The 65B Fine-Tuned Champion and the 7B Base Merge King
The Hugging Face Open LLM Leaderboard collection "Open LLM Leaderboard best models โค๏ธโ๐ฅ" has become the go-to daily snapshot of which models are actually winning on benchmark text generation. Updated March 13, 2025, the collection sorts models into two clear tag categories: ๐ถ fine-tuned and ๐ค base merges/moerges. Each entry lists the model name, parameter size, task (Text Generation), update date, upvote count, and a note specifying the category and approximate parameter range. The list is community-voted, with upvotes reflecting real human attention โ 700 total upvotes for the collection, and individual models ranging from a handful to tens of thousands of votes.
The Best ๐ถ Fine-Tuned Model of Around 65B
Among the fine-tuned entries, one model stands out as the clear leader for the 65B parameter range: sequelbox/Llama2-70B-SunsetBoulevard. With 69B parameters, it was ranked as the Best ๐ถ fine-tuned model of around 65B on the leaderboard today, updated January 5, 2025, and has already collected 48 upvotes. Its closest fine-tuned competitor in the same ballpark is MaziyarPanahi/calme-3.2-instruct-78b, which holds the Best ๐ถ fine-tuned on domain-specific datasets tag with 78B parameters, updated January 20, 2025, and 470 upvotes โ a massive gap in community recognition. Another domain-specific fine-tuned model, dnhkng/RYS-Llama3.1-Large, runs at 82B parameters, updated September 3, 2024, with just 40 upvotes, showing that scale alone doesn't drive votes; relevance to the benchmark does.
The fine-tuned tag ๐ถ seems to reward models that have been carefully adapted on specific datasets or tasks, rather than simply released at large scale. The SunsetBoulevard model's 48 upvotes suggest it has found a niche among users who want a Llama 2 derivative that still performs well on the leaderboard's text generation prompts. The calme-3.2-instruct model's surprisingly high vote count (470) indicates that domain-specific instruction tuning resonates strongly with the community, even though its 78B parameter count puts it well above the "around 65B" range the topic asks about.
The Best ๐ค Base Merges and Moerges Model of Around 7B
On the base merges and moerges side, the model that earns the ๐ค tag for the 7B approximate range is tensopolis/virtuoso-lite-tensopolis-v1. With 10B parameters, it was ranked as the Best ๐ค base merges and moerges model of around 7B on the leaderboard today, updated March 8, 2025, and has 20 upvotes. Its nearest peer in the smaller merge space is wanlige/li-14b-v0.4, which holds the Best ๐ค base merges and moerges model of around 13B tag with 15B parameters, updated April 30, 2025, and 1.02k upvotes โ by far the most-voted base merge model in the entire collection. The large gap between 20 and 1,020 upvotes tells me that the li-14b-v0.4 model has achieved critical adoption, even though its "around 13B" categorization puts it above the "around 7B" scope of this topic.
Why do some base merge models gather orders of magnitude more upvotes than others? The li-14b-v0.4 model appears to have struck a balance: a parameter count that's large enough to be competitive on text generation benchmarks, but not so massive that it becomes impractical to run or distribute. Its April 2025 update suggests active maintenance, which likely contributes to the vote surge. The virtuoso-lite model, by contrast, has only 20 upvotes despite being in the target "around 7B" range โ perhaps its lite version feels too diminished to the original, or maybe the merge technique doesn't yet have a proven track record for the community.
What the Tags Actually Mean
The three tag categories on the leaderboard each signal a different production story. The ๐ถ fine-tuned tag means the model started as a base architecture and was then adapted on a targeted dataset โ this is the path taken by sequelbox/Llama2-70B-SunsetBoulevard and MaziyarPanahi/calme-3.2-instruct-78b. The ๐ค base merges and moerges tag indicates the model is the result of merging multiple checkpoints or bases into one, a technique that can preserve strengths from each parent while smoothing over weaknesses. The ๐ข pretrained and ๐ฉ continuously pretrained tags represent models that have either not been fine-tuned or have undergone ongoing pre-training on new data, and they dominate the higher parameter ranges (34B through 111B).
Looking at the Contenders Just Outside the Scope
Several models land just outside the "around 65B" and "around 7B" windows but are worth noting for context. On the fine-tuned side, MaziyarPanahi/calme-3.2-instruct-78b at 78B parameters and dnhkng/RYS-Llama3.1-Large at 82B parameters both carry the ๐ถ tag, showing that the fine-tuned ecosystem has settled into a pattern of very large models. On the base merge side, wanlige/li-14b-v0.4 at 15B, FINGU-AI/RomboUltima-32B at 33B, and CombinHorizon/zetasepic-abliteratedV2-Qwen2.5-32B-Inst-BaseMerge-TIES at 33B all carry the ๐ค tag, demonstrating that the merge technique scales upward โ but the community's voting attention seems to concentrate on the lower-parameter examples where the merge offers the most noticeable practical benefit.
The Vote Distribution Tells a Story
Total upvotes for the collection stand at 700, with individual models ranging from single digits to over a thousand. The fine-tuned model sequelbox/Llama2-70B-SunsetBoulevard has 48 upvotes, the domain-specific fine-tuned calm e-3.2-instruct-78b has 470, and the base merge tensopolis/virtuoso-lite-tensopolis-v1 has 20. Meanwhile, wanlige/li-14b-v0.4 leads all base merges with 1,020 upvotes. This distribution suggests that fine-tuned models in the 70B+ range can still attract substantial community attention, but base merge models only break the thousand-vote barrier when they hit approximately 15B parameters โ large enough to be useful but small enough to be accessible.
Bottom Line
If you're looking for the single best fine-tuned model in the 65B neighborhood on the current Open LLM Leaderboard, sequelbox/Llama2-70B-SunsetBoulevard is the one with the ๐ถ tag, 69B parameters, and 48 upvotes as of the March 13, 2025 update. If you want the top base merges/moerges model in the 7B approximate range, tensopolis/virtuoso-lite-tensopolis-v1 carries the ๐ค tag with 10B parameters and 20 upvotes, though wanlige/li-14b-v0.4 at 15B parameters and 1,020 upvotes is the undeniable community favorite โ it just falls on the "around 13B" side of the categorization. The leaderboard collection makes these comparisons easy by tagging each model with its category and approximate parameter band, but the vote counts reveal which models are actually winning the human attention game.
Source: https://huggingface.co/collections/open-llm-leaderboard/open-llm-leaderboard-best-models (full text extracted via web_extract, paged through via read_source until EOF at char 7676)
