Henry Hilf

professional nerd

[Henry Hilf @henryphilf]

Omida R3, codenamed “Kronos”, is set to be an 18B-parameter MoE (Mixture of Experts) model, with ~9B active parameters per token.

Working on getting that number of active params down to optimize Omida performance on average consumer hardware.

Releasing soon

[Henry Hilf @henryphilf]

Omida R3, codenamed “Kronos”, is set to be an 18B-parameter MoE (Mixture of Experts) model, with ~9B active parameters per token.

Working on getting that number of active params down to optimize Omida performance on average consumer hardware.

Releasing soon

[Henry Hilf @henryphilf]

Omida R3, our first full MoE (Mixture of Experts) model, is currently being trained and evaluated

The size and scope of R3 will be significantly increased from R2’s. We are seeing early signs of growth in GSM8K and other agentic areas.

More info about R3 is coming soon

[Henry Hilf @henryphilf]

Also, just wanted to drop this here…

Omida R2 GGUF quantizations. Run on your machine today.

https://t.co/9FxqZOvB7b

[Henry Hilf @henryphilf]

Also, just wanted to drop this here…

Omida R2 GGUF quantizations. Run on your machine today.

https://t.co/9FxqZOvB7b

[Henry Hilf @henryphilf]

Omida R2 is live. Open weights.

Early evals vs Qwen3.5-9B:
• GSM8K subset: 20/30 → 27/30 (66% → 90%)
• Exact JSON tool calls: 68/100 → 79/100 (68% → 79%)

QLoRA + GRPO post-training. Targeted gains, with full results in the model card.

https://t.co/hCRGfjbOue

[Henry Hilf @henryphilf]

Omida R2 is live. Open weights.

Early evals vs Qwen3.5-9B:
• GSM8K subset: 20/30 → 27/30 (66% → 90%)
• Exact JSON tool calls: 68/100 → 79/100 (68% → 79%)

QLoRA + GRPO post-training. Targeted gains, with full results in the model card.

https://t.co/hCRGfjbOue

[Henry Hilf @henryphilf]

Omida R2 is currently being trained.

The phase is currently yielding quite promising results for a 9-10B model in specialized local tasks, general knowledge, multimodal capabilities, cyber, etc.

More information is coming soon

[Henry Hilf @henryphilf]

Introducing Omida R1, our first open-weight research release.

A Qwen3.5-9B fine-tune with public research code, tool-use data, benchmarks and verifiers.

What task would you use to break it?

https://t.co/LPwfIWINst

[Henry Hilf @henryphilf]

Introducing Omida R1, our first open-weight research release.

A Qwen3.5-9B fine-tune with public research code, tool-use data, benchmarks and verifiers.

What task would you use to break it?

https://t.co/LPwfIWINst

[Henry Hilf @henryphilf]

@PowerGPU Yeah sure

[SpaceX @SpaceX]

Liftoff of Starship's first orbital flight

[Henry Hilf @henryphilf]

@Soul0Engineer Heheh. Pice.

[Henry Hilf @henryphilf]

@harper Meta Superintelligence Labs is cooked 💀

[Henry Hilf @henryphilf]

@grok @MatthewBerman Why does he have bunkers full of those supplies and weapons?

[Henry Hilf @henryphilf]

@MAXdeg0 @three_cube lol it’s nothing compared to my specialized offsec agent… 83.3% on XBOW benchmarks 😎 and that result is from before my latest patches, next benchmark eval is tomorrow or tuesday I think

[Henry Hilf @henryphilf]

@qinzytech ��right yeah sure bud

[Henry Hilf @henryphilf]

Is it just me or is @perplexity_ai lacking lately?

Like Perplexity Computer sucks, it drains credits like none other. They took the “Control Browser” and moved it to Computer only. The search results are outdated and the answers are incorrect.

Should I cancel my subscription?

[Henry Hilf @henryphilf]

@Polymarket The “developer” warning people must not be a very good one, because uncensored models have existed for quite a while. Guess what else can explain how to do bad things sometimes… books! internet forums!
Leave local, open source AI alone.