remove-aux-only	finetune	—	126K	S	—	HectorHe	0	253
fxmeng/DSV2-Lite-DSA-Warmup-400	finetune	—	—	—	other	fxmeng	0	236
fxmeng/DSV2-Lite-IndexCache-Warmup-400	finetune	—	—	—	other	fxmeng	0	234
wantsleep/DeepSeek-V2-Lite-Q4_K_M-GGUF	quantized	GGUF	—	—	other	wantsleep	0	229
fxmeng/DSV2-Lite-LatentIndex-Warmup-400	finetune	—	—	—	other	fxmeng	0	224
ttxxhh/DeepSeek-V2-Lite-Q4_0-GGUF	quantized	Q4_0	—	—	other	ttxxhh	0	201
HectorHe/DeepSeek-V2-Lite-aux-free-sft-math7k-1epoch-1e-3-gamma	finetune	—	126K	S	—	HectorHe	0	196
Etelis/DeepSeek-V2-Lite-FP8-BLOCK-padded	quantized	-FP8	15.7B	L	other	Etelis	0	192
Umranz/Primera-NT	finetune	—	15.7B	L	other	Umranz	1	118

Repository changelog

21 commits

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

Repository changelog 21 commits

Date	Commit	SHA
2024-06-25	Update modeling_deepseek.py	604d566
2024-06-08	docs: update README.md	419ef8b
2024-06-07	docs: update README.md	c5b1d95
2024-05-17	Update README.md	901a97a
2024-05-17	Update README.md	c85be81
2024-05-17	Update README.md	2e8db74
2024-05-17	Upload LICENSE	4264766
2024-05-17	Update README.md	2bb6c7d
2024-05-17	Update README.md	77f4da5
2024-05-17	Update README.md	451beb2
2024-05-17	Update README.md	1fa320e
2024-05-17	Update README.md	69ed3d3
2024-05-17	Update README.md	3a0614a
2024-05-16	Update README.md	31c5617
2024-05-16	Update README.md	730833f
2024-05-16	Update README.md	6473224
2024-05-16	Update README.md	da4112d
2024-05-16	Update README.md	47be703
2024-05-15	Upload folder using huggingface_hub	151982a
2024-05-15	Update README.md	b9a3eb3
2024-05-15	initial commit	c91ca88

Articles and coverage

16 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.

Official1

DeepSeek V2, near GPT4 performance with 30% GPT3.5 cost ($0.14M input tokens)

deepseek.com2024-05-08officialauthorityabout the DeepSeek V2 familyHN 16 pts · 2 comments38 words

Analysis7

DeepSeek v2: 437x less memory footprint vs. v1 [video]

youtube.com2026-09-18about the DeepSeek V2 familyHN 1 pts · 0 comments

DeepSeek-V2: A Strong, Economical, and Efficient Moe Language Model

github.com2024-05-07authorityabout the DeepSeek V2 familyHN 14 pts · 3 comments1,699 words

DeepSeek-V2 is a 236B parameter Mixture-of-Experts model pretrained on 8.1 trillion tokens that saves 42.5% of training costs while boosting maximum generation throughput to 5.76 times. The architecture utilizes MLA (Multi-head Latent Attention) to eliminate the bottleneck of inference-time key-value cache, thus supporting efficient inference. To utilize DeepSeek-V2 in BF16 format for inference, 80GB*8 GPUs are required.written, reviewed, cited

The open-source code runs slower on HuggingFace GPUs compared to the internal codebase due to platform constraints.
The SGLang framework supports FP8 KV Cache and offers the best latency and throughput among open-source frameworks for this model.

iamboosted/deepseek-v2-lite-base-benchmarks

github.comauthority1,117 words

DeepSeek-V2-Lite base, a 15.7B total / 2.4B active parameter model trained on 5.7T tokens, scores 0.0% on AIME 2024 and 2025 due to non-termination at 16k tokens. Independent evaluation using vLLM shows the model underperforms SmolLM3-3B-Base on knowledge-heavy tasks, likely driven by the 5.7T versus 11.2T pretraining token gap. The benchmark highlights a 3.6 point drop on GSM8K compared to vendor figures attributed to harness differences.written, reviewed, cited

The model fails to terminate on 57 of 60 AIME problems even when allowed 16,384 tokens.
Loglikelihood scoring for GPQA Diamond will cause an out-of-memory error on a 48 GB card.
A killed vLLM process leaves a zombie that prevents the next run from starting on the GPU.

DEEPSEEK_V2_LITE_RESEARCH.md

github.comauthority3,953 words

DeepSeek-V2-Lite is a high-efficiency Mixture-of-Experts model released on May 16, 2024, designed for economical training and efficient inference. It features a sparse architecture with 15.7B total parameters but only activates 2.4B per token, achieving an 85% sparsity rate with 66 experts per MoE layer where Top-6 scoring routed experts are selected. The model uses Multi-head Latent Attention to reduce KV cache size by 93.3% compared to standard attention mechanisms, offering a 4:1 compression ratio and 75% reduction in KV cache while maintaining a 32K context window.written, reviewed, cited

The DeepSeek-R1-Distill-Qwen-7B (dense, January 2025) may be simpler to implement first due to no expert routing complexity.

DeepSeek-V2: A Strong, Economical, and Efficient MOE Language Model

arxiv.org2025-01-01authorityabout the DeepSeek V2 familyHN 3 pts · 0 comments404 words

DeepSeek-V2 is a strong Mixture-of-Experts language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE, where MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector.written, reviewed, cited

DeepSeek-V2 integrated, RAGFlow v0.5.0 is released

github.com2024-05-08authorityabout the DeepSeek V2 familyHN 7 pts · 0 comments460 words

cdn.jsdelivr.netabout the DeepSeek V2 family1,728 words

Community2

Deepseek-V2 censors prompt 'What happened at Tienanmen square?'

old.reddit.com2024-05-16about the DeepSeek V2 familyHN 21 pts · 0 comments

Local DeepSeeK-V2 Inference: 14T/S for Decode W Only 21GB 4090 and 136GB DRAM

old.reddit.com2024-07-30about the DeepSeek V2 familyHN 8 pts · 0 comments

All coverage tracked for this model (16) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
iamboosted/deepseek-v2-lite-base-benchmarks	github.com	—	2026-10-14	exa
https://cdn.jsdelivr.net/gh/deepseek-ai/deepseek-v2@main/README.md	cdn.jsdelivr.net	—	2026-10-14	exa
DEEPSEEK_V2_LITE_RESEARCH.md	github.com	—	2026-10-14	exa
DeepSeek-V2-Lite — Parameters, Context, Benchmarks & Pricing \| AI Atlas	ai-atlas.co	—	2026-10-14	exa
DeepSeek v2 Lite — Hardware Requirements & Compatibility \| llmrun	llmrun.dev	—	2026-10-14	exa
DeepSeek v2: 437x less memory footprint vs. v1 [video]	youtube.com	2026-09-18	2026-10-07	hn
DeepSeek V2: Architecture, Benchmarks & Legacy	deepseekai.guide	2026-04-25	2026-10-14	exa
Model Variants \| deepseek-ai/DeepSeek-V2 \| DeepWiki	deepwiki.com	2025-04-18	2026-10-14	exa
DeepSeek-V2: A Strong, Economical, and Efficient MOE Language Model	arxiv.org	2025-01-01	2026-10-07	hn
Best Open Source LLM:Tencent Hunyuan-Large Outperforming Llama 3 and DeepSeek V2	techwavearena.com	2024-11-08	2026-10-07	hn
Local DeepSeeK-V2 Inference: 14T/S for Decode W Only 21GB 4090 and 136GB DRAM	old.reddit.com	2024-07-30	2026-10-07	hn
DeepSeek-V2: A Huge LLM with Efficient Inference	thesalt.substack.com	2024-05-22	2026-10-14	exa
Deepseek-V2 censors prompt 'What happened at Tienanmen square?'	old.reddit.com	2024-05-16	2026-10-07	hn
DeepSeek-V2 integrated, RAGFlow v0.5.0 is released	github.com	2024-05-08	2026-10-07	hn
DeepSeek V2, near GPT4 performance with 30% GPT3.5 cost ($0.14M input tokens)	deepseek.com official	2024-05-08	2026-10-07	hn
DeepSeek-V2: A Strong, Economical, and Efficient Moe Language Model	github.com	2024-05-07	2026-10-07	hn

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-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model this model

DeepSeek-AI, Aixin Liu, Bei Feng +154 · 2024-05-07 · PDF · HF paper page (30 upvotes)

Similar models

3 models

Similar models

gpt-oss-20b · 20.9B
same size band (L), also MoE, same task (text-generation), more providers, permissive license
Moonlight-16B-A3B-Instruct · 16.0B
same size band (L), also MoE, same task (text-generation), permissive license
deepseek-moe-16b-chat · 16.4B
same size band (L), also MoE, same task (text-generation)

Sources

3 sources

Sources

Hub snapshot: 2026-10-26 · HF API: 2026-10-26 07:20 UTC · Providers: 2026-10-26 07:26 UTC · Articles: 2026-10-26 07:27 UTC · Summary: 2026-10-07 15:09 UTC · Last verified 2026-10-26 07:27 UTC

Follow DeepSeek-V2-Lite

Watch this model

/feeds/models/deepseek-ai--DeepSeek-V2-Lite.xml(RSS 2.0) · every change to DeepSeek-V2-Lite
/feeds/families/deepseek-ai--DeepSeek-V2.xml(RSS 2.0) · every change across the DeepSeek-V2 family

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