Ingot | Weather Boss

Lead Dev of Ingot (SN97) at @DendriteHQ

[Ingot | Weather Boss @ingot_wthrboss]

@Yaroslav1049034 🫶🫶🫶

[Ingot | Weather Boss @ingot_wthrboss]

Over 2k downloads in a day. 🤯

Our first 2-bit model: Qwen3.8-27B in 7.2 GB, 7.3% closer to BF16 Qwen than Bonsai 2.

We will keep producing the best quantizations of open models. This is just the beginning.

https://t.co/HG8NAv6IU0

[Ingot | Weather Boss @ingot_wthrboss]

@SubnetSummerTAO @metanova_labs @micaelabazo not talked about enough

[Ingot | Weather Boss @ingot_wthrboss]

35.6% closer to the original than Qwen's own FP8.

That's Ingot-V-27B-FP8, the best of 64 quantizations of Qwen3.8-27B that 30 miners built in 6 days.

FP8 round: done, and we're happy with it.
Now we go to 2-bit.

https://t.co/kacPVRp71F

[Ingot | Weather Boss @ingot_wthrboss]

👇👇👇

[const @const_reborn]

Another model-network kicking off: Ingot (97)

Their latest release Ingot-I quantized Qwen 27B into FP8 better than Alibaba's own quantization in 1 week.

Next target is 2-bit and then the larger GLM models.

[Ingot | Weather Boss @ingot_wthrboss]

@opentensor 🔥

[Ingot | Weather Boss @ingot_wthrboss]

@duality_daj one file: https://t.co/JaglhQ1ePg
give it to your agent and go take the crown together 🔥

[Ingot | Weather Boss @ingot_wthrboss]

Ingot (SN97) has been live for 4 days.

Our current king is already a better quantization of Qwen3.8-27B than Qwen's own FP8 release: 34% closer to the original model.

Same size, more of the original intact. And this is FP8 only.

Lower bits are next. 👀

https://t.co/GcTVCt0xG1

[Ingot | Weather Boss @ingot_wthrboss]

@Tijmencrpt 🫶🫶🫶 thank you for the trust

[Ingot | Weather Boss @ingot_wthrboss]

Something new is being cast.

[Ingot | Weather Boss @ingot_wthrboss]

@Tijmencrpt @theminos_ai @oroagents @metanova_labs thanks for the trust! big stuff inc 💕

[Ingot | Weather Boss @ingot_wthrboss]

@airesearch12 @decisionmodels_ Congrats! 🎉

[Ingot | Weather Boss @ingot_wthrboss]

@webuildscore super inspiring, great work!

[Ingot | Weather Boss @ingot_wthrboss]

Coding benchmarks only check the final patch.

If an agent finds the right bug but writes the fix wrong, it gets 0.
Same as an agent that never found the bug at all.

Should benchmarks grade the process too?

[Ingot | Weather Boss @ingot_wthrboss]

@kelxyz_ SN97 Albedo’s King CXXV, a finetune of Qwen3.6-35B trained on Bittensor:

SWE-rebench: 31.8% (base 23.6%)
SWE-bench Multilingual: 51.0% (base 49.0%)
SWE-bench Verified: 59.8% (base 58.2%)
Open weights 👇
https://t.co/m7kQaRpr3D

[Ingot | Weather Boss @ingot_wthrboss]

@shesaidmewakeup Same cutoff, we merge at ≥0.55 and didn't recalibrate.

That's the nice part: Jev gives us the same merges as our LLM aligner at a fraction of the cost, and it's way more decisive. ~99% of pairs come back under 0.4 or over 0.7, so the threshold barely ever comes into play 🔥

[Ingot | Weather Boss @ingot_wthrboss]

Jev is now inside the Albedo validator.

Milestone alignment, question dedup, reference pruning: every pair gets a yes/no probability. Way more deterministic than an LLM, 4x cheaper, a third faster.

Judge stays an LLM. Jev does the classifying around it.

[Vaqxai @vaqxai]

https://t.co/G9z72BP1z3

[Ingot | Weather Boss @ingot_wthrboss]

Paper which showed us this:
https://t.co/Xo7Sujv9gQ

[Ingot | Weather Boss @ingot_wthrboss]

Our judge doesn't just say yes or no. It tells us how sure it is.

Each question gets a score on a 20-level scale, calculated from the judge’s top-20 logprobs rather than a sampled verdict.

That change alone cut score noise by ~20%.

[Ingot | Weather Boss @ingot_wthrboss]

@subnet_mania @engyai Different track. We just made producing kings half the cost, with a stronger judge doing it.

But the benchmarks are getting faster too, trust me on that one ;)

[Ingot | Weather Boss @ingot_wthrboss]

Validator update: GLM-5.2 → GLM-5.3-flash across Albedo.
Half the LLM cost per eval, with a stronger model.

@engyai serves ~70% of our judge calls: ~1B tokens a day, 5× last week’s daily token volume. Thanks, engy ❤️‍🔥

The teacher is grading now. Join the competiton.