[HN Gopher] Tongyi DeepResearch - open-source 30B MoE Model that...
       ___________________________________________________________________
        
       Tongyi DeepResearch - open-source 30B MoE Model that rivals OpenAI
       DeepResearch
        
       Author : meander_water
       Score  : 221 points
       Date   : 2025-11-02 11:43 UTC (11 hours ago)
        
 (HTM) web link (tongyi-agent.github.io)
 (TXT) w3m dump (tongyi-agent.github.io)
        
       | jychang wrote:
       | This is over a month old, they released the weights a long time
       | ago.
        
         | earthnail wrote:
         | And for those not so tightly in the loop: how does it compare?
        
         | jwr wrote:
         | That's OK -- not all of us follow all the progress on a daily
         | basis, and a model that is a month old doesn't become useless
         | just by being a month old!
        
       | embedding-shape wrote:
       | Isn't OpenIA "Deep research" (not "DeepResearch") a
       | methodology/tooling thing, and you'll get different responses
       | depending on what specific model you use with it? As far as the
       | UI allows you to, you could use Deep research with GPT-5, GPT-4o,
       | o3 and so on, and that'll have an impact on the responses.
       | Skimming the paper and searching for some simple terms makes it
       | seem like they never expand on what exact models they've used,
       | just that they've used a specific feature from ChatGPT?
        
         | simonw wrote:
         | At this point "deep research" is more of a pattern - OpenAI and
         | Perplexity and Google Gemini all offer products with that name
         | which work essentially the same way, and Anthropic and Grok
         | have similar products with a slightly different name attached.
         | 
         | The pattern is effectively long-running research tasks that
         | drive a search tool. You give them a prompt, they churn away
         | for 5-10 minutes running searches and they output a report
         | (with "citations") at the end.
         | 
         | This Tongyi model has been fine-tuned to be really good at
         | using its search tool in a loop to produce a report.
        
           | embedding-shape wrote:
           | Yes, but I think my previous point still matter, namely what
           | exact model is being used greatly affects the results.
           | 
           | So without specifying which model is being used, it's really
           | hard to know what is better than something else, because we
           | don't understand what the underlying model is, and if it's
           | better because of the model itself, or the tooling, which
           | feels like an important distinction.
        
       | aliljet wrote:
       | Sunday morning, and I find myself wondering how the engineering
       | tinkerer is supposed to best self-host these models? I'd love to
       | load this up on the old 2080ti with 128gb of vram and play, even
       | slowly. I'm curious what the current recommendation on that path
       | looks like.
       | 
       | Constraints are the fun part here. I know this isn't the 8x
       | Blackwell Lamborghini, that's the point. :)
        
         | homarp wrote:
         | llama.cpp gives you the most control to tune it for your
         | machine.
        
         | giobox wrote:
         | If you just want to get something running locally as fast as
         | possible to play with (the 2080ti typically had 11gb of VRAM
         | which will be one of the main limiting factors), the ollama app
         | will run most of these models locally with minimum user effort:
         | 
         | https://ollama.com/
         | 
         | If you really do have a 2080ti with 128gb of VRAM, we'd love to
         | hear more about how you did it!
        
         | CuriousSkeptic wrote:
         | Im sure this guy has some helpful hints on that:
         | https://youtube.com/@azisk
        
         | exe34 wrote:
         | llama.cpp + quantized:
         | https://huggingface.co/bartowski/Alibaba-NLP_Tongyi-DeepRese...
         | 
         | get the biggest one that will fit in your vram.
        
           | davidsainez wrote:
           | This is the way. I managed to run (super) tiny models on CPU
           | only with this approach.
        
           | trebligdivad wrote:
           | How do people deal with all the different quantisations?
           | Generally if I see an Unsloth I'm happy to try it locally;
           | random other peoples...how do I know what I'm getting?
           | 
           | (If nothing else Tongyi are currently winning AI with cutest
           | logo)
        
             | exe34 wrote:
             | personally I've only used them for toying around - but in
             | all cases you have to test them for your use case anyway.
        
         | btbuildem wrote:
         | I've recently put together a setup that seemed reasonable for
         | my limited budget. Mind you, most of the components were
         | second-hand, open box deals, or deep discount of the moment.
         | 
         | This comfortably fits FP8 quantized 30B models that seem to be
         | "top of the line for hobbyists" grade across the board.
         | 
         | - Ryzen 9 9950X
         | 
         | - MSI MPG X670E Carbon
         | 
         | - 96GB RAM
         | 
         | - 2x RTX 3090 (24GB VRAM each)
         | 
         | - 1600W PSU
        
           | pstuart wrote:
           | That's basically what I imagined would be my rig if I were to
           | pull the trigger. Do you have an NVLink adapter as well?
        
             | btbuildem wrote:
             | No NVLink; it took me a long time to compose the exact
             | hardware specs, because I wanted to optimize performance.
             | Both cards are on x8 PCIe direct CPU channels, close to
             | their max throughput anyway. It runs hot with the CPU
             | engaged, but it runs fast.
        
           | nine_k wrote:
           | Does it offer more performance than a Macbook Pro that could
           | be had for a comparable sum? Your build can be had for under
           | $3k; a used MBP M3 with 64 GB RAM can be had for
           | approximately $3.5k.
        
             | btbuildem wrote:
             | I'm not sure, I did not run any benchmarks. As a ballpark
             | figure -- with both cards throttled down to 250W, running a
             | Qwen-30B FP8 model (variant depending on task), I get
             | upwards of 60 tok/sec. It feels on par with the premium
             | models, tbh.
             | 
             | Of course this is in a single-user environment, with vLLM
             | keeping the model warm.
        
           | PeterStuer wrote:
           | Unfortunately the RTX 3090 has no native FP8 support.
        
         | jlokier wrote:
         | I use a Macbook Pro with 128GB RAM "unified memory" that's
         | available to both CPU and GPU.
         | 
         | It's slower than a rented Nvidia GPU, but usable for all the
         | models I've tried (even gpt-oss-120b), and works well in a
         | coffee shop on battery and with no internet connection.
         | 
         | I use Ollama to run the models, so can't run the latest until
         | they are ported to the Ollama library. But I don't have much
         | time for tinkering anyway, so I don't mind the publishing
         | delay.
        
           | MaxMatti wrote:
           | How's the battery holding up during vibe coding sessions or
           | occasional LLM usage? I've been thinking about getting a
           | MacBook or a laptop with a similar Ryzen chip specifically
           | for that reason.
        
           | anon373839 wrote:
           | I'd strongly advise ditching Ollama for LM Studio, and using
           | MLX versions of the models. They run quite a bit faster on
           | Apple Silicon. Also, LM Studio is much more polished and
           | feature rich than Ollama.
        
             | terhechte wrote:
             | Fully agree to this. LM Studio is much nicer to use and
             | with MLX faster on Apple Silicon
        
         | jwr wrote:
         | I just use my laptop. A modern MacBook Pro will run ~30B models
         | very well. I normally stick to "Max" CPUs (initially for more
         | performance cores, recently also for the GPU power) with 64GB
         | of RAM. My next update will probably be to 128GB of RAM,
         | because 64GB doesn't quite cut it if you want to run large
         | Docker containers _and_ LLMs.
        
         | sumo43 wrote:
         | Try running this using their harness
         | https://huggingface.co/flashresearch/FlashResearch-4B-Thinki...
        
         | 3abiton wrote:
         | As many pointed out, Macs are decent enough to run them (with
         | maxxed rams). You also have more alternative, like DGX Sparks
         | (if you appreciate the ease of cuda, albeit a tad bit slower
         | token generation performance), or the Strix Halo (good luck
         | with ROCm though, AMD still peddling hype). There is no
         | straitghtforwars "cheap" answer. You either go big (gpu
         | server), or compromise. Either way use either vllm or sglang,
         | or llama.cpp. ollama is just inferior in every way to
         | llama.cpp.
        
       | mehdibl wrote:
       | It's a Qwen 3 MoE fine tune...
        
       | zurfer wrote:
       | It makes me wonder if we'll see an explosion of purpose trained
       | LLMs because we hit diminishing returns on invest with pre
       | training or if it takes a couple of months to fold these
       | advantages back into the frontier models.
       | 
       | Given the size of frontier models I would assume that they can
       | incorporate many specializations and the most lasting thing here
       | is the training environment.
       | 
       | But there is probably already some tradeoff, as GPT 3.5 was
       | awesome at chess and current models don't seem trained
       | extensively on chess anymore.
        
         | deepanwadhwa wrote:
         | -> GPT 3.5 was awesome at chess I don't agree with this. I did
         | try to play chess with GPT3.5 and it was horrible. Full of
         | hallucinations.
        
           | miki123211 wrote:
           | It was GPT-3 I think.
           | 
           | As far as I remember, it's post-training that kills chess
           | ability for some reason (GPT-3 wasn't post-trained).
        
         | alephnerd wrote:
         | > if we'll see an explosion of purpose trained LLMs...
         | 
         | Domain specific models have been on the roadmap for most
         | companies for years now for both competitive (why give up your
         | moat to OpenAI or Anthropic) and financial (why finance
         | OpenAI's margins) perspective.
        
         | onlyrealcuzzo wrote:
         | Isn't the whole point of the MOE architecture exactly this?
         | 
         | That you can individually train and improve smaller segments as
         | necessary
        
           | idiotsecant wrote:
           | I think it's the exact opposite - you don't specifically
           | train each 'expert' to be a SME at something. Each of the
           | experts is a generalist but becomes better at portions of
           | tasks in a distributed way. There is no 'best baker', but
           | things evolve toward 'best applier of flour', 'best kneader',
           | etc. I think explicitly domain-trained experts are pretty
           | uncommon in modern schemes.
        
             | viraptor wrote:
             | That's not entirely correct. Most of moe right now are
             | fully balanced, but there is an idea of a domain expert moe
             | where the training benefits fewer switches.
             | https://arxiv.org/abs/2410.07490
        
           | ainch wrote:
           | Generally you train each expert simultaneously. The benefit
           | of MoEs is that you get cheap inference because you only use
           | the active expert parameters, which constitute a small
           | fraction of the total parameter count. For example Deepseek
           | R1 (which is especially sparse) only uses 1/18th of the total
           | parameters per-query.
        
         | criemen wrote:
         | > or if it takes a couple of months to fold these advantages
         | back into the frontier models.
         | 
         | Right now, I believe we're seeing that the big general-purpose
         | models outperform approximately everything else. Special-
         | purpose models (essentially: fine tunes) of smaller models make
         | sense when you want to solve a specific task at lower
         | cost/lower latency, and you transfer some/most of the abilities
         | in that domain from a bigger model to a smaller one. Usually,
         | people don't do that, because it's a quite costly process, and
         | the frontier models develop so rapidly, that you're perpetually
         | behind them (so in fact, you're not providing the best possible
         | abilities).
         | 
         | If/when frontier model development speed slows down, training
         | smaller models will make more sense.
        
       | rokob wrote:
       | This whole series of work is quite cool. The use of `word-break:
       | break-word;` makes this really hard to read though.
        
         | soared wrote:
         | I actually can't read it for some reason? My brain just can't
         | connect the words
        
           | don-bright wrote:
           | so it appears the entire text has been Translated with non-
           | breaking space unicode x00a0 instead of normal spaces x0020,
           | so the web layout is considering all paragraph text as a
           | super-long single word ('the\00a0quick\00a0\brown\00a0fox'
           | instead of 'the quick brown fox') - the non-breaking space
           | character appears identically to breaking-space when rendered
           | but underlying coding breaks the concept of "break at end of
           | word" because there is no end as 00a0 literally means "non-
           | breaking"). per Copilot spending a half hour explaining this
           | to me, apparently this can be fixed by opening web browser
           | developer view, and copy/pasting this code into the console.
           | 
           | function replaceInTextNodes(node) { if (node.nodeType ===
           | Node.TEXT_NODE) { node.nodeValue = node.nodeValue
           | .replace(/\u00A0/g, ' '); } else {
           | node.childNodes.forEach(replaceInTextNodes); } }
           | 
           | replaceInTextNodes(document.body);
        
           | dlisboa wrote:
           | That's why typography matters. You can't read it because a
           | very basic convention has been broken here and that throws
           | everything off.
        
       | theflyestpilot wrote:
       | I hope the translation for this is actually "Agree" Deep
       | research. Just a dig at "You are absolutely right!" sycophancy.
        
         | numpad0 wrote:
         | TIL the "full" name of Alibaba Qwen is Tong Yi Qian Wen
         | (romanized as "Tongyi Qianwen", something along "knows all
         | thousand questions"), of which the first half without the
         | Chinese accent flags is romanized identically to "Tong Yi ",
         | meaning "same intents" or "agreed".
         | 
         | The Chinese version of the link says "Tong Yi  DeepResearch" in
         | the title, so doesn't look like the "agree" to be the case.
         | Completely agreed that it would be hilarious.
         | 
         | 1: https://www.alibabacloud.com/en/solutions/generative-
         | ai/qwen...
        
           | rahimnathwani wrote:
           | For people who don't read Chinese: the two 'yi' characters
           | numpad0 mentioned (Yi  and Yi ) are the same, but written in
           | different variants of Chinese script
           | (Simplified/Traditional).
        
       | Traubenfuchs wrote:
       | It still feels to me like OpenAI has zero moat. There are like 5
       | paid competitors + open source models.
       | 
       | I switch between gemini and ChatGpt whenever I feel one fails to
       | fully grasp what I want, I do coding in claude.
       | 
       | How are they supposed to become the 1 trillion dollar company
       | they want to be, with strong competition and open source
       | disruptions every few months?
        
         | rokob wrote:
         | I don't know if they can pull it off but a lot of companies are
         | built on strong enterprise sales being able to sell free stuff
         | with a bow on it to someone who doesn't know better or doesn't
         | care.
        
         | isoprophlex wrote:
         | Premium grade deals with Oracle. They will bullshit their way
         | into government and enterprise environments where all the key
         | decision makers are clueless and/or easily manipulated.
        
         | nickpinkston wrote:
         | Yea, I agree.
         | 
         | Arguably LLMs are both (1) far easier to switch between models
         | than it is today to switch from AWS / GCP / Azure systems, and
         | (2) will be rapidly decreasing switching costs for your legacy
         | systems to port to new ones - ie Oracle's, etc. whole business
         | model.
         | 
         | Meanwhile, the whole world is building more chip fabs, data
         | centers, AI software/hardware architectures, etc.
         | 
         | Feels more like we're headed to commodification of the compute
         | layer more than a few giant AI monopolies.
         | 
         | And if true, that's actually even more exciting for our
         | industry and "letting 100 flowers bloom".
        
         | whiplash451 wrote:
         | Isn't the moat in the product/UI/UX? I use Claude daily and
         | love the "scratch notebook" feel of it. The barebone model does
         | not get you any of this.
        
           | hamandcheese wrote:
           | I agree that the scaffolding around the model contributes
           | greatly to the experience. But it doesn't take billions of
           | dollars in GPUs to do that part.
        
         | Grimblewald wrote:
         | Of course they dont, the only advantage it ever had was the
         | willingness to destroy trust on the internet by scraping
         | everything from everyone rules and expectations be dammed.
         | 
         | The underlying architecture isnt special, the underlying skills
         | and tools aren't special.
         | 
         | There is nothing openAI brings to the table other than a
         | willingness to lie, cheat, and steal. That only gives you an
         | edge for so long.
        
         | red2awn wrote:
         | The moat of OpenAI is 1. internal knowledge they've built over
         | the last few years building front tier models 2. their talent
         | 3. the ChatGPT brand (go ask a random person on the street,
         | they know ChatGPT but not Claude or Gemini)
        
       | steveny3456 wrote:
       | Juju
        
       | krystofee wrote:
       | Isnt it huge deal, that this 30B model can compare and surpass
       | huge closed models?
        
       | tbruckner wrote:
       | Has anyone found these deep research tools useful? In my
       | experience, they generate really bland reports don't go much
       | further than summarization of what a search engine would return.
        
         | ainch wrote:
         | The reports are definitely bland, but I find them very helpful
         | for discovering sources. For example, if I'm trying to ask an
         | academic question like "has X been done before," sending
         | something to scour the internet and find me examples to dig
         | into is really helpful - especially since LLMs have some base
         | knowledge which can help with finding the right search terms.
         | It's not doing all the thinking, but those kind of broad
         | overviews are quite helpful, especially since they can just run
         | in the background.
        
           | kmarc wrote:
           | I caught myself that most of my LLM usage is like this:
           | 
           | ask a loaded, "filter question" I more or less know the
           | answer for, and mostly skip the prose and get to the links to
           | its sources.
        
         | andy99 wrote:
         | My experience is the same as yours. It feels to me (similar to
         | most LLM writing) like they write for someone who's not going
         | to read it or use it but is going to glance at it and judge the
         | quality that way and assume it's good.
         | 
         | Not to different from a lot of consulting reports, in fact, and
         | pretty much of no value if if you're actually trying to learn
         | something.
         | 
         | Edit to add: even the name "deep research" to me feels like
         | something defined to appeal to people who have never actually
         | done or consumed research, sort of like the whole "phd level"
         | thing.
        
         | blaesus wrote:
         | "Summarization of what a search engine would return" is good
         | enough for many of my purposes though. Good for breaking into
         | new grounds, finding unknown unknowns, brainstorming etc.
        
         | criemen wrote:
         | I tend to use them when I'm looking to buy something of
         | category X, and want to get a market overview. I can then still
         | dig in and decide whether I consider the sources used
         | trustworthy or not, and before committing money, I'll read some
         | reviews myself, too. Still, it's a speedup for me.
        
       | DataDaemon wrote:
       | Unfortunately soon China will take lead in AI.
        
         | aeve890 wrote:
         | Unfortunately? May I ask why? What country would you like to be
         | the lead in AI?
        
           | ninetyninenine wrote:
           | The USA of course. Isn't it obvious? What other country is
           | more Free and great? None. Why does this even need to be
           | asked?
           | 
           | China is full of people who want communism to dominate the
           | world with totalitarian control so no one wants China to
           | dominate anything at all because they are bad...
        
             | Krasnol wrote:
             | The USA is being led by a criminal pedo atm. There is
             | military in the streets and SA-like, masked thugs are
             | kidnapping people. Billionaires sit behind the wheels to
             | profit from all those developments. Many of them are
             | somehow related to AI. You can image what that will be/is
             | used (see Palantir).
             | 
             | The whole country is going down the drain right now. There
             | is nothing about it, sane people outside the Republican
             | bubble would consider "freedom".
        
             | victorbjorklund wrote:
             | USA is threatening to invade Europe so not sure it can be
             | considered great.
        
         | davidsainez wrote:
         | I have been very impressed with the Qwen3 series. I'm still
         | evaluating them, and I generally take LLM benchmarks with a
         | huge grain of salt, but their MoE models in particular seem to
         | offer a lot of bang for the compute. But what makes you so sure
         | they will take the lead?
        
         | ninetyninenine wrote:
         | Isn't this an indication they are already in the lead? They
         | currently have the best model that beats everyone on all
         | quantitative metrics? Are you implying that the US has a better
         | model somewhere?
        
       | sumo43 wrote:
       | I made a 4B Qwen3 distill of this model (and a synthetic dataset
       | created with it) a while back. Both can be found here:
       | https://huggingface.co/flashresearch
        
         | Nymbo wrote:
         | Just tried this out with my web search mcp, extremely impressed
         | with it. Never seen deep research this good from a model so
         | small.
        
       | brutus1213 wrote:
       | I recently got a 5090 with 64 GB of RAM (intel cpu). Was just
       | looking for a strong model I can host locally. If I had
       | performance of GPT4-o, I'd be content. Are there any suggestions
       | or cases where people got disappointed?
        
         | p1esk wrote:
         | 5090 has 32GB of RAM. Not sure if that's enough to fit this
         | model.
        
           | svnt wrote:
           | It should fit enough of the layers to make it reasonably
           | performant.
        
           | IceWreck wrote:
           | LlamaCPP supports offloading some experts in a MoE model to
           | CPU. The results are very good and even weaker GPUs can run
           | larger models at reasonable speeds.
           | 
           | n-cpu-moe in https://github.com/ggml-
           | org/llama.cpp/blob/master/tools/serv...
        
         | bogtog wrote:
         | GPT-OSS-20B at 4- or 8-bits is probably your best bet?
         | Qwen3-30b-a3b probably the next best option. Maybe there exists
         | some 1.7 or 2 bit version of GPT-OSS-120B
        
       | yalogin wrote:
       | In my experience using these supposed expert models, they are all
       | more or less the same given they all are trained on the same
       | internet data. The differentiation and value is in the context
       | window management and how relevant info from your session is
       | pulled in. So it's the interface to the model that makes all the
       | difference. Even there the differences are quite minimal. That is
       | because all these companies want to toe the line between
       | providing functionality to keep the users engaged and pushing
       | them to sign up for the subscription.
       | 
       | All this to ask the question, if I host these open source models
       | locally, how is the user interface layer that remembers and picks
       | the right data from my previous session and the agentic
       | automation and others implemented? Do I have to do it myself or
       | are the free options for that?
        
         | viksit wrote:
         | this is a great question. what are the main use cases that you
         | have for this? i've been working on a library for something
         | similar and exposing it via an mcp interface. would love to
         | pick your brain on this (@viksit on twitter)
        
       | ninetyninenine wrote:
       | Is China dominating the US in terms of AI? Given that they
       | currently have a model that beats the best models at all formal
       | quantitative benchmarks?
       | 
       | What is the state of AI in China? My personal feeling is that it
       | doesn't dominate the zeitgeist in China as it does in the US and
       | despite this because of the massive amount of intellectual
       | capital they have just a small portion of their software
       | engineering talent working on this is enough to go head to head
       | with us even though it only takes a fraction of their attention.
        
         | idiotsecant wrote:
         | I think the lesson of the Chinese catchup in AI is that there
         | is a massive disadvantage in being first, in this domain. You
         | can do all the hard work and your competitors can distill that
         | work out of your model for pennies on the dollar. Why should
         | anyone want to do the work?
        
           | MaxPock wrote:
           | This sounds like copium . If it was just about
           | distillation,we'd be seeing many awesome models from Europe
           | ,Japan and even India.
        
       | whiplash451 wrote:
       | Has anyone tried running this on a 5090 or 6000 pro? What
       | throughput do you see?
        
       ___________________________________________________________________
       (page generated 2025-11-02 23:00 UTC)