Stellan Haglund

Entrepreneur and code ninja.

[Stellan Haglund @Stellanhaglund]

@Alibaba_Qwen 27b soon?

[Stellan Haglund @Stellanhaglund]

@LottoLabs From my small tests in the playground I'm not convinced.
Have you tried it?

[Stellan Haglund @Stellanhaglund]

Where is qwen 3.7 27?

[Stellan Haglund @Stellanhaglund]

@nadeem_jaleel @liquidai @huggingface 😂

[Stellan Haglund @Stellanhaglund]

Impressive benchmarks but I must be doing something wrong in the playground as usual - it’s not doing what I want it to

[Liquid AI @liquidai]

Today, we're releasing LFM2.5-8B-A1B, a device-optimized model designed to power real-life applications on phones, laptops, PCs, robots, and fast & lightweight server-side use-cases.

> 8B MoE, 1.5B active
> Expanded 128K context
> LFM2.5 flagship hybrid MoE architecture
> Trained on 38T tokens + large-scale RL
> fast, reliable tool calling, punching above its weight, comparable to models with up to 4x its size
> customizable on a single GPU for any specialized task
> LFM2 open-weight license

🧵

[Stellan Haglund @Stellanhaglund]

@liquidai Asked a question in Swedish got a French response

[Stellan Haglund @Stellanhaglund]

@liquidai Just English?

[Stellan Haglund @Stellanhaglund]

@GoogleAIStudio @antigravity Can I do other apps than react there now?

[Stellan Haglund @Stellanhaglund]

@ibuildthecloud @leetcoderbr But under the hood it’s swapping the content the same way I guess?

[Stellan Haglund @Stellanhaglund]

@ibuildthecloud https://t.co/Bo1nIPftt7

[Stellan Haglund @Stellanhaglund]

@ibuildthecloud Yeah it’s not easy, sveltekit has a really simple and good state management with reactivity.

Have you looked at electricsql?
Basically crdt with a local browser db connected to the real db.
So offline support with managed sync and everything in place

[Stellan Haglund @Stellanhaglund]

@ibuildthecloud I see, I was just thinking about that, multiple routes and a nested layouts is solvable I guess but I’m not sure how that would be structured.
What are your state synchronization issues and what have you tried?

[Stellan Haglund @Stellanhaglund]

@mr_r0b0t @magiknono From what I’ve tried it’s very similar to tailscale.
The big upside I see is that it’s fully self hostable with everything needed to manage devices and users.
Wireguard is more one connection at a time.
While in tailscale/netbird you can manage tons of devices in the same mesh

[Stellan Haglund @Stellanhaglund]

@chrisdrit Notice any difference in acceptance rate on fp8?

[Stellan Haglund @Stellanhaglund]

@tmaiaroto I see an enormous gain locally with vllm but not on everything.
For simple tasks like structure this into json I can go very high in mtp.
For more complex stuff it drops fast.
I guess the drafter could be fine tuned to be better on specific stuff too

[Stellan Haglund @Stellanhaglund]

@atomic_chat_hq Do you notice any difference in acceptance rate on the quantized, the drafter is trained for the fp16 after all

[Stellan Haglund @Stellanhaglund]

@tmaiaroto @atomic_chat_hq Very big difference depending on the type of prompt I notice.

[Stellan Haglund @Stellanhaglund]

@mr_r0b0t @magiknono Tailscale uses wireguard for the peer to peer connection you establish within your mesh.
Netbird basically has the full tailscale setup open sourced

[Stellan Haglund @Stellanhaglund]

@googleaidevs How would I fine tune the drafter?