Inductive ML

Independent research lab

[Inductive ML @inductive_ml]

6/ Reports, pre-registrations and analysis scripts are public: https://t.co/QRYcj2bghE. We run independent audits of RL environments before they're sold to labs: https://t.co/y4GBUVbY6T

[Inductive ML @inductive_ml]

1/ In 85% of Xiaomi's MiMo RL coding tasks, ls -t points straight at the files the agent needs to change. We audited their environments. Here's what we found, and why it's not the scandal it sounds like 🧵 https://t.co/WUWaTnm4fJ

[Inductive ML @inductive_ml]

5/ The fix is one line at image build: find <repo> -xdev -exec touch -h -d @0 {} +. We sent it to @XiaomiMiMo today.

[Inductive ML @inductive_ml]

me and opus 5.5 on medium effort against the world

[Inductive ML @inductive_ml]

Fable limits on the max plan has gone to shits, 20% used through just chatting on the web?? @ClaudeDevs

[Inductive ML @inductive_ml]

Why can't I use my Codex reset if I have 84% limit remaining but my reset expiring tonight? @thsottiaux ?
This used to work before wtf

[Inductive ML @inductive_ml]

Bonsai 2 gets the same treatment 🌳
27B + vision on MLX, with 19.5% less loaded model memory: 8.60 → 6.92 GB.
Native 1.75-bit weights, losslessly packed. Original vision preserved. All tested outputs match the official checkpoint.
https://t.co/oK1mXFnkxP

[PrismML @PrismML]

Today, we’re announcing Ternary Bonsai 2 27B.

Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance.

Two months after the first Bonsai 27B release, the biggest change is quality. The footprint remains 5.9 GB, but the gap to full precision has narrowed materially, with particularly strong gains in agentic coding, multimodal reasoning, and long-horizon tool use.

Ternary Bonsai 2 27B is available today under Apache 2.0.

[Inductive ML @inductive_ml]

@maximelabonne @huggingface check this out as well, 230M at 1400 tps on the browser

https://t.co/7oUIBs9Q7B

[Inductive ML @inductive_ml]

what tps do you get? (all local btw)
https://t.co/tLRvtQvk8X

[Inductive ML @inductive_ml]

what tps do you get? (all local btw)
https://t.co/tLRvtQvk8X

[Inductive ML @inductive_ml]

@liquidai LFM-230M in the browser at 1400 tok/s , possibilities are endless, try it out yourself

https://t.co/DOl3hxMpij

[Inductive ML @inductive_ml]

@liquidai LFM-230M in the browser at 1400 tok/s , possibilities are endless, try it out yourself

https://t.co/DOl3hxMpij

[Inductive ML @inductive_ml]

Bonsai 27B now runs in just 5.9 GB of RAM 🤯

22% less memory, 100% lossless token-for-token identical output to the official checkpoint, verified across all 26.9B weights.

Runs on any 16GB Apple Silicon Mac:

pip install mlx-lm
mlx_lm.generate --model inductiveML/Ternary-Bonsai-27B-mlx-lossless-1.75bpw

check out here

https://t.co/Ng2CRowFsC

[Inductive ML @inductive_ml]

@PrismML

[Inductive ML @inductive_ml]

Bonsai 27B now runs in just 5.9 GB of RAM 🤯

22% less memory, 100% lossless token-for-token identical output to the official checkpoint, verified across all 26.9B weights.

Runs on any 16GB Apple Silicon Mac:

pip install mlx-lm
mlx_lm.generate --model inductiveML/Ternary-Bonsai-27B-mlx-lossless-1.75bpw

check out here

https://t.co/Ng2CRowFsC

[Inductive ML @inductive_ml]

claude down ahhhhh, was having a pretty deep convo with fable

[Inductive ML @inductive_ml]

ever heard of batch inference?

[toly 🇺🇸 @toly]

My estimate is that it would take a $100k setup to get kimi to run a home at 70-100 tokens per second. It’s nuts that anthropic and OpenAI are giving it away for $200 a month.

[Inductive ML @inductive_ml]

chatgpt on web is the most laggy experience ever how have they not fixed it yet

[Inductive ML @inductive_ml]

got a @thsottiaux reset LFGGG