SillyTavern weekly news for the week of June 16 to June 22, 2026. The model story this week is still GLM 5.2 *yawn* (kidding, I’m just salty from z.AI’s shift). Putting that to the side, the real story is the user-side tooling cluster: DeepLore v2.5, Natural Extended v0.9, Leonardo, Zebede, and a quiet but real character-card workflow shift. Megumin V8. GitHub quiet. Also, what model is Owl Alpha on OR?.. That’s not part of the news. That’s my personal question. Stealth models drive me nuts.
A good week for tools
Last week the SillyTavern story was a model release (GLM 5.2 finally landing on OpenRouter, DeepSeek v4 punching above its weight, Fable 5 starting a culture war). This week the model story was a sequel: GLM 5.2 is now the consensus pick, 5.1 is the cautionary tale, the 5.1 vs 5.2 thread settled into “yeah, 5.2 won,” and the is the Z.ai subscription worth it thread is the one to read if you are still deciding whether to pay. I didn’t, but I’m a brokie. I’m sitting with Gemma 4 31b right now, but I’ll write a separate thing for that. Mileage may vary, as always, but the curve has bent.
The actual story of the week, though, is everything around the model. The community tooling layer got dense in a way that suggests ST is hitting a maturity phase where the official client ships less, and the people who build on top of it ship more.
the community tooling cluster
If you only install one thing for ST this week, make it DeepLore v2.5. The release post is the rare kind of thread that ships a feature list, a use case, and a screenshot all in the first post. Keyword plus AI lore retrieval from your Obsidian vault, seven languages, a graph health view, editable prompts. If you write lore-heavy RPs in Obsidian, this is the one.
Natural Extended v0.9 is the group-chat response control extension and it is, frankly, overdue. Group chat is the most under-served part of SillyTavern, and any extension that takes it seriously is a good thing indeed.
Leonardo is the creator card that makes its own preset, with a randomizer for scenario generation. That is shamelessly ambitious and I respect it. Cards-that-make-presets is a category I did not have on my bingo card for 2026, and we are getting two of them this month.
Pura’s Director Preset 14.0 went CoT-less and picked up RPG Elements. If you are already running a director-style preset, the changelog is worth a read. The 14.0 version number is a flex and a half for a community preset.
The character-card workflow has its own thread this week. Zebede’s Roleplay Tool is the renamed character generator, and the rename is the story: tools that survive long enough to get renamed are tools that earned it. A new JanitorAI to SillyTavern Character Card Generator Chrome extension shows up to make the cross-platform card shuffle easier. And a post about making a character card by hand for the first time reads as the community quietly telling itself, and anyone listening, that the AI-generated card era is over and the human-curated one is back (facts).
megumin v8
Last week Megumin Suite V8 showed up in the “new tools” section. I wrote about it last week. I’ve been messing with it. I love it. I think it’s pretty high-quality. I’ve been using it with Gemma. On Obsidian. Sue me.
The MiMo gag from a few days later, Finally I understand what ‘breath hitching’ actually is. Thanks MiMo., reads as related, in the sense that the community is in a phase where the new tools are making people laugh at their own old writing.
github: all quiet on the northern front
The official SillyTavern repo is still on a quiet stretch, with the active PR stream focused on plumbing: a 700-tokens option for tokenizers, a refactor of the extension management and assets download menu, a malformed-JSON option in extractJsonData, the usual npm audit cleanup, and rate limit work on the basic auth middleware. None of that is going to trend on Reddit. All of it is the kind of work that prevents a release from falling over six months later.
The 700-token preset option pairs neatly with what Megumin V8 is exposing on top of it, which suggests the two are converging on the same token-budget conventions. I am the kind of person who reads these PRs and gets excited about a rate limit on auth middleware, so make of that what you will.
housekeeping
A continuity checker PSA argued that a 2-4B model running alongside your main RP model is basically free and catches plot holes better than the big one. I am filing this under “probably correct, definitely going to try.” A NIM cracking down on OpenClaw thread is worth a bookmark if you route any of your traffic through NVIDIA’s hosted endpoints. That said, the title is more dramatic than the contents. I read it. You read it. Like a baby bird eating food from its momma (gross!)
Consume. Create. Obsess.
More tools, guides, and rabbit holes at rpfiend.com.