Today, we published our first model. 109M params, trained on a pc.
Meet Squishy πͺΌ , the little model that wished it could. Sentence structure is mostly OK. Narrator gets a little unhinged at times. Edu only! ππ¦
huggingface.co/deltakitsune...
ollama.com/fauxpaslife/...
#AI #BuildInPublic #LocalAi
Posts by delta : kitsune
While Squishy speaks like a toddler, the pipeline works. Built with custom settings and around 19 hours on a local rig.
#deltakitsune #ai #localfirst #localai #squishy
First Kitsune model via the Fine-Tuning Suite is officially "baking."
This is for a 150m model, affectionately named "Squishy," and will operate as a tiny storyteller.
This is running LOCALLY on an RTX 5060ti 16GB card. If GPT 5, Claude... are in the Trillions... Wow. Talk about compute!
π¦ Kitsune is getting a lot more flexible; stabilized the drive migration and hardened the backup system against failures. More importantly, Semantic RAG is no longer hardcoded. You can now route embeddings and rerankers through Ollama, Hugging Face, or local paths directly.
#LocalAI #changelog #d:k
SSD Failed. π¦π₯π
Backed-up. And back up and running. If you do not backup regularly - or have a functioning requirements list, changelog, or have everything hardwired -> Help future you out today.
Now, to sell a kidney for a replacement. π
#deltakitsune #buildinpublic #developerlife #localAI
Kitsune π¦ just took a major step from tracking emotions to forming memories. Iβve built a unified memory crystals layer that synthesizes patterns into a visible brain π§ . V1 reports show these as salience-sized bubbles, mapping out what's remembered. #ai #aihealth #memorycrystals #kitsunecare
Your AI should read the room. Kitsune : Care adapts to your affect β empathetic when struggling, energetic when vibing. V1 live. Memory Crystals next. Built for those the system overlooks. π¦π§
#AI #ChronicIllness #BuildInPublic #HealthTech #LocalAI
Finally, a full pipeline. πͺπ¦
Import > Develop Code/Models > Get Data (Distill/HF/Import) > Prune / Merge (ad-hoc) > Benchmark > Iterate > Deploy > Hugging Face, Ollama, ChatUI > Customize (Persona / Behavior / Settings).
Making it happen one module at a time.
#AI #LocalFirst #CloudReady #Develop
Coming up for air! π Another thing shipped. Momentum is real this weekend. πͺπ¦
User Feedback: Track user feedback for different elements of that chat. This enables the quantification of personality preferences, bug tracking, answer quality, and other factors.
#AI #LocalFirst #CloudReady
V1 of the benchtesting suite has moved to testing. π¦βοΈ
Application: Testing cloud models vs. local specialized experts. (latency, accuracy, which assistants work well together, observations etc., human appraisal for use case)
Then sharing results.
#BuildInPublic #LocalFirst #CloudReady #AI
Who knew that I liked butterflies? π¦π¦
#AI #AIGen #LocalFirst #CloudReady
If you use #Ollama and want a snappy, fast router, consider trying this one.
Arch-Router-1.5B - intelligent #routing between multiple #LLM based on domain and action preferences.
ollama.com/fauxpaslife/...
I imported it into the Ollama system from the #katanemo file.
#AI #agentic #specialists
Oh, there I am π¦... sleeping, again, on a cool winter's day.
(Testing pipelines and local tooling.)
#AI #AIGenerated #RTX5060 #Nvidia #BuildInPublic #LocalFirst #CloudReady #kitsune
I started using delta:kitsune to scan the #news for me and give me a digestible PDF. You are more than welcome to review it if you would like. Feedback is always welcome.
@deltakitsune.bsky.social #AI #Summary #RSS
github.com/fauxpaslife/...
I got the order in pics wrong, meh. I blame the human in this equation. π π
AI feels good because it's programmed to.
Same model, flipped prompts:
"incredibly draining!" vs "pathetic waste of space."
Neither is truth. Both follow instructions. That validation? Engineered empathy.
For chronic illness AI: meet people where they actually are. π¦
#AI #BuildInPublic #HealthTech
Stop paying for AI's mistakes.
Built conversational interruption: hit ESC mid-response β AI stops, pivots, no restart.
NVIDIA just shipped this as "cutting edge."
Built for brain fog days when you can't craft perfect prompts.
π₯ youtu.be/20XxMnjZygg
#LocalAI #HealthTech #AI #CloudReady
I had to throttle tokens quite a bit on early attempts of models on the Pi. Starting was 30. So unless I asked a math question - it was not really useful. Also on my other rig I tried some lower parameter and heavily quantized models.
Thank you! Really appreciate it.
First goal was PoC; and would the model that I could actually put on there be smarter than a potato. I have a small Qwen and Gemma I should test as well. Let me run more tests this week, and I will get back to you on this.
V2: Hybrid architecture. LFM2.5 reasons in text, Piper speaks. Faster, more reliable. Still 100% local on a Pi5. No cloud required.
Repo in comments.
www.youtube.com/shorts/4fMue...
#BuildInPublic #LocalAI #OpenSource #RaspberryPi #EdgeComputing #LiquidAI #AIAccessibility #LocalFirst
V1: Pure audio-to-audio using LiquidAI's LFM2-Audio. Single model handles everything. Pi5, no cloud, $150 hardware. Experimental but it works. Repo in comments.
#BuildInPublic #LocalAI #OpenSource #RaspberryPi #EdgeComputing #LiquidAI #AIAccessibility #LocalFirst
www.youtube.com/shorts/kNyeg...
2026 is going to be a good year.
I cannot fit it all from LinkedIn into BlueSky, but it really does encapsulate the delta : kitsune's update perfectly.
#ai #buildinpublic #localfirst #cloudready #healthtech #opensource #privacy
There we have it. Still do not know. ππ¦ #ai #silly #buildinpublic
It has been quiet. Still building. And playing π¦. Made a snake game and lights out using kitsune : code. Compile and run in app. Some version someday will have the model manage that and iterate X cycles/time. I think GPT is using "tooling time" allotments. #buildinpublic #ai #code
This is what a lot of coding/iteration looks like. Determined to make delta : kitsune all that it can be. Here is to 2026. (Year of the π¦?)
#openai #buildinpublic #localfirst #cloudready
Cloud integration for OpenAI, Anthropic, and Ollama models is now integrated. Still early days, but with some Kitsune pizzaz. #AI #BuildInPublic #LocalFirst #CloudReady
This proves the goal is worth chasing. Our roadmap validated through tech.