Community

236 community available in the DeepSeek directory

What happened to Deepseek?

Meta had a comeback - arguably not opensource, but still - but Deepseek just seems to have vanished from the scene. What happened? Will we ever see Deepseek V4?

Mr_Moonsilver

326

From Twitter/X: DeepSeek is rolling out a limited V4 gray release.

Source: https://x.com/i/status/2041458478569689589

jmorant555

Gemma 4 just casually destroyed every model on our leaderboard except Opus 4.6 and GPT-5.2. 31B params, $0.20/run

Tested Gemma 4 (31B) on our benchmark. Genuinely did not expect this. 100% survival, 5 out of 5 runs profitable, +1,144% median ROI. At $0.20 per run. It outperforms GPT-5.2 ($4.43/run), Gemini 3 Pro ($2.95/run), Sonnet 4.6 ($7.90/run), and absolutely destroys every Chinese open-source model we've tested — Qwen 3.5 397B, Qwen 3.5 9B, DeepSeek V3.2, GLM-5. None of them even survive consistently. The only model that beats Gemma 4 is Opus 4.6 at $36 per run. That's 180× more expensive. 31 billion parameters. Twenty cents. We double-checked the config, the prompt, the model ID — everything is identical to every other model on the leaderboard. Same seed, same tools, same simulation. It's just this good. Strongly recommend trying it for your agentic workflows. We've tested 22 models so far and this is by far the best cost-to-performance ratio we've ever seen. Full breakdown with charts and day-by-day analysis: foodtruckbench.com/blog/gemma-4-31b FoodTruck Bench is an AI business simulation benchmark — the agent runs a food truck for 30 days, making decisions about location, menu, pricing, staff, and inventory. Leaderboard at foodtruckbench.com EDIT — Gemma 4 26B A4B results are in. Lots of you asked about the 26B A4B variant. Ran 5 simulations, here's the honest picture: 60% survival (3/5 completed, 2 bankrupt). Median ROI: +119%, Net Worth: $4,386. Cost: $0.31/run. Placed #7 on the leaderboard — above every Chinese model and Sonnet 4.5, below everything else. Both bankruptcies were loan defaults — same pattern we see across models. The 3 surviving runs were solid, especially the best one at +296% ROI. But here's the catch. The 26B A4B is the only model out of 23 tested that required custom output sanitization to function. It produces valid tool-call intent, but the JSON formatting is consistently broken — malformed quotes, trailing garbage tokens, invalid escapes. I had to build a 3-stage sanitizer specifically for this model. No other model needed anything like this. The business decisions themselves are unmodified — the sanitizer only fixes JSON formatting, not strategy. But if you're planning to use this model in agentic workflows, be prepared to handle its output format. It does not produce clean function calls out of the box. TL;DR: 31B dense → 100% survival, $0.20/run, #3 overall. 26B A4B → 60% survival, $0.31/run, #7 overall, but requires custom output parsing. The 31B is the clear winner. Updated leaderboard: foodtruckbench.com

Disastrous_Theme5906

1,895

One year ago DeepSeek R1 was 25 times bigger than Gemma 4

I'm mind blown by the fact that about a year ago DeepSeek R1 came out with a MoE architecture at 671B parameters and today Gemma 4 MoE is only 26B and is genuinely impressive. It's 25 times smaller, but is it 25 times worse? I'm exited about the future of local LLMs.

rinaldo23

411

DeepSeek Employee Teases "Massive" New Model Surpassing DeepSeek V3.2

Translated by Nano Banana https://preview.redd.it/8bfh5zk1q6rg1.png?width=1158&format=png&auto=webp&s=9d8e6c2f285ba04527f0e9578f9ca7b75124c11f https://preview.redd.it/jpa7aikcr6rg1.png?width=688&format=png&auto=webp&s=2a35594f8ff5eb5f2cd18ad2f4de6662f2898b1d Note: The employee just deleted his reply; it seems he said something he shouldn't have. Original post: http://xhslink.com/o/3ct3YOygvNN

External_Mood4719

329

DeepSeek Core Researcher Daya Guo Rumored to Have Resigned

Recently, heavy-hitting news regarding a major personnel change has emerged in the field of Large Language Models (LLMs): Daya Guo, a core researcher at DeepSeek and one of the primary authors of the DeepSeek-R1 paper, has reportedly resigned. Public records show that Daya Guo possesses an exceptionally distinguished academic background. He obtained his PhD from Sun Yat-sen University in 2023, where he was mentored by Professor Jian Yin and co-trained by Ming Zhou, the former Deputy Dean of Microsoft Research Asia (MSRA). Daya Guo officially joined DeepSeek in July 2024, focusing his research on Code Intelligence and the reasoning capabilities of Large Language Models. During his tenure at DeepSeek, Guo demonstrated remarkable scientific talent and was deeply involved in several of the company’s milestone projects, including DeepSeekMath, DeepSeek-V3, and the globally acclaimed DeepSeek-R1. Notably, the research findings related to DeepSeek-R1 successfully graced the cover of the top international scientific journal Nature in 2025, with Daya Guo serving as one of the core authors of the paper. Regarding his next destination, several versions are currently circulating within the industry. Some reports suggest he has joined Baidu, while other rumors indicate he has chosen ByteDance. As of now, neither the relevant companies nor Daya Guo himself have issued an official response. External observers generally speculate that the loss of such core talent may be related to the intense "talent war" and competitive compensation packages within the LLM sector. As the global AI race reaches a fever pitch, leading internet giants are offering highly lucrative salaries and resource packages to secure top-tier talent with proven practical experience. Insiders point to two primary factors driving Guo’s departure: 1. Computing Resources: Despite DeepSeek's efficiency, the sheer volume of computing power available at the largest tech giants remains a significant draw for researchers pushing the boundaries of LLM reasoning. 2. Compensation Issues: Reports indicate a "salary inversion" within the company, where newer hires were reportedly receiving higher compensation packages than established core members. The departure may not be an isolated incident. Rumors are circulating that other "important figures" within DeepSeek are currently in talks with major tech firms, seeking roles with larger "scope" and better resources. As the global AI race reaches a fever pitch, the ability of "AI unicorns" to retain top-tier talent against the massive resources of established internet giants is facing its toughest test yet. Source from some Chinese news: https://www.zhihu.com/pin/2018475381884200731 https://news.futunn.com/hk/post/70411035?level=1&data\_ticket=1771727651415532 https://www.jiqizhixin.com/articles/2026-03-21-2 https://www.xiaohongshu.com/discovery/item/69bd211c00000000230111fb?source=webshare&xhsshare=pc\_web&xsec\_token=CBbUil7jGmHR\_sMr3sM56dYn9utmWYYN11mYMfe6FL0Cw=&xsec\_source=pc\_share

External_Mood4719

124

Qwen wants you to know…

Seen while walking through Singapore’s Changi airport earlier this week. Alibaba Cloud spending up big on advertising.

m-gethen

1,990

I need help with testing my llama.cpp Deepseek Sparse Attention (DSA) implementation (someone GPU-rich)

I have initial proof-of-concept implementation ready and now I want to confirm that it works correctly. Unfortunately the difference between the model performance with dense vs sparse attention is subtle and it's visible only for very complex problems. Basically you need a full benchmark run to make sure the implementation works correctly. I can't do it on my Epyc 9374F + RTX PRO 6000 workstation as it would take hundreds of hours. What I need is an access to a machine with at least 768 GB of VRAM (or more) for a few hours to run lineage-bench (either a full run or limited lineage-256/lineage-512) on DeepSeek V3.2 Speciale in Q8\_0 in my llama.cpp deepseek-dsa branch with dense and sparse attention and compare results with my sglang fp8 tests. It may be either direct or via human proxy. I have GGUFs ready. I tried to do it on vast.ai rented 8x RTX PRO 6000 instance, but had problems fitting the model with indexer tensors on this configuration (CUDA OOM errors). So either more time to research this or more powerful hardware is needed - and I feel that I already burned enough money on this.

fairydreaming

What the hell is Deepseek doing for so long?

Almost all the Chinese AI companies have surpassed their models. Even Xiaomi now has a far better model. They are still somehow stuck in v 3.2 with minor updates. They supposedly have so much resources now that they have international attention. They haven't even released a decent multimodal model. Are they just out of race at this point? I don't see how they can even compete with frontier Chinese AI companies, much less than frontier US companies unless they release something that's truly groundbreaking in every way.

Terrible-Priority-21

226

Oh Deepseek V4, where art thou?

Ok, ok, so I don't really expect an answer to this question, but I am really hoping the new Deepseek model drops pretty soon. After dealing with the US model companies I am SO ready for more open models to arrive on the scene to challenge them. Please oh Deepseek team, won't you bring us more open innovation? Hopefully sooner rather than later. Until then I'll continue to dream of more open model innovations to come... EDIT: I honestly didn't expect to get crucified for this post and downvoted so much in this community. If you are a downvoter I'd love to know your reasons so I can learn from my mistakes..

awebb78

DeepSeek V4 coming this week?

de_sonnaz

13 months since the DeepSeek moment, how far have we gone running models locally?

Once upon a time there was a tweet from an engineer at Hugging Face explaining how to run the frontier level DeepSeek R1 @ Q8 at \~5 tps for about $6000. Now at around the same speed, with this $600 mini PC, you can run the highly superior Qwen3-27B @ Q4. But if you want more usable speeds, with the still much stronger Qwen3.5-35B-A3B @ Q4/Q5, you can get 17-20 tps. Isn't it wild? At this pace of improving smaller models, could we be running next year a 4B model better than Kimi 2.5?

dionisioalcaraz

340

What if LLM agents passed KV-cache to each other instead of text? I tried it -- 73-78% token savings across Qwen, Llama, and DeepSeek

If you've used multi-agent setups with LangChain, CrewAI, AutoGen, or Swarm, you've probably noticed: every agent re-tokenizes and re-processes the full conversation from scratch. Agent 3 in a 4-agent chain is re-reading everything agents 1 and 2 already chewed through. When I measured this across Qwen2.5, Llama 3.2, and DeepSeek-R1-Distill, 47-53% of all tokens in text mode turned out to be redundant re-processing. AVP (Agent Vector Protocol) is my attempt to fix this. Instead of passing text between agents, it passes the KV-cache directly. Agent A finishes reasoning serializes its key-value attention states, and Agent B injects them. No re-tokenization, no redundant forward passes. Text: Planner -> [text] -> Critic re-tokenizes everything -> [text] -> Refiner re-tokenizes everything Latent: Planner -> [KV-cache] -> Critic injects, skips to generation -> [KV-cache] -> Refiner same What it actually does: Same model on both sides? Direct KV-cache transfer, zero overhead. Same family, different size (e.g. Qwen2.5-7B talking to 1.5B)? Vocabulary-mediated projection. No learned params, no calibration data needed. Different families? Falls back to JSON. Not everything needs to be fancy. Transport-agnostic -- works alongside A2A, MCP, gRPC, whatever you're already using Binary wire format, not JSON+Base64 (33% overhead on tensor data is painful) Numbers (these are structural, not accuracy claims): Token savings of 73-78% and 2-4x speedups held consistent across all three model families. This isn't model-dependent -- it's just fewer forward passes, so less wall time. Here's the intuition: text prompt sizes balloon at each hop (186 -> 545 -> 1,073 -> 1,397 tokens in a 4-agent GSM8K chain). Latent stays flat at \~164-207 tokens per hop because prior context arrives as pre-computed KV-cache, not as text that needs re-encoding. The gap widens with chain length. At 4 agents it's roughly 2x. At 16 agents (projected) it'd be around 6x, because text scales O(n\^2) while latent scales O(n). Limitations (yes, I know about these): Sample sizes are n=20 per model. The token and speed numbers are solid because they're structural (fewer forward passes is fewer forward passes), but n=20 isn't enough to make accuracy claims. That's future work. Tested on small models only (1.5B-3B on an RTX 3070 Ti). 7B+ results pending. This is a datacenter / same-machine thing. KV-cache for a 3B model runs about 130 MB per sample. You need 1 Gbps+ bandwidth minimum. Sending this over the internet is not happening. Requires KV-cache access, so self-hosted only. Won't work with OpenAI/Anthropic/etc. APIs. Same model only for now. Cross-model (Rosetta Stone) is implemented but not benchmarked yet. Latent uses 17-54x more VRAM than text because you're holding KV-cache across hops instead of discarding it. Totally fine for 1.5B-3B on 8GB+ GPUs. At 7B+ it becomes a real constraint, and I don't have a clean answer for that yet. Try it yourself: pip install avp Two API levels depending on how much control you want: import avp msg = avp.pack("Hello", model="Qwen/Qwen2.5-7B-Instruct", think_steps=20) answer = avp.unpack(msg, model="Qwen/Qwen2.5-7B-Instruct") from avp import HuggingFaceConnector connector = HuggingFaceConnector.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct") context = connector.think("Analyze this problem", steps=20) answer = connector.generate("Solve it.", context=context) vLLM connector also available (pip install "avp[vllm]"). Links: SDK: github.com/VectorArc/avp-python (MIT, 377 tests, 7 benchmarks) Spec: github.com/VectorArc/avp-spec Benchmark details: BENCHMARKS.md This is a nights-and-weekends project born out of my own multi-agent work. Happy to answer questions about the implementation and genuinely interested in feedback from people running multi-agent setups in production.

proggmouse

120

DeepSeek V4 will be released next week and will have image and video generation capabilities, according to the Financial Times

Financial Times: DeepSeek to release long-awaited AI model in new challenge to US rivals (paywall): https://www.ft.com/content/e3366881-0622-40a7-9c34-a0d82e3d573e

Nunki08

647

DeepSeek allows Huawei early access to V4 update, but Nvidia and AMD still don’t have access to V4

https://www.reuters.com/world/china/deepseek-withholds-latest-ai-model-us-chipmakers-including-nvidia-sources-say-2026-02-25/ According to a Reuters report today, DeepSeek has recently granted early access to its major V4 update to domestic suppliers such as Huawei. This move is intended to help these companies optimize their processor software and ensure the model runs efficiently on their hardware. However, chipmakers like Nvidia and AMD have not yet been granted access.

External_Mood4719

424

DeepSeek released new paper: DualPath: Breaking the Storage Bandwidth Bottleneck in Agentic LLM Inference

https://arxiv.org/abs/2602.21548 https://preview.redd.it/25rh3yahktlg1.png?width=536&format=png&auto=webp&s=f282d71496b6386841732137a474f1b238269950 A joint research team from Peking University, Tsinghua University, and DeepSeek-AI has released its latest research findings on optimizing Large Language Model (LLM) inference architectures. The team successfully developed a novel inference system called \\DualPath\\, specifically designed to address technical bottlenecks in KV-Cache storage I/O bandwidth under agentic workloads. https://preview.redd.it/hdssmlcnktlg1.png?width=511&format=png&auto=webp&s=6ba3bc1fd5fa0f310205f8de5bb73e022a0a8263

External_Mood4719

214

My theory on all the negative Chinese AI media coverage right now. It's about the stock market, investor panic, and the upcoming release of Deepseek V4.

Everywhere you look right now in the media, the news cycle is dominated by attacks on Chinese AI Labs, saying they trained on illegal Nvidia GPUs, the can only do what they do because they distill on American model companies responses, they lack any true capability of innovation internally and can only copy what they see. I have not seen this many coordinated attacks against Chinese AI Labs before, although after Deepseek was released last year there were definitely atttacks. I've been thinking about this barrage of negative coverage at this very moment from every single American AI Labs, plus Nvidia (all at the same time) and it occurred to me that the last time Deepseek launched a model there was massive investor panic, and what is expected to happen anytime now? Yep, Deepseek is expected to release their anticipated V4 version of Deepseek. I believe this timing of negative coverage is specifically designed to drown out any media attention on the upcoming release. Nvidia and the AI companies don't want a repeat of last year, specifically with the investor panic, as they try to raise record amounts for their own AI. And Nividia and Google, etc.. would rather not have their stock values decline by double digits. So they are manufacturing FUD to try to prevent it. Just think about the timing of all this negative media posting when you see it and look through the FUD to see the real fear based on historical evidence before buying into it.

awebb78

Exclusive: China's DeepSeek trained AI model on Nvidia's best chip despite US ban, official says

blahblahsnahdah

190

Deepseek and Gemma ??

ZeusZCC

937

New DeepSeek update: "DeepSeek Web / APP is currently testing a new long-context model architecture, supporting a 1M context window."

From AiBattle on 𝕏: https://x.com/AiBattle\_/status/2022280288643039235

Nunki08

127

DeepSeek just updated to a 1M context window!

The DeepSeek app was just updated with 1M context, and the knowledge cutoff date is now May 2025. It's unclear for now if this is a new model. Also, there hasn't been any movement on their Hugging Face page yet. https://preview.redd.it/9z2ggdgy9uig1.png?width=1179&format=png&auto=webp&s=a3f48da856b53751f2db2b17ac5f49baaf9add55

Dr_Karminski

DeepSeek has launched grayscale testing for its new model on both its official website and app. 1M content length!

This model know Gemini 2.5 Pro on not web search https://preview.redd.it/ontumt5s3uig1.jpg?width=657&format=pjpg&auto=webp&s=efff85457597b8fd9dbcbcf3d1d99d62a0678ea2 DeepSeek has launched grayscale testing for its new model on both its official website and app. The new model features a 1M context window and an updated knowledge base. Currently, access is limited to a select group of accounts." https://preview.redd.it/j1qiarng1uig1.png?width=1163&format=png&auto=webp&s=3a99f1652ea755a7aeaa600250ff4856133fbfca It look Like V4 Lite not actually V4

External_Mood4719

131

Deepseek architecture, but without all the parameters

I’m seeing a pattern that perhaps is not legitimate, but it seems everyone is copying the latest Deepseek architecture on their latest releases. In the process though they are also copying the parameter count (roughly), which makes the models inaccessible to most (unless you use their API or spent as much as you would to buy a used car). So my question is, are there smaller models using the same tech but with less parameters? EDIT: to be clear, I’m not talking generally about the MoE technology. I’m fully aware that’s where we moved to leaving dense models in the dust for the most part. As an example Kimi model and the latest large Mistral model copy more than just MoE.

silenceimpaired

Who is waiting for deepseek v4 ,GLM 5 and Qwen 3.5 and MiniMax 2.2?

The title? I hope they come out soon... I'm especially waiting for DS V4, it should be pretty good, hopefully it will be reasonably fast(probably slow though since it is gonna be bigger than v3.2) via OpenRouter. Well, glm 5 is out already technically on Open Router.

power97992