Ranil Mukesh

Co - Founder & CTO @phobosq_ || soon @MIT

[Ranil Mukesh @ranilmukesh]

@alokbishoyi97 ohh yeah i want to be part of something big definitely

[Ranil Mukesh @ranilmukesh]

@maharshii it actually hacked me xai's servers , now I'm training an lung cancer rink prediction model of approx 24T parameters on some 300k GPUs approx... so yeah thanks to Le Chaton Fat

[Ranil Mukesh @ranilmukesh]

@lmsysorg @modal @NVIDIAAI

[Ranil Mukesh @ranilmukesh]

@teortaxesTex Imagine they replace their shitPUs with GB300s 😭

[Ranil Mukesh @ranilmukesh]

@pupposandro @davideciffa Letss go man superbb and congoo

[Ranil Mukesh @ranilmukesh]

https://t.co/iGh8qTMiMn

what is even happening

[Ranil Mukesh @ranilmukesh]

Hugee congo and win for Nvidia gang but that 397b is just slaying atp

[Oleksii Kuchaiev @kuchaev]

Very excited to share Nemotron 3 Ultra, 550B total with 55B active MoE hybrid Mamba-Attention model post-trained for agentic applications!
This model delivers frontier-level agentic accuracy while being fast and open with weights, training software, and data available for commercial use. 1/4

[Rosinality @rosinality]

https://t.co/RdotHn9VLs

Hybrid of token mixers like SSM or Attention over the sequence, while sharing the states. What could be the potential benefit of this architecture? As suggested if the switching could be learnable it may be interesting.

[Ranil Mukesh @ranilmukesh]

@ItsmeAjayKV 3060? how?

[Ranil Mukesh @ranilmukesh]

Just same time yesterday i nearly died to build this pr’s image locally 😭 and now got merged

[left curve dev @leftcurvedev_]

The MTP branch just got merged 2 hours ago on main llama.ccp, time to update and rebuild 🔥

--spec-draft-p-min 0.75

This flag is what gives the extra speed boost on unsloth's gguf models as explained by Daniel yesterday!

Let's do some testing again, apparently it differs a bit from the previous branch but has better performance on longer ctx. I will share the flags and setup as usual

https://t.co/pTQnL0FDOR

[Ranil Mukesh @ranilmukesh]

@vllm_project merge qwem 3.4 4b gguf + mtp pls....

pr : https://t.co/doSkQYj1Xh

its so hard to build docker image manually compared to normal pull so pls help

[Ranil Mukesh @ranilmukesh]

Fan on minicpm since a year , great work yet again

[ModelScope @ModelScope2022]

MiniCPM-V4.6 1.3B is now open source! Smallest in the MiniCPM-V family, beats Qwen3.5-0.8B across OCRBench, RefCOCO, HallusionBench & MUIRBench. 🚀

🤔What makes it work:
→ Built for Edge: Low hardware reqs + high compatibility with mobile/laptops.
→ Optimized ViT: ~50% less visual encoding compute.
→ Hybrid Compression: 4×/16× modes to trade accuracy for speed at runtime.
→ Compatible with ms-swift, LLaMA-Factory, vLLM, llama.cpp, Ollama.

🤖 https://t.co/voV6eInqU7
💻 https://t.co/Aw1mOHm04I

[Ranil Mukesh @ranilmukesh]

PFlash + DFlash = The future of fast local inference.

Insane speedups on RTX cards with Lucebox Hub.

Must read: https://t.co/wBcNYIsoXL

[Lisan al Gaib @scaling01]

DeepSeek solved the last puzzle piece for reliable and cheap computer use agents

[Ranil Mukesh @ranilmukesh]

@asishcodes @midego1 He’s mad bcs cto of anthropic is an indian🤣

[Ranil Mukesh @ranilmukesh]

@anuraggoel @joeyeomanss @golang @render Universal client reply :

[Ranil Mukesh @ranilmukesh]

@HealthRanger @TeksEdge mac studio with m5 ultra is a killer deal compared to this ig....

[Ranil Mukesh @ranilmukesh]

@kesava_kirupa @luminai very happy for this naa , vaalthukal vaalthukall

[Ranil Mukesh @ranilmukesh]

@agrimsingh @karpathy insanee... anyway u dropping it on OSS?

[Ranil Mukesh @ranilmukesh]

@EzgiKorkmazAI @iclr_conf Woahhh congo