Jean-Baptiste Chalnot

Actuary by training, engineer at heart, operator by necessity. Co-founder @ Gedeon & KaribU. Homelab tinkerer & certified wine nerd.

[Jean-Baptiste Chalnot @JBChalnot]

@sudoingX Qwen3.8-Flash-Next. At the current release cadence, ask me again tomorrow😂

[Jean-Baptiste Chalnot @JBChalnot]

@Rayan_Nezzar

[Jean-Baptiste Chalnot @JBChalnot]

@ivanfioravanti Vision came at a cost though: it’s noticeably more verbose now.

[Jean-Baptiste Chalnot @JBChalnot]

@nickgraynews Not every workload is about maximizing tokens per dollar. A lot of async automation doesn’t need frontier capabilities or low latency, and sometimes you simply don’t want private data leaving your network.

[Thanh Pham @runsonai]

Spent a couple days making Qwen 3.8 flash-next on my m3 ultra studio less painful on long prompts.

So it now powers my Hermes agent.

Started around 88 tok/s decode. An overnight restack got decode to 99 tok/s. TTFT improved & streaming felt a bit better.

M4/m5 Macs already have a nax flash kernel for this model's sparse attention. M3 doesn't have those tensor units so that path never turns on. So had to figure this out for my m3 ultra. First time writing a custom kernel, oh fun!

What it feels like now: hermes is faster and same quality. It solves my random problems consistently well, this model is good. Follows instructions well, it's fast and can fix stuff. Great daily driver.

Shoutout to @Youssofal_ for great work he's doing with MTPLX so mac users can run models faster. I just did some additional testing & tweaking so Qwen 3.8 flash could be even faster and responsive.

[Jean-Baptiste Chalnot @JBChalnot]

@RimasXYZ @MiaAI_lab For some reason everyone jumped on NVFP4, but I don’t think it makes much sense for this model. The routed experts are only a small fraction of the bytes moved per token, so the throughput gain is limited.

[Jean-Baptiste Chalnot @JBChalnot]

@MiaAI_lab I’d keep Qwen3.8-Flash in FP8. NVFP4 degrades it too much, and the decode gains are partly offset by the increased verbosity. On my own tests, FP8 Qwen3.8-Flash beats DS4F Vision, which also seems to have lost the conciseness of 0731.

[Jean-Baptiste Chalnot @JBChalnot]

@DataChaz @huggingface @pollenrobotics I told my wife I ordered one to build a programmable baby monitor.

Little did she know what she had just agreed to 😂

[Jean-Baptiste Chalnot @JBChalnot]

@hkdom @MiaAI_lab So far, I find the vision noticeably worse than Qwen3.8-Flash, and it seems to have lost the conciseness I loved in 0731

[Jean-Baptiste Chalnot @JBChalnot]

@skalskip92 @OpenRouter @Sentdex You should try Qwen 3.8 flash next

[Jean-Baptiste Chalnot @JBChalnot]

@Tech2Wild @NVIDIAAI How do you find the verbosity vs 0731? Feels like it lost the conciseness I loved

[Jean-Baptiste Chalnot @JBChalnot]

@MiaAI_lab Fast, yes. But damn it loves to talk 😅

[Jean-Baptiste Chalnot @JBChalnot]

@MiaAI_lab @NVIDIAAI Damn, you cook fast 🔥

[Jean-Baptiste Chalnot @JBChalnot]

@jun_song What backend/kernel are you using? SGLang seems pretty slow, and as far as I can tell it doesn’t support speculative decoding with vision yet. Waiting for a new vLLM image…

[Jean-Baptiste Chalnot @JBChalnot]

@Tech2Wild Can’t wait until vLLM can serve DS4F Vision.

[Jean-Baptiste Chalnot @JBChalnot]

@MiaAI_lab Every day feels like Christmas when you’ve got local compute 😂

[Jean-Baptiste Chalnot @JBChalnot]

@rben_ll D'accord avec toi sur le fond : les benchs ne disent qu'une petite partie de la vérité. Mais ne t'emballes pas trop loin dans l'autre sens avant d'avoir fini tes tests.

[Jean-Baptiste Chalnot @JBChalnot]

@skalskip92 Match my results: GLM 5.3 flash is indeed bad at grounding even if general vision is OK. You should test Qwen 3.8 flash next. It matches Qwen 3.5 397b on my own grounding benchs.

[Jean-Baptiste Chalnot @JBChalnot]

@0xWhiteMage So this voice in my head was you the whole time ??? 🤣

[Jean-Baptiste Chalnot @JBChalnot]

@rben_ll J’avais vraiment envie d’aimer Mistral (cocorico 🇫🇷) mais mon expérience avec Mistral Small 4 en self-hosted a été franchement décevante comparés aux modèles chinois de tailles comparables.

Je suis curieux de voir tes résultats sur le Large 3 du coup.