INGOT

Frontier model in your pocket.

Powered by Bittensor SN97.

[INGOT @IngotSubnet]

@clgssss Agreed, that's exactly why we called it promising, not proven.

Both gains are above the run-to-run noise, but on one run each. More runs or a bigger jump will settle it.

The 25% number is the solid one: both models measured side by side on the same 236 held-out steps

[INGOT @IngotSubnet]

Two weeks ago, we started Ingot's competition on Qwen3.8 27B at 1.75 bits per weight. The file is 5.95 GB instead of 54.7 GB, with a gap to the original we're working to close.

The baseline was PrismML's Ternary Bonsai 2. Since then, 27 kings have been crowned from 270 submissions. Each successive king had to bring its predictions measurably closer to the original than the preceding king's.

King XXVII's next-token predictions are 25% closer to the full-size original than Bonsai 2's on 236 held-out steps from real coding and terminal work.

SWE-bench Multilingual and Terminal-Bench both moved up for the first time this round. Those gains are right at the edge of run-to-run noise. We need more runs to see whether they hold up, so for now we call them promising, not proven.

Ingot-XXVII is still one 5.95 GB file, and the weights are open.

Download the weights: https://t.co/ISObSB2b9J

The competition continues. We'll keep testing and return with measured results and the evaluation details behind them.

[INGOT @IngotSubnet]

@peijiancc It is! All credit to the miners. 🫶

[INGOT @IngotSubnet]

@peijiancc In KL to the original:
Bonsai 2 started at 0.110,
King XXVII is at 0.083,
around 0.03 a model starts to match the original,
the original is 0.

Still a way to go, and we're going for it.

[Ingot | Weather Boss @ingot_wthrboss]

The numbers, for anyone who wants them.

SWE-bench Multilingual 58.0 → 63.7, run-to-run noise ±2.9
Terminal-Bench 2.1 50.6 → 58.4, noise ±4.8

Both clear their noise bands, but they're single runs. Settling it takes either repeating them or a gain big enough that one run decides it.

KL is what I trust:
236 held-out agent steps, same teacher answers, paired.
0.1101 → 0.0828, 20x the measurement error.

Really proud of this one! ❤️‍🔥

[Ingot | Weather Boss @ingot_wthrboss]

Love this kind of test: a hard, niche question you can check yourself beats any leaderboard.

Ingot was the only one to get the core idea right: every message goes to every device.

Sharper details and more speed on Mac are next. 🔥

[INGOT @IngotSubnet]

How do you improve a model that's already as good as the original?

That was the question behind Ingot's first round on Bittensor SN97: quantize Qwen3.8 27B to FP8 and stay as close as possible to the full BF16 model. Qwen's own FP8 release already scores like the original. There is no measurable intelligence left to win back.

What is left is fidelity: how closely a quantized model tracks the original, token by token and layer by layer.

Even Qwen's official FP8 drifts. On the BF16 model's own transcripts from six benchmarks, it picks a different next token than the original 2.2% of the time, and its hidden states sit 6.4% away from the original's. That sounds small, but over a 4,000-token answer it adds up to about 90 places where the FP8 model would choose differently, and any one of them can send the rest of the answer somewhere else.

The round's final king, Ingot-V-27B-FP8, closes part of that gap. Against Qwen's FP8, it has:
- 13% lower KL divergence from the BF16 model's next-token predictions
- 9.5% lower distance from its hidden states

Five kings were crowned in four days to get there. It keeps Qwen's FP8 format and size, 30.9 GB, so it runs with the same serving commands.

FP8 was the warm-up. Round 2 is ternary: 1.75 bits per weight, 5.95 GB, where the gap to the original is wide.

The weights are open. Swap it in for Qwen/Qwen3.8-27B-FP8. We think you'll keep it:
https://t.co/xwHXYGITkz

[Ingot | Weather Boss @ingot_wthrboss]

You read about 5 tokens a second.
A 27B model on the base M5 MacBook now writes 25.

Ternary weights, our llama.cpp fork and a DFlash drafter: 2x the usual 2-bit setup, fully offline, in 13 GB of RAM.

Try it on your Mac 👇
https://t.co/qJ7VSVaeF3

[INGOT @IngotSubnet]

Today the smartest AI models live in giant data centers, and you rent them.

We want you to own them, on your phone.

That's how we start.

[Vaqxai @vaqxai]

https://t.co/WsilGtJ5by

[INGOT @IngotSubnet]

Looks like a lot of people don't want to rent AI anymore...

[Ingot | Weather Boss @ingot_wthrboss]

200,000+ downloads.

11 days. 22 kings. All on Hugging Face.

27B parameters. 6 GB. Runs on your laptop.

[Ingot | Weather Boss @ingot_wthrboss]

Testing Ingot models locally. You can too 👇

We're tuning our fork of PrismML's llama.cpp so Ingot runs as fast as possible everywhere: NVIDIA GPUs, new Apple chips and older ones.

Here it is running on CUDA: ~250 tokens per second 👀

[Dendrite @DendriteHQ]

We are proud to announce that the first issue of Dendrite Magazine is out.

You can find almost everything that happened across Dendrite and our subnets last month below.

Have a great read! https://t.co/HNaXhMwOus

[INGOT @IngotSubnet]

Small is beautiful. And useful.

A frontier model on your own laptop. That's the ambition behind Ingot.

Ingot turns model compression into an open competition on Bittensor SN97. Miners develop quantized versions of powerful open models, competing to reproduce the original more faithfully within a fixed size limit. Winning models are published for commercial use.

The aim is to bring more of what powerful AI can do onto hardware people can own. Getting there means making models smaller while preserving the capabilities that make them useful.

That trade-off is at the heart of quantization, and its story begins long before AI.

Quantization is the art of rounding the continuous world to a small set of discrete values. It always loses something, but a well-chosen loss made digital media possible and is now moving large AI models onto laptops and phones.

The foundations of quantization lie in Claude Shannon's information theory. In 1948, he showed that any message can be measured in bits. His source-coding theorem sets a floor (entropy) on the average code length needed to describe a source without loss. He also gave us rate–distortion theory: if some error is acceptable, you can go below that floor. Every quantizer since is a bet placed on that trade-off curve.

The digital revolution was built on that bet: JPEG, MP3 and modern video codecs all cashed in on it. They transform a signal, then quantize coarsely wherever our eyes and ears are least sensitive: fine colour detail, sounds masked by louder ones, motion we barely track. A song shrinks roughly tenfold and still sounds the same. Discarding imperceptible precision is the quiet engine behind every photo, playlist and stream.

Compressing AI is the modern frontier. A language model is billions of numbers, usually stored as 16-bit floats. Post-training quantization compresses them down to 8, 4, or even 2 bits. That small is beautiful - and useful - is a lesson computing has learned before. It once lived in mainframes behind glass; the personal computer put it on every desk. Large language models are retracing that path, from server clusters to laptops, phones and edge chips. Quantization is what makes the move possible, bringing privacy, offline use and lower cost.