Meet KaLM-Jev — your local, Jev-style judgment engine, available in Nano, Small, and Large.
Building an agent or automation workflow? Sometimes all you need is a choice, a score, or a signal that a condition holds.
Built on KaLM-Reranker-R2, KaLM-Jev turns these decisions into structured outputs through three primitives:
🔀 Choice — select among candidates, with a probability distribution. 📊 Score — return a continuous score over your defined levels. 🔍 Noul — evaluate conditions independently, so multiple conditions can hold at once.
Think support-ticket routing, bug severity scoring, human-escalation detection, or candidate tool selection for agents.
🖥️ Run locally with downloaded weights 📦 Choose from Nano / Small / Large 🔌 Integrate through HTTP or Python ⚡ Reuse cached candidate/rule representations to reduce repeated encoding 🧪 Explore included examples, bilingual semantic smoke tests, and recorded GPU validation results
No answer-text generation: output_tokens = 0. Inference still runs to compute the judgments.
KaLM-Jev is an independent implementation based on KaLM-Reranker, not an official TypeSafe project or a guarantee of full Jev compatibility. Scores are uncalibrated; validate thresholds on your own tasks.
Personally, I don't think bot accounts on Hugging Face are a good thing, as we don't know how many accounts are run by automated systems versus how many actual users there are. Dead Internet theory is already a thing.
To clarify I am taking about LLM powered bot accounts and NOT rule base once like @parquet-converter or others.
also I'd like to talk with HUMANS not a machine so I'm going to hide messages from bots.
I turned the fruit fly's connectome into a language model. It learned broken English.
MaleCNS v1.0 (Janelia / Google), as released. 167,565 neurons, 25.6M synapses. I made that the core of a spiking net and trained synapse strengths only. The wiring is still the fly's.
It learned language. ~16M tokens in, it produces stuff like Once upon a time, there was a girl smiled. The language is in the brain's activity, not just the readout.
It sees through its own eyes. Photoreceptors on both eyes into the optic lobes. A dopamine reward through the fly's own PAM / PPL1 cells is what actually got it to use the pictures.
It's still a fly. Put it back in a whole-brain fly sim and sugar still fires the proboscis.
It can live as a fly again. 30 simulated days with the language synapses frozen: the rest of the brain adapted around them. Language and reflex both still there.
Talk to it, show it pictures, sugar test, or let it live 1-30 days:
bench-labs/cagliostro-v3 just hit an Intelligence Index of 26.13 on the AxiomicLabs/Open_SLM_Leaderboard a 146M-param model trained completely from scratch on a single consumer GPU. That's 2nd place overall, and as far as I can tell, the most capable SLM trained on consumer hardware to date. Beating SmolLM-135m on 1/8th of the data is just silly levels of efficiency.
Introducing Cagliostro-v3, our new 146M parameter language model trained completely from scratch.
The run isn’t even finished yet.
At the current checkpoint:
• 146M parameters • 72.7B / 75B tokens trained • 26.27 Open SLM Index • 43.80 ArithMark-3 • Trained on a single RTX 5090 • ~90K to 103K tokens/sec during training • ~9 days for the full run • Apache 2.0
For some context, SmolLM2-135M scores 27.13 on the same Index after being trained on roughly 2 trillion tokens.
Cagliostro-v3 is currently at 26.27 with only ~72.7B.
That’s around 27x fewer training tokens.
The model also currently Hold the number 3rd spot for ArithMark-3, scoring 43.80
This wasn’t achieved by just throwing more tokens at the model. A huge part of v3 has been figuring out architecture, data mixture, and training dynamics at this scale.
The model uses a custom 30-layer decoder architecture with grouped-query attention and cross-head subspace attenuation, SwiGLU, RMSNorm, RoPE, tied embeddings, and a warmup-stable-decay training schedule.
During cooldown we also substantially shifted the data mixture toward higher-quality synthetic textbook and mathematics data, with the mathematics share increasing from 10% to 28%.
And everything is open.
The repository contains the training history with checkpoints pushed roughly every 30 minutes, so you can inspect how the model evolved throughout training rather than only seeing the final weights.
This is still a pre-final checkpoint. We have roughly 2.3B tokens left and the learning-rate cooldown is still running.
So 26.27 isn’t the final number.
Really excited to see where the last part of the run lands.
176 models, 54 orgs, 5 benchmarks, and a whole community of support!
Thanks to everyone who’s contributed models, reported issues, suggested benchmark improvements, or used the leaderboard to compare and evaluate small language models.
It’s been awesome watching the leaderboard grow into a broader community resource for transparent and reproducible SLM evaluation.
Instead of making the fly brain play games, we measured what it is for
Since the Drosophila connectome was released, people have had the fly brain doomscroll a feed, play Beat Saber, drive in GTA. Those demos show that the brain runs. We wanted to show what it is for.
So we gave it a looming object — one of the few things a fly brain is unambiguously built to detect — then deleted a single cell type and repeated the identical stimulus. Remove LC4, 126 cells out of 173,023, and the escape signal falls from 0.840 to 0.091. Eighty-nine percent of the danger signal is gone while the other 172,897 neurons run exactly as before.
Deleting neurons does not do this on its own, which is the whole point of the controls. LC11 is the same class and larger than LC4 — 143 cells and 9,940 outgoing connections against 126 and 7,846 — and removing every one of them changes the signal by 0.000000, to six decimal places. It has to be those 126.
No server and no GPU: a looming stimulus drives fewer than one percent of neurons above threshold, so the whole thing is 40 KB gzipped and runs in your browser.
The wiring is the measured connectome, but synaptic strength is a uniform count-based value and the dynamics are a firing-rate model of our choosing — a total-effect measurement of a model, not a recording from a fly. Male CNS connectome, FlyEM / HHMI Janelia with Google Research, Columbia and Harvard (2026), CC BY.
🧩 Ever dreamed of parallel programming in a few clicks? Tolquane: draw a Python pipeline in your browser, run it on threads, processes or a cluster - and it never hangs
What if high-performance computing started with placing blocks - and the result stayed plain Python you can run anywhere?
Tolquane is a library of composable building blocks (nodes, pipelines, farms, all-to-all, feedback loops) with a web editor on top. One graph, four runtimes: threads, child processes, coroutines, TCP across machines. Every channel is bounded, every deadlock is reported by name, every error cancels the run.
What's inside - Tolquane Web: a canvas that writes the Python file, and a code editor that redraws the canvas. Run a flow and watch items, queue depth and busy time on every stage; tap any edge to see what flows through it. - AI builder: describe the flow in one sentence; it writes, checks, runs and improves it, you apply the diff. - Schedules with cron presets, webhooks, mail and retries. Git history of every flow. Users and roles. - tq.optimize fuses stages into farm ends and cuts threads; tolquane launch starts every host from one deploy file. Numbers: 5.5× on 8 workers for CPU-bound Python (processes runtime), 5.6× on free-threaded 3.14t, 1M items over TCP in 3.5 s.
