Herlon Aguiar

I like computers =)

[Herlon Aguiar @mxfp0]

For Local AI I'm using GLM 5.3 Flash (on 2x DGX Sparks) + Qwen 3.8 27B (RTX 5090).
This gives me a very nice combination of Intelligence + Speed.

GLM 5.3 Flash can spawn sub-agents that are very fast and control them.

GLM 5.3 Flash usually gives me around 30 tok/s and can spawn sub-agents using Qwen 3.8 27B with around 130 tok/s each.

[Herlon Aguiar @mxfp0]

@groypte https://t.co/tpaFMA1No7

[Herlon Aguiar @mxfp0]

First time using an Qwen 3.8 27B Abliterated, just 🤯

[Herlon Aguiar @mxfp0]

@JoesInvestments @sfxnz @O80925253 could that improve yours too?

[Herlon Aguiar @mxfp0]

OpenAI want to get into the bot battle for sure 😅

[OpenAI @OpenAI]

Get ready.

[Herlon Aguiar @mxfp0]

@keennay lol this caught me off guard 😂

[Herlon Aguiar @mxfp0]

I still can't believe Anthropic still doesn't allow using Claude in another harness 🤦‍♂️

[Claude @claudeai]

Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family.

It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.

[Herlon Aguiar @mxfp0]

@defilan This is amazing!

[Herlon Aguiar @mxfp0]

@mmastrac Really great work! Thanks!

[Herlon Aguiar @mxfp0]

@landontgreen @Zai_org GLM 5.3 Flash is GOAT for local. I'm looking forward for the future iterations 🚀
@Zai_org soon ™️ ?

[Herlon Aguiar @mxfp0]

This verbosity control in opencode cli is so good, idk why they took so long to add it!

[Herlon Aguiar @mxfp0]

@TechMDAI @O80925253 I'm really glad to see people testing this, totally deserves more attention!

[Herlon Aguiar @mxfp0]

@MrSage

[Herlon Aguiar @mxfp0]

@MrSage Make sure to use PETG material, because of the heat

[Herlon Aguiar @mxfp0]

@MrSage Yeah, it's just the 3d files, I uploaded them to https://t.co/v7HDZLsnZ7 then tried to find the cheapest price in the world 😂 for me it's around 36 euros. Which sounds pretty expensive, given this is something very small, but I'm not going to buy a 3d printer only for that haha

[Herlon Aguiar @mxfp0]

@MrSage I'm using this https://t.co/YQHdoQVJqA for the pairs, temps are pretty good during load

[Herlon Aguiar @mxfp0]

@mmastrac Pretty nice! I'll keep eye on this too, looking forward to see it breaking the 80 tok/s prose C=1 🚀

[Herlon Aguiar @mxfp0]

@O80925253 how can I buy you a coffee? Amazing work!

[Herlon Aguiar @mxfp0]

WOW, this is the best one for 4 sparks switchless! 🤯🤯🤯🤯🤯
!! 80 tok/s C=1 PROSE, 5K tok/s prefill !!

=== Decode ===
Prose:
×1 80.1 agg 80.1/str · TTFT 237ms
×2 105.5 agg 53.7/str · TTFT 178ms
×3 130.0 agg 45.9/str · TTFT 217ms
×4 144.0 agg 39.6/str · TTFT 243ms

Structured
×1 154.3 agg 154.3/str · TTFT 104ms
×2 192.1 agg 98.3/str · TTFT 195ms
×3 267.2 agg 90.3/str · TTFT 179ms
×4 287.4 agg 75.0/str · TTFT 235ms

Code
×1 148.2 agg 148.2/str · TTFT 181ms
×2 195.6 agg 100.6/str · TTFT 286ms
×3 247.5 agg 86.0/str · TTFT 356ms
×4 263.6 agg 78.3/str · TTFT 1.07s

=== Prefill ===
1k 2385.1 tok/s · TTFT 440ms · 1049 tok
2k 2728.3 tok/s · TTFT 760ms · 2073 tok
4k 4021.7 tok/s · TTFT 1.03s · 4123 tok
8k 4661.2 tok/s · TTFT 1.76s · 8219 tok
16k 4988.3 tok/s · TTFT 3.29s · 16409 tok
32k 5079.9 tok/s · TTFT 6.46s · 32791 tok
64k 5139.5 tok/s · TTFT 12.76s · 65561 tok
128k 5124.3 tok/s · TTFT 25.58s · 131096 tok

[O @O80925253]

GLM-5.3-Flash NVFP4 recipe wip: performance-optimized DGX Spark TP 2/4

TP4: RigMark prose 61.7
sparkDash prose 82.1 (1.52×),
32K prefill 5,069 tok/sec
Peak concurrency 346 tok/sec at x16

[Herlon Aguiar @mxfp0]

@MrSage check this out!

[Herlon Aguiar @mxfp0]

@alexellisuk @jnardiello I'm setting up this one right now. Seems quite promising!