jefflau.eth

yaklaşık 2 saat önce

@MiaAI_lab Can’t wait for this one!

0

138

yaklaşık 3 saat önce

Mia cooking 👩‍🍳

yaklaşık 4 saat önce

DeepSeek-V4.1-Flash EXL3-EAQ v1.1 is out 🧪 What's different vs v1? Sparse attention uses an indexer to pick which past tokens matter. v1.1 keeps that whole path in full BF16 instead of quantizing it (the indexer, shared query projection, and compressor). When fed the original model's inputs, v1.1 makes the exact same picks 100% of the time past 16k (403,560 / 403,560), vs 94.95% on v1. In a full quantized run, both land near 73%. That's because the inputs reaching the indexer have already shifted from quantizing the rest of the model, so a BF16 indexer alone can't fix that gap. v1.1 weights just ~0.15 GB more than v1. Please note that this is experimental ‼️ Get it here: https://t.co/yMmNvpb4gj

MiaAI_lab's tweet photo. DeepSeek-V4.1-Flash EXL3-EAQ v1.1 is out 🧪

What's different vs v1?

Sparse attention uses an indexer to pick which past tokens matter. v1.1 keeps that whole path in full BF16 instead of quantizing it (the indexer, shared query projection, and compressor).

When fed the original model's inputs, v1.1 makes the exact same picks 100% of the time past 16k (403,560 / 403,560), vs 94.95% on v1.

In a full quantized run, both land near 73%. That's because the inputs reaching the indexer have already shifted from quantizing the rest of the model, so a BF16 indexer alone can't fix that gap.

v1.1 weights just ~0.15 GB more than v1.

Please note that this is experimental ‼️

Get it here:
https://t.co/yMmNvpb4gj

13

164

9

62

9K

1

2

0

2

340

cryptography engineer and researcher. prev. @renegade_fi @BainCapCrypto, @__geometry__, @PrivacyEthereum, @Consensys

yaklaşık 22 saat önce

@EnsBtc @nvidia @ashxhart @MiaAI_lab Run local ai. Keep your data private. Sovereign intelligence

0

1

0

32

Takip edebileceğin hesaplar

weijie.eth

@weijie_eth

Justin Drake

@drakefjustin

Bitcoin security researcher

nick.eth

@nicksdjohnson

Lead developer of ENS & Ethereum Foundation alum. Certified rat tickler. he/him. bsky: https://t.co/RdZTUkWNq2 wc: https://t.co/nvP0lHPJqz

1 gün önce

I've already convinced more than 1 person to buy a cluster of @nvidia DGX sparks and following @ashxhart and @MiaAI_lab to start their local AI journey. And so the evangelism begins...

2

5

0

2

448

yaklaşık 24 saat önce

@sudoingX how?!

0

683

1 gün önce

@MiaAI_lab Looking forwars to it

0

1

0

206

1 gün önce

@sonderby @Razumasu Is 6 enough to run full glm 5.3?

1

0

61

1 gün önce

@ashxhart @ciprianveg If you get this to 100 tps i might have to get some more sparks…

0

1

0

35

1 gün önce

@jun_song It just looks better…

0

1

0

106

1 gün önce

@MiaAI_lab Wordpress runs half the internet. I started my career on wordpress!

0

1

0

76

1 gün önce

@JamesMeadlock Is the EXL3 also running on 4 sparks?

1

0

99

1 gün önce

Just a tiny little dock makes me want to buy 4 more sparks...

1 gün önce

I want that switch

40

403

15

39

27K

1

5

0

456

2 gün önce

Can confirm, this is a true story.

Jun Song

@jun_song

3 gün önce

Local AI psychosis : If you buy one DGX Spark, you'll be adding a second one to your cart within a few weeks

85

434

10

23

22K

0

4

0

552

3 gün önce

There are a few moments in ENS history that are as momentous as this. We’ve been building for almost 10 years already and this is really the foundations of the next few decades where a programmable, permisionless naming system will be required to shape the next sovereign internet

ens.eth @ensdomains

4 gün önce

The vote to activate ENSv2 on Ethereum mainnet is now live. This is a major step toward launching the next generation of ENS, with the decision in the hands of the DAO. Read the proposal and vote ⤵ https://t.co/zZ7Mj0lyOU

9

142

32

15

13K

10

73

14

5

3K

3 gün önce

Local AI reminds me of Ethereum in 2015. We don’t do it because it’s the easiest, the cheapest or the fastest. We do it because it is right. Sovereign intelligence.

0

18

4

0

573

3 gün önce

It’s kind of mad that i could lose internet at home and i still have access to pretty much frontier level inference.

0

6

0

459

3 gün önce

Leaders in free speech

3 gün önce

I like what DeepSeek is doing. - They have a top of the line open-model. - Top of the line harness. - And now top of the line inference engine. All open-source and available out of the open. Thumbs up for them, I love it.

78

4K

166

177

62K

0

4

1

0

612

3 gün önce

Code is freedom of speech. This applies to smart contracts. That also applies to open weight models. The fight hasn’t changed.

1

7

0

1

549

4 gün önce

It’s kind of wild that in 2026 you can get high quality code for literally free.

Yum⋆₊˚

@yuhasbeentaken

4 gün önce

a few more things i’ve learned replacing Muse Spark 1.3 in OpenCode: 1. DeepSeek V4.1 Flash can last me ~8 hours of work before i hit the daily free limit, as long as i’m not constantly spawning subagents. 2. the 200K context cap is manageable, but i usually hit compaction several times during a long session. for bigger repos, that matters. 3. when context starts becoming the bottleneck, i switch to Space Bunny for the 1M window instead of forcing DeepSeek through repeated compactions. 4. for planning, i still prefer GLM 5.3 Flash, but i don’t use it for everything. on simple build tasks, it can take 2–3x more turns than DeepSeek. 5. if DeepSeek starts making weird implementation choices, MiMo 2.6 Flash is my fallback before i jump to a paid frontier model. right now my free stack is basically: GLM → plan DeepSeek → build Space Bunny → long context MiMo → backup

19

676

25

374

33K

0

5

0

3

716

4 gün önce

@MiaAI_lab Will take a look! Currently still running the 4 spark official weights and taking some benchmarks before i move across!

0

1

0

27

4 gün önce

Been running Mia's recipes on my sparks the last week or so since mine arrived. It's been really nice to see how on it the local AI community is