[HN Gopher] OpenAI Leaks 120B Open Model on Hugging Face
       ___________________________________________________________________
        
       OpenAI Leaks 120B Open Model on Hugging Face
        
       Author : skadamat
       Score  : 110 points
       Date   : 2025-08-01 15:44 UTC (7 hours ago)
        
 (HTM) web link (twitter.com)
 (TXT) w3m dump (twitter.com)
        
       | yieldcrv wrote:
       | *uploads
        
         | ipsum2 wrote:
         | Accidentally reveals.
        
           | yieldcrv wrote:
           | accidentally on purpose
        
             | vntok wrote:
             | Why not mere ineptitude?
        
         | seydor wrote:
         | But with an NDA so that leaks can be legit
        
           | kristianp wrote:
           | What's the chances someone under NDA has leaked the url to
           | the xeeter in question?
        
       | 44za12 wrote:
       | I don't get the hype with OpenAI OSS, they would never make a
       | model better than their proprietary models open source, and the
       | other open source models beat GPT and family so why the wait?
        
         | jphoward wrote:
         | I think they could release non-agentic models that are as good
         | as 4o, and have almost no repercussions on sales tbh.
         | 
         | I have Ollama installed (only a small proportion of their
         | clients would have a large enough GPU for this) and have
         | download deepseek and played with it, but I still pay for an
         | OpenAI subscription because I want the speed of a hosted model,
         | and never mind the luxuries of things like Codex's diffs/pull
         | request support, agents on new models, deep research etc. - I
         | use them all at least weekly.
        
           | 44za12 wrote:
           | I pay for Cursor, OpenAI and kimi (to use with Claude Code),
           | OpenAI is good with quickly refining my thoughts, Cursor's
           | subscription I'm reconsidering to cancel bought it for Claude
           | but the rate limits are making it impossible for me to find
           | it useful. Kimi is what truly surprises me, Claude code shows
           | this conversation costed you $500 (based on Opus usage which
           | is mapped to kimi k2) while I've barely spent $2. I have
           | Ollama as well, majorly to quickly test small models that
           | could be improved for our usecase through finetuning.
        
             | garciasn wrote:
             | What am I doing wrong that I'm never hitting the rate
             | limits on the $100 Max plan?
        
               | Topfi wrote:
               | Considering my personal heavy use also not leading to
               | rate limits and what I've seen by some users over the
               | past months, I suspect a mix of actually thinking about
               | your code before writing a prompt, managing context by
               | documenting and running stuff like git, npm install, etc.
               | yourself instead of "Hey Claude, setup React with Radix
               | and install a few packages". I have genuinely seen
               | someone use ultrathink for setting up a starter repo
               | hosted on Github, despite the commands being listed in
               | the readme, so I can see how certain people may hit the
               | limits quicker than others. Still, I will cancel my
               | Claude Max subscription if they remain intransparent
               | concerning the amount of use we actually get, especially
               | regarding the mail they sent out recently which stated
               | that 20x Max users do only get 10x in terms of expected
               | usable hours. Same goes for still not providing an
               | official way to track how much use one has left in a
               | week.
        
               | garciasn wrote:
               | > running stuff like git, npm install, etc. yourself
               | 
               | Ah; this definitely makes sense! I do this myself and
               | then paste back only the relevant part of the log so as
               | to limit this. I suspect I am being more conservative
               | than others.
        
               | 44za12 wrote:
               | I am on the pro plan, I was considering Max, but then i
               | found kimi and I'm getting used to it.
        
               | nico wrote:
               | Are you using kimi with Claude Code? Are you using it via
               | OpenRouter?
        
               | 44za12 wrote:
               | With claude code and Kilo as well. I'm using moonshot's
               | API.
        
             | ewoodrich wrote:
             | I've been using Kimi with Roo via OpenRouter and have been
             | very surprised at how capable it is. It's the first open
             | model I've tried that actually lives up to claims I see
             | online that's it on par with this or that previous gen
             | proprietary model. Context window has been the only
             | negative, at least with the providers OpenRouter has been
             | giving me but forgiveable given how absurdly cheap it is.
        
               | nico wrote:
               | > but forgiveable given how absurdly cheap it is
               | 
               | Are you using it everyday for programming? If so, how
               | much more or less does it cost you per month? More or
               | less than $100?
        
           | vineyardmike wrote:
           | They would definitely have sales repercussions, but it might
           | be worth it.
           | 
           | They are fully trying to be a consumer product, developer
           | services be damned. But they can't just get rid of the API
           | because it's a good incremental source of revenue, and thanks
           | to the Microsoft deal, all that revenue would end up in
           | Azure. Maintaining their API is basically just a way to get a
           | slice of that revenue.
           | 
           | But if they open sourced everything, it might sour the
           | relationship more with Microsoft, who would lose azure
           | revenue and might be willing to part ways. It would also
           | ensure that they compete on consumer product quality not
           | (directly) model quality. At this point, they could basically
           | put any decent model in their app and maintain the user base,
           | they don't actually need to develop their own.
        
         | Topfi wrote:
         | Pure performance isn't necessarily everything. Context window,
         | speed and local use are just some of the upsides this model may
         | have. We still know next to nothing so anything is possible,
         | but if it is an MoE at 120B, that could enable some interesting
         | local use cases, even if it less capable than e.g. Deepseek V3,
         | simply by running on more hardware/at higher tokens/sec.
         | GPT-4.1s code focus has also shown that OpenAI does have a
         | knack for models with a more narrow use case, maybe this will
         | do well in specific tasks. More so since GPT-4.1 was that much
         | better than the massive GPT-4.5, I am cautiously optimistic.
         | 
         | Even if it does poorly in all areas (like Llama 4 [0]), there
         | is still a lot the community and industry can learn from even
         | an uncompetitive model.
         | 
         | [0] Llama 4 technically has a massive 10M token context as a
         | differentiator, however in my experience, it is not reliably
         | usable beyond 100k.
        
         | jstummbillig wrote:
         | I don't see how it would be in OpenAIs selfish interest to
         | release an open source model that sucks. Unless you can
         | cohesively explain how that would work in their favor, it seems
         | a lot smarter to assume that they won't.
        
         | granitepail wrote:
         | While the benchmarks all say open source models Kimi and Qwen
         | outpace proprietary models like GPT 4.1, GPT 4o, or even o3, my
         | (and just about everyone I know's) boots on the ground
         | experience suggests they're not even close. This is for tool
         | calling agentic tasks, like coding, but also in other contexts
         | (research, glue between services, etc). I feel like it's worth
         | putting that out there--it's pretty clear there's a lot of
         | benchmark hacking happening. I'm not really convinced it's
         | purposeful/deceitful, but it's definitely happening. Qwen3
         | Coder, for example, is basically incompetent for any real
         | coding tasks and frequently gets caught in death spirals of bad
         | tool calls. I try all the OSS models regularly, because I'm
         | really excited for them to get better. Right now Kimi K2 is the
         | most usable one, and I'd rate it at a few ticks worse than GPT
         | 4.1.
        
           | jimbo808 wrote:
           | I would have assumed anyone frequenting HN would have figured
           | out by now that benchmarks are 100% bullshit. I guess I'd be
           | wroing.
        
             | dist-epoch wrote:
             | So what do you propose? Gut feel, N=1 tests?
        
               | spullara wrote:
               | it currently beats depending on the benchmarks
        
               | BoorishBears wrote:
               | I mean, in other environments people say that.
               | 
               | If you asked "What's the best bicycle", most enthusiasts
               | would say one you tried, works for your usecase, etc.
               | 
               | Benchmarks should be for pruning models you try at the
               | absolute highest level, because at the end of the day
               | it's way too easy to hack them without breaking any rules
               | (post-train on the public, generate a ton of synthetic
               | examples, train on those, repeat)
        
               | int_19h wrote:
               | At the moment, the only way you can tell if the model is
               | good for a particular task is by trying it at that task.
               | Gut feel is how you pick the models to test first, and
               | that is also based largely on past experience and
               | educated guesses as to what strengths translate between
               | tasks.
               | 
               | You should also remember that there's no free lunch. If
               | you see models below a certain size fail consistently,
               | don't expect a model that is even smaller to somehow
               | magically succeed, no matter how much pixie dust the
               | developer advertises.
        
             | andrewmcwatters wrote:
             | I think anyone frequenting HN and actually using these
             | tools absolutely knows these benchmarks are 100% bullshit
             | and the only real way to test these things is to just use
             | them yourself.
             | 
             | Many small models are supposedly good for controlled tasks,
             | but given a detailed prompt, I can't get any of them to
             | follow simple instructions. They usually just regurgitate
             | the examples in the system prompt. Useless.
        
           | daft_pink wrote:
           | isn't the problem with the benchmarks that most people
           | running ai locally are running way lower weights?
           | 
           | i have an m4 studio with a lot of unified memory and i'm
           | still no where near running a 120b model. i'm at like 30b
           | 
           | apple or nvidia's going to have to sell 1.5 tb ram machines
           | before benchmark performance is going to be comparable
           | 
           | Plus when you use claude or openai, these days it's
           | performing google searches etc that my local model isn't
           | doing.
        
             | BoorishBears wrote:
             | No, I've deployed a lot of open weight models and the gap
             | between closed source is there even at larger sizes.
             | 
             | I'm running a 400B parameter model at FP8 and it still took
             | a lot of post-training to get an even somewhat comparable
             | performance
             | 
             | -
             | 
             | I think a lot of people implicitly bake in some grace
             | because the models are open weights, and that's not
             | unreasonable because of the flexibility... but in terms of
             | raw performance it's not even close.
             | 
             | GPT-3.5 has better world knowledge than some 70B models,
             | and a few even larger.
        
               | daft_pink wrote:
               | you're killing my dream of blowing $50-100k on a desktop
               | supercomputer next year and being able to do everything
               | locally ;)
               | 
               | "the hacker news dream" - a house, 2 kids, and a desktop
               | supercomputer that can run a 700B model.
        
               | meaydinli wrote:
               | Take a look at: https://www.nvidia.com/en-
               | us/products/workstations/dgx-spark... . IIRC, it was
               | about ~$4K.
        
               | phonon wrote:
               | An M4 Max twice the memory bandwidth (which is typically
               | the limiting factor)
        
               | BoorishBears wrote:
               | I'll say neither of them will do anything for you if
               | you're currently using SOTA closed models in anger and
               | expect that performance to hold.
               | 
               | I'm on a 128GB M4 Max, and running models locally is a
               | curiosity at best given the relative performance.
        
             | refulgentis wrote:
             | I'm so darn confused on local LLMs and M-series inference
             | speed, the perf jump from M2 Max to M4 Max was negligible,
             | 10-20%. (both times MBP, 64 GB and max gpu cores)
        
             | granitepail wrote:
             | In my case, I'm paying for inference on the original models
             | from e.g. Fireworks. So it's not a quantization problem.
             | The Qwen3 I was using was the new 458B (i think that's the
             | size?) model that was their top performer for code.
             | 
             | I agree with other comments that there are productive uses
             | for them. Just not on the scale of o4-mini/o3/claude 4
             | sonnet/opus.
             | 
             | So imo open weights larger models from big US labs is a big
             | deal! Glad to see it. Gemma models, for example, are great
             | for their size. They're just quite small.
        
           | n_kr wrote:
           | It may be the way I use it, but qwen3-coder (30b with ollama)
           | is actually helping me with real world tasks. Its a bit worse
           | than big models for the way I use it, but absolutely useful.
           | I do use ai tools with very specific instructions though,
           | like file paths, line numbers if I can, and specific
           | direction about what to do, my own tools, etc. so that may be
           | why I don't see such a huge difference from big models.
           | 
           | I should try Kimi K2 too.
        
             | refulgentis wrote:
             | You'll see good results, Kimi is basically a micro dosing
             | Sonnet lol. V v v reliable tool calls, but, because it's
             | micro dosing, you don't wanna use it for implementing
             | OAuth, maybe adding comments or strict direction (i.e. a
             | series of text mutations)
        
           | torginus wrote:
           | Not sure about benchmarks, but I did use Deepseek when it was
           | novel and cool for a variety of tasks before going back to
           | Claude, and in my experience it was OK, not significantly
           | worse for what I use these models for (writing code small
           | functions at a time, learning about libraries etc.), tham
           | closed stuff at the time.
        
         | rdtsc wrote:
         | > they would never make a model better than their proprietary
         | models open source
         | 
         | Not their proprietary model, but maybe other open source
         | models, or closed source models of their competitors. That way
         | they can first ensure they are the only player on both sides,
         | and then can kneecap their open source models just enough to
         | drive the revenue to their proprietary one.
        
           | 44za12 wrote:
           | Making a model better than proprietary models is in fact
           | making a model better than their closed source models if you
           | believe the benchmarks.
        
       | Nerd_Nest wrote:
       | Whoa, 120B? That's huge.
        
         | qeternity wrote:
         | 120B MoE. The 20B is dense.
         | 
         | As far as dense models go, it's larger than many but Mistral
         | has released multiple 120B dense models, not to mention Llama3
         | 405B.
        
           | sciencesama wrote:
           | How much ram do you need to run this !!??
        
             | cubefox wrote:
             | Probably about one byte per weight (parameter) plus a bit
             | extra for the key-value cache (depends on the size of the
             | context window).
        
               | int_19h wrote:
               | You can go below one byte per parameter. 4-bit
               | quantization is fairly popular. It does affect quality -
               | for some models more so than others - but, generally
               | speaking, a 4-bit quantized model is still going to do
               | significantly better than an 8-bit model with 1/2
               | parameters.
        
       | arnaudsm wrote:
       | Who's the target of 120B open-weights models? You can only run
       | this in the cloud, is it just PR?
       | 
       | I wish they released a nano model for local hackers instead
        
         | xandrius wrote:
         | For people who run stuff on the cloud?
        
         | kccqzy wrote:
         | They are probably hoping that someone else will distill it into
         | smaller models, much like DeepSeek released a giant 671B model
         | but there are useful distillations down to 30B.
        
         | oldge wrote:
         | This sized model is trivial to run on a modern workstation
        
           | dmonitor wrote:
           | You'll have to define modern workstation for me, because I
           | was under the impression that unless you've purpose-built
           | your machine to run LLMs, this size model is impossible.
        
             | wincy wrote:
             | You can run a 4 bit quantized 120B model on a 96GB
             | workstation card, the Blackwell Pro workstation, which are
             | $7500. Considering the 5090 is bought by gamers for $3300
             | it's definitely attainable, even though it's obviously
             | expensive.
             | 
             | I'm running a gaming rig and could swap one in right now
             | without having to change anything compared to my 5090, so
             | no $5000 Threadripper or a $1000 HEDT motherboard with a
             | ton of RAM slots, just a 1000 watt PSU and a dream.
        
             | 0x457 wrote:
             | When people say "modern workstation" in context of LLM,
             | they usually mean its consumer(pro-sumer?) grade hardware
             | on a single machine. As opposed to racks of GPUs that you
             | can even buy as a mere mortal (min order size)
             | 
             | It doesn't mean you can grab your work laptop from 5 years
             | ago and run it there.
        
             | int_19h wrote:
             | Get a Mac Studio with however much memory you need, and
             | ideally an Ultra chip (for max memory bandwidth), and
             | there's your workstation. I regularly run quantized 100b+
             | models on my M1 Ultra with 128Gb RAM.
        
         | 152334H wrote:
         | They have a 20b for GPU poors, too.
         | 
         | I will be running the 120B on my 2x4090-48GB, though.
        
         | segmondy wrote:
         | You can run it locally too. Below are a few of my local models,
         | this is coming in light compared to them. At Q4 it's ~60B.
         | Furthermore being a MoE, most of it can be in system memory and
         | only the shared experts needs to go to GPU, provided you have a
         | decent system with decent memory bandwidth, you can get decent
         | performance. I'm running on GPUs, folks with Apple can run this
         | if they have enough ram with minimal effort.
         | 126G /llmzoo/models/Qwen3-235B-InstructQ4       126G
         | /llmzoo/models/Qwen3-235B-ThinkingQ4       189G
         | /llmzoo/models/Qwen3-235B-InstructQ6       219G
         | /llmzoo/models/glm-4.5-air       240G /llmzoo/models/Ernie
         | 257G /llmzoo/models/Qwen3-Coder-480B       276G
         | /llmzoo/models/DeepSeek-R1-0528-UD-Q3_K_XL.b.gguf       276G
         | /llmzoo/models/DeepSeek-TNG       276G
         | /llmzoo/models/DeepSeek-V3-0324-UD-Q3_K_XL.gguf       422G
         | /llmzoo/models/KimiK2
        
         | jlokier wrote:
         | You can run models the size of this one locally, even on a
         | laptop, it's just not a great experience compared with an
         | optimised cloud service. But it is local.
         | 
         | The size in bytes of this 120B model is about 65 GB according
         | to the screenshot, and elsewhere it's said to be trained in
         | FP4, which matches.
         | 
         | That makes this model small enough to run locally on some
         | laptops without reading from SSD.
         | 
         | The Apple M2 Max 96GB from January 2023, which is two
         | generations old now, has enough GPU-capable RAM to handle it,
         | albeit slowly. Any PC with 96 GB of RAM can run it on the CPU,
         | probably more slowly. Even a PC with less than 64 GB of RAM can
         | run it but it will be much slower due to having to read from
         | the SSD constantly.
         | 
         | If it's a 20B MoE, it will read about one fifth of the data per
         | token, making it about 5x faster than a 120B FP4 non-MoE would
         | be, but it still needs all the data readily available for
         | multiple tokens.
         | 
         | Alternatively, someone can distill and/or quantize the model
         | themselves to make a smaller model. These things can be done
         | locally, even on a CPU if necessary if you don't mind how long
         | it takes to produce the smaller model. Or on a cloud machine
         | rented long enough to make the smaller model, which you can
         | then run locally.
        
       | m_ke wrote:
       | Would be interesting if this was a coding focused model optimized
       | for Mac inference. Would be a great way to undercut Anthropic.
       | 
       | Pretty much give away Sonnet level coding model and have it work
       | with GPT-5 for harder tasks / planning.
        
         | CharlesW wrote:
         | Out of curiosity, have you tried running Qwen3 Coder 30B
         | locally?
         | https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-...
        
       | natas wrote:
       | okay, so where do I download this now that it's been removed from
       | huggingface?
        
       ___________________________________________________________________
       (page generated 2025-08-01 23:01 UTC)