There are 10 types of people in the world - those who know binary and those who don’t | coding DAGKnight eprint.iacr.org/2022/1494.pd…
Based in North America
- Tweets 1,848
- Following 183
- Followers 3,922
- Likes 2,348
I made the fastest inference stack for Qwen3.8-27b on a RTX 3090.
Up to 195k context, 82 tps single request, 672 tps peak 64 concurrent.
github.com/syv-ai/qwen38-27b…
4
2
38
2,568
Thanks for this, I’ve been testing it at 64k it’s amazingly fast. At CTX=long it’s still performing really well (eyeball figures around 50+tps on a single 3090).
1
8
557
Learning about $KAS today.
Haven’t found a project that interested me in years, will have to continue to research this one.
Community is strong, lots of genuine people.
POW + on chain settlement that scales- 10 blocks per second- is an interesting combination.
Feels like Bitcoin base layer + lightning in one.
Delving more into issuance, difficulty adjustment, and security.
Kaspium is a solid wallet, easy to use and no KYC bs. @KaspaPowerLaw @realvijayk @NakaSompo
Bitcoin=Orange Pill, Kaspa=Teal Pill? 🟢
86
181
741
23,362
The quoted text doesn’t directly answer the question so here goes:
1. GD protocol rules say that you (1) point to all tips you know and (2) propagate blocks asap. (1) makes it so all* tips eventually get pointed at so let’s say you have 15 tips, your mined block is expected to point to all of them. Simply put this process converges. See attached part of the Phantom Ghostdag paper.
(*there’s some more nuance to this but I’ll skip that for now)
2. GD and also DK has this concept of “selected parent chain” and particularly important is the **virtual** selected parent chain. This chain has the property that it doesn’t change ordering as blocks confirm apart from near the tips (I’m glossing over details and nuances here).
3. Ordering of blocks is based on this robust chain, so the global ordering is also robust (after some number of confs)
Hope that helps.
1
1
8
135
To help visualize this “virtual selected parent chain” take a look at the blue line connected blocks here: kgi.kaspad.net/
1
9
107
Qwen3.8-27B can now be run locally! ✨
Run on 17GB RAM via Unsloth Dynamic GGUFs.
Qwen3.8-27B is by far the strongest model for its size. We also uploaded NVFP4 quants.
GGUF: huggingface.co/unsloth/Qwen3…
Guide: unsloth.ai/docs/models/qwen3…
We promised open weights for Qwen3.8. Now, time to meet them! 🎉
⚡ Qwen3.8-27B:
- A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows.
- 262K native context, easily extendable to 1M tokens via YaRN.
- Built for builders. Highly efficient, high-quality, and licensed under Apache 2.0.
🚀 The open weights for Qwen3.8-2.4T-A95B (Max-level) have also been released recently.
Whether you're shipping lightweight applications with Qwen3.8-27B locally or building agents with Qwen3.8-2.4T-A95B, they're yours now!
Download, deploy, and build something we haven't imagined yet. 👀👇
- Hugging Face：
huggingface.co/collections/Q…
- ModelScope：
modelscope.cn/collections/Qw…
204
622
5,170
1,327,846
Nice to be back home after 8 days of intense devving sessions with several contributors to Kaspa’s core development, some of whom I had never met before. We exchanged a l o t of information and brainstormed different paths forward and their tradeoffs across several parallel topics.
On a personal level, my public coding activity was somewhat on hold during this period, including Argent progress and coding sessions I wanted to do (and share via @KaspaNative). My next few weeks will center on initiating the DK group effort -- with several ppl joining @coderofstuff_ in his already long journey deep in the trenches (actually, @FreshAir08 is already deep in as well with a very promising incremental cascade-voting algo). Alongside that, I’ll also find time to move Argent towards a release and perhaps take a more active role in helping the ongoing covenant-related spec & conventions work (KCCs) converge
40
170
675
15,940
coderofstuff retweeted
Nice to be back home after 8 days of intense devving sessions with several contributors to Kaspa’s core development, some of whom I had never met before. We exchanged a l o t of information and brainstormed different paths forward and their tradeoffs across several parallel topics.
On a personal level, my public coding activity was somewhat on hold during this period, including Argent progress and coding sessions I wanted to do (and share via @KaspaNative). My next few weeks will center on initiating the DK group effort -- with several ppl joining @coderofstuff_ in his already long journey deep in the trenches (actually, @FreshAir08 is already deep in as well with a very promising incremental cascade-voting algo). Alongside that, I’ll also find time to move Argent towards a release and perhaps take a more active role in helping the ongoing covenant-related spec & conventions work (KCCs) converge
40
170
675
15,940
Even with just 24gb vram you could run the qwen3.6-27b already. There’s good info on that here
"how do you fit qwen 3.6 27b q4 on 24gb at 262k context" lands in my dms 5 times a week. here is the exact memory math.
model bytes at idle = 16gb (q4_k_m of 27b dense)
kv cache at 262k context with q4_0 for both k and v = 5gb
total = 21gb on the card
headroom = 3gb for prompts and tool call traces
the magic is the kv cache type. most people leave it at default fp16 or push to q8 thinking quality wins. on qwen 3.6 27b dense at 262k:
- fp16 kv cache = does not fit at all
- q8 kv cache = fits at 23gb but runs 3x slower (double penalty: more vram, less speed)
- q4_0 kv cache = fits at 21gb at full speed (40 tok/s flat curve, same speed at 4k or 262k)
most builders never test the kv cache type because tutorials never mention it. it is the single biggest unlock on consumer 24gb hardware.
flags i run:
./llama-server -m Qwen3.6-27B-Q4_K_M.gguf -ngl 99 -c 262144 -np 1 -fa on --cache-type-k q4_0 --cache-type-v q4_0
what they do:
-ngl 99 = offload everything to gpu
-c 262144 = 262k context window
-np 1 = single user slot (do not enable multi-slot, eats headroom)
-fa on = flash attention on (memory and speed both win)
--cache-type-k q4_0 --cache-type-v q4_0 = the unlock
if you are sitting on 24gb and not running this config, you are leaving 250k of context on the table. or worse, you are running q8 kv cache and burning 3x your speed for nothing.
q4 is not a compromise on consumer hardware. it is the right call.
78
Cant wait for dagknight😍
Thank you all core devs for hard work.
Follow coder of stuff for more dagknight insight😍
$kas
Behind this merge and recent annoucement are countless hours of sifting through every code path, tests, retests and head down coding by @michaelsuttonil @Max143672 @OriNewman @IzioDev et al. Finally, toccata is merged in master. Releases coming up soon (tm).
Any guesses for the mainnet DAA activation score?
1
7
42
2,038
"Let's do a quick Dagknight lecture"
Reality: 1h30 recording of @michaelsuttonil deep diving into ordering challenges, tie-breaking and grey colouring.
Time for a cold beer now 🤯
13
40
239
4,662
Vlad, how can you be so right about so many Bitcoin topics?
Well, I’m glad you asked. I have 3 simple principles:
1. I look up to the builders, the visionaries & the guys I consider smarter than me. Not those who have been historically right, not the narrative gatekeepers. The guys who want Bitcoin to do everything and have a plan to make it happen.
2. I don’t care about what the others think. The high priests of the church of maximalism, the plebs, the pseudo celebrities like Max Keiser and the consensus craftsmen like Greg Maxwell or Mr. Hodl. I do my own thing without caring about public perception or whatever.
3. I always assume that I might be wrong and prepare accordingly. You can call me out for not having enough conviction, I don’t care. I’m here to breathe fresh air, not smell brainfarts in the same room philosophizing about their purity. This means sometimes I have to leave the room and take a look (and a sniff) from outside.
9
7
90
4,196
coderofstuff retweeted
To busy to share stuff live but really excited about this week-long meetup we have going on “somewhere in Europe” like @hashdag says. DK, P2P, vprogs, silver & argent, fascinating ppl, great food, what not. Kaspa has a bright future bcs it has many dreamers with far-reaching vision and execution capabilities behind it. Let me just put it that way.
55
285
920
44,740
