[HN Gopher] Bagel: Open-source unified multimodal model
       ___________________________________________________________________
        
       Bagel: Open-source unified multimodal model
        
       Author : tosh
       Score  : 205 points
       Date   : 2025-05-26 05:51 UTC (17 hours ago)
        
 (HTM) web link (bagel-ai.org)
 (TXT) w3m dump (bagel-ai.org)
        
       | mnky9800n wrote:
       | I couldn't find it, what are the hardware expectations for bagel?
        
         | gunalx wrote:
         | If you follow the hugginface link at the bottom you get to te
         | actual model. Here https://huggingface.co/ByteDance-
         | Seed/BAGEL-7B-MoT
         | 
         | It seems to be 7b, but like with other new architectures expect
         | to not be able to run it quantizised.
        
         | tonii141 wrote:
         | If the model uses FP16 precision and has 7 billion active
         | parameters, it would require approximately 14 GB of VRAM. I
         | didn't read the paper.
        
           | sfphoton wrote:
           | How can you calculate required VRAM from precision and
           | parameter number?
        
             | a_t48 wrote:
             | A float16 is 2 bytes. 7B * 2 bytes = 14GB. I can't say if
             | that's an accurate number, but that's almost certainly how
             | tonii141 calculated it.
        
               | sfphoton wrote:
               | Oh, so FP16 means FloatingPoint16? I'm glad to learn
               | something today, thanks!
        
             | NitpickLawyer wrote:
             | Rule of thumb is parameter_count * precision. Precision can
             | be anything [32,16,8,4] bits. 32bits is sometimes used in
             | training (although less now I guess), and rarely in
             | inference. For a while now "full" precision is 16bit (fp16,
             | bf16), fp8 is 8bit, int4 is 4bit, and so on. Everything
             | that's not "full" precision is also known as quantised. fp8
             | is a quantised version of the "full" model.
             | 
             | So quick napkin math can give you the VRAM usage for
             | _loading_ the model. 7b can be ~14GB full, 7GB in fp8 and
             | ~3.5GB in 4bit (AWQ, int4, q4_k_m, etc). But that 's just
             | to load the model in VRAM. You also need some available
             | VRAM to run inference, and there are a lot of things to
             | consider there too. You need to be able to run a forward
             | pass on the required context, you can keep a kv cache to
             | speed up inference, you can do multiple sessions in
             | parallel, and so on.
             | 
             | Context length is important to take into account because
             | images take a lot of tokens. So what you could do with a 7b
             | LLM at full precision on a 16GB VRAM GPU might not be
             | possible with a VLM, because the context of your query
             | might not fit into the remaining 2GB.
        
             | Havoc wrote:
             | Realistically you probably just want to look at the file
             | size on huggingface and add ~2 gigs for OS/Firefox tabs and
             | and a bit for context (depends but lets say 1-2)
             | 
             | The direct parm conversion math tends to be much less
             | reliable than one would expect once quants are involved.
             | 
             | e.g.
             | 
             | 7B @ Q8 = 7.1gb [0]
             | 
             | 30B @ Q8 = 34.6gb [1]
             | 
             | btw you can also roughly estimate expected output speed too
             | if you know the device memory throughput. Noting that this
             | doesn't work for MoEs
             | 
             | Also recently discovered that in CPU mode llama.cpp does
             | memory mapping. For some models it loads less than a
             | quarter into memory.
             | 
             | https://huggingface.co/TheBloke/Llama-2-7B-GGUF/tree/main
             | 
             | https://huggingface.co/TheBloke/LLaMA-30b-GGUF/tree/main
        
       | LourensT wrote:
       | These days, papers come with an advertisement video
        
         | jxjnskkzxxhx wrote:
         | As someone who used to be in the academia, I think is isn't bad
         | _in itself_ , I just worry that by comparison it raises the
         | burden of effort that one has to make in order to get their
         | work noticed.
        
           | kleiba wrote:
           | Compared to the effort required to play in that field _at
           | all_ , making a video is almost negligible.
        
             | jxjnskkzxxhx wrote:
             | If it's an obligation, it's admin. Scientists _hate_ admin.
        
         | lern_too_spel wrote:
         | This has been common for CG papers for two decades. Image
         | generation is CG.
        
       | charcircuit wrote:
       | The demo shows pretty weak performance compared to other small
       | models. It misunderstood my question due to picking an uncommon
       | way to interpret it. After clarifying what I wanted it lost all
       | context I had provided in the previous message. My benchmark
       | query intentionally ambiguous and I use it to see how models
       | handle ambiguity, handle information which can be outdated, and
       | handle avoiding hallucination. Usually weak models will just
       | hallucinate an answer, but this model was the first who want able
       | to understand the question.
        
       | saretup wrote:
       | > Scala _b_ le Perceptu _a_ l _G_ enerative Mod _el_
       | 
       | If you wanna call it Bagel, just call it Bagel. No need to make
       | up a justification.
        
       | spuz wrote:
       | I'm interested in potential alternatives to ChatGPT's advanced
       | voice mode. When I see the word "multimodal" I'm hopeful the
       | model understands text + voice but instead it almost always seems
       | to refer to text + images. Is there a keyword that I can use to
       | look for models that work with voice similar to ChatGPT's
       | advanced voice mode?
        
         | amrrs wrote:
         | Google Gemini Live is pretty good.
         | 
         | If you want to try only voice, Try unmute.sh by Kyutai which
         | will be eventually open-sourced
        
         | cjbprime wrote:
         | I don't know that ChatGPT's voice mode is using audio as a
         | transformer input directly.
         | 
         | It could just be using speech to text (e.g. Whisper) on your
         | input, and then using its text model on the text of your words.
         | Or has OpenAI said that they aren't doing this?
        
           | mrshu wrote:
           | OpenAI does not provide many details about their models these
           | days but they do mention that the "Advanced voice" within
           | ChatGPT operates on audio input directly:
           | 
           | > Advanced voice uses natively multimodal models, such as
           | GPT-4o, which means that it directly "hears" and generates
           | audio, providing for more natural, real-time conversations
           | that pick up on non-verbal cues, such as the speed you're
           | talking, and can respond with emotion.
           | 
           | From https://help.openai.com/en/articles/8400625-voice-mode-
           | faq
        
       | moffkalast wrote:
       | Oh no it's The Everything Bagel.
        
       | gregjw wrote:
       | bagel
        
       | pleone wrote:
       | Is it from ByteDance Team, right? The team behind TikTok, CapCut,
       | BuzzVideo and more. Any thoughts on that?
        
         | rvnx wrote:
         | Like BYD vs Tesla. US is getting more and more late and more
         | closed than ever (e.g. Chinese Qwen LLM versus LLaMA). So long-
         | term, China may emerge as the dominating force in tech.
        
       | mdrzn wrote:
       | A quick test in the "demo" link doesn't show it to be "as smart"
       | as it appeared in the demos on the page. I really hope it does
       | all it's promising to do, but I'm skeptic so far.
        
         | mrec wrote:
         | I found it surprising that even one of the demos on the page
         | appeared to get it wrong. (Chat example #5, explaining the "My
         | Handwriting In Exams" meme.) Not _horribly_ wrong, but still an
         | odd example to cherry-pick for publicity material.
         | 
         | ETA: oof, and it's still getting hands wrong. (Editing demo
         | #12)
        
       | sandra_vu wrote:
       | Hi good job, team. Any plans to commercialize the model?
        
       | wsintra2022 wrote:
       | https://news.ycombinator.com/item?id=44063602
        
       | jjrv wrote:
       | I found a losslessly compressed version:
       | https://github.com/LeanModels/Bagel-DFloat11
       | 
       | It works following readme instructions at least on Ubuntu, on my
       | RTX 3090 GPU with 24 gigs of memory, just barely. Have to close
       | most other windows and lower screen resolution to be able to load
       | the model. Then it generates or edits images in 2-3 minutes. I
       | only have this one GPU and am using Chrome to use the browser
       | interface on the same machine.
       | 
       | The original release won't run on this hardware, but the
       | compressed one is supposed to give identical results.
        
         | jjrv wrote:
         | I also asked it to explain what's funny in some newspaper comic
         | strips in Finnish. It misunderstands some words and makes up
         | nutty explanations, but most phrases still get translated
         | correctly and its explanations do fit the drawn scenes once you
         | factor in those misunderstandings. For such a small model that
         | seemed impressive.
        
       | GrantMoyer wrote:
       | Nice, it's really an open source model, Apache 2.0.
        
       | akacrobat wrote:
       | This looks exciting! There is a serious dearth of high-quality
       | open-source models with multimodal capabilities. So, really
       | looking forward to playing with this one.
       | 
       | Has anyone here experimented with fine-tuning this for domain-
       | specific applications?
        
       | akoculu wrote:
       | Good summary of the paper:
       | https://x.com/build__ship/status/1926930191185580176
        
       ___________________________________________________________________
       (page generated 2025-05-26 23:02 UTC)