List of open models released or updated this week on this sub, just in case you missed one.
Posted by aifeed-fyi@reddit | LocalLLaMA | View on Reddit | 28 comments
A quick list of models updates and new releases mentioned in several posts during the week on LocalLLama. I wanted to include links to posts/models but it didn't go through.
- Kimi K2-0905 – new release from Moonshot AI
- Wayfarer 2 12B & Nova 70B – open-sourced narrative roleplay models from AI Dungeon
- EmbeddingGemma (300M) – Google’s compact multilingual embedding model
- Apertus – new open multilingual LLM from ETH Zürich (40%+ non-English training data)
- WEBGEN-4B – web design generation model trained on 100k synthetic samples
- Lille (130M) – a truly open-source small language model (trained fully from
- Hunyuan-MT-7B & Hunyuan-MT-Chimera-7B – Tencent’s new translation & ensemble models
- GPT-OSS-120B – benchmarks updates
- Beens-MiniMax (103M MoE) – scratch-built, SFT + LoRA experiments
KaroYadgar@reddit
Thanks for this post! It's difficult keeping up with everything.
One-Employment3759@reddit
Especially on this sub which is now mostly full of slop and closed models.
Zc5Gwu@reddit
Agreed, super helpful. Links would be even more killer.
aifeed-fyi@reddit (OP)
Fully, it had links but it was auto removed so I posted without
aifeed-fyi@reddit (OP)
I can relate :)
ThomasPhilli@reddit
Do you have the link for Beens-Minimax? I cannot find it and would love to try out!
pmttyji@reddit
Please update your list with below items
AND
Thedrummer is on fire!!!
swagonflyyyy@reddit
Webgen-4b was interesting but dissappointing when I tested it online. It just got stuck in an infinite loop of code and sometimes the markdown broke in between.
Its an interesting fine-tune but they still have a ways to go.
No_Efficiency_1144@reddit
New Kimi great
LivingHighAndWise@reddit
You running it in Ollama? I tried pulling it but it doesn't appear to exist.
Spectrum1523@reddit
isn't it absolutely huge? what are you running it on
LivingHighAndWise@reddit
Yes.. I did find it and started downloading it, but it was over 700 GB lol.
No_Efficiency_1144@reddit
Haven’t tried yet
simplir@reddit
Excited to test Gemma embeddings performance on a clustering project I'm working one.
-Cubie-@reddit
Remember to use the prompts. There's some explanations on the HF blogpost about it.
OnanationUnderGod@reddit
can you link it? this is the internet. thanks!
aifeed-fyi@reddit (OP)
Thanks for letting me know. Still didn't see it, I just saw the release last night
kantydir@reddit
It's pretty good for its size but not even close to bigger models like Qwen3 Embedding 4B. It's very picky with the query prompt though, be careful.
Klutzy-Snow8016@reddit
A couple more new ones:
Meituan Longcat Flash: https://huggingface.co/meituan-longcat/LongCat-Flash-Chat
Kwai Clear: https://huggingface.co/collections/Kwai-Klear/klear10-68ba61398a0a4eb392ec6ab1
Plums_Raider@reddit
Still waiting to try apertus... hopefully soon updates will follow for us gguf guys
aifeed-fyi@reddit (OP)
Same here, want to see if its up to the promise
FullOf_Bad_Ideas@reddit
Did they promise much? 70B is below Olmo 32B on English tasks, but a bit above Olmo 32B on multilingual.
It's still way behind Llama 3 70B for example.
They have a good license on it, and it's more open so it's interesting, but I don't think there's much stuff about the model weights themselves that sound exciting.
Plums_Raider@reddit
Even if not. I as a swiss person want to test if its better at speaking swiss german than the imo swiss german sota model gemini lol
aifeed-fyi@reddit (OP)
Lol, makes sense
FullOf_Bad_Ideas@reddit
Baichuan-M2 is a new RL-trained model optimized for HealthBench.
It released about a month ago, but paper came out just now.
https://arxiv.org/abs/2509.02208
smirkishere@reddit
Thanks for the shoutout for WEBGEN :) We're working hard on getting the non-preview version out and more UIGEN and a new series out soon!! (And bigger models 4Bs are our internal thermometers to check how our pipelines working).
aifeed-fyi@reddit (OP)
You guys are doing great, I saw few demos on X. Still need to test it myself but seems very promising
Namra_7@reddit
Waiting for big qwen model 🙂