Alex

building sovereign infrastructure and working on a model-specific gfx1151 inference engine at

[KRAFTON AI @Krafton_AI]

We’re releasing Raon-1 nano, an effective 1.5B dense LLM with per-layer embeddings and 3.1B total parameters. ⚡

Our journey from PUBG Ally toward on-device LLMs continues.

We pruned and distilled Raon-1 mini preview (30B-A4B) using our MoE-to-Dense approach, then further trained it on 3T tokens, followed by SFT, OPD, and RL.

Designed for natural conversations, maintaining conversational context, instruction following, and tool use in both Korean and English.

Small model. Smart interactions. Enjoy! 😊

🤗 Weights (Apache 2.0 license): https://t.co/ZVMZmlSNlz
📄 MoE-to-Dense (NeurIPS 2026 Oral): https://t.co/eoYH86lBs0
🪖 Learn more about PUBG Ally: https://t.co/A3iZvP2X7c

[Alex @DevOuterReaches]

hell of a day for big j

[Alex @DevOuterReaches]

@elder_plinius we all have those thoughts sometimes

[Alex @DevOuterReaches]

added tensor parallelism to my internal llama.cpp fork.

decode now hits 32 t/s for ds4.1 flash running at tp=2 on gorgon halo + strix halo connected by usb4.

[Alex @DevOuterReaches]

@KireiStudios @iBosgame yeah, stock thermals aren't great on it, even with the updated fan curve. i had to keep it in balanced mode because of that

repasting may help, but i haven't tried that yet

[Alex @DevOuterReaches]

@prestonjbyrne kinda pointless considering anyone can get ds4.1 flash running on 3060 12gb and 64gb of ddr3 now

could put a box together for less than 1k usd

[Alex @DevOuterReaches]

@Evankuoj @theojaffee possible too. but it invites some questions regarding their judgement and suitability for the role if they just accept everything uncritically

[Alex @DevOuterReaches]

@MrKapitalist @prestonjbyrne home setups costing less than $10k usd can already serve ds4.1 flash to a couple of users at a reasonable speed using no more than 200w

and if you don't mind the slow speed, can stream k3.1 pro off an ssd on a ~$3100 128gb strix halo (see my github)

can't stop the signal

[Alex @DevOuterReaches]

@hwchase17 may as well skip to incentive engineering

[Alex @DevOuterReaches]

mostly for larping on hikes and camping trips. could be useful in a real off-grid survival situation too, but hopefully it won't come to that.

currently trying to get lfm3-omni-2b to run well enough to perform light agentic tasks. had to resort to weight-space programming to fix its tool calls

[Alex @DevOuterReaches]

picked up a panasonic toughpad fz-g1. i5-7300u, 8gb of ram, 512gb ssd. most interesting parts are the flir thermal camera and the gps.

built out a game-inspired field terminal app for it. map, sensors, gps-based proximity alerts, tak, journal, and soon agent integration via mcp/api.

code coming soon

[Alex @DevOuterReaches]

to be fair most people would end up crying if forced to listen to dario for long

[Luke Burgis @lukeburgis]

When I spoke to an Anthropic employee who invited me to their HQ for a "Christian leaders" gathering, they told me that when they hear Dario speak about morality, they often cry. That was one of the key moments I realized I was out. If you watch this man speak and are moved to tears, you probably need to talk to a psychologist. And you should not be in a position of power at any frontier lab, because we need anti-mimetic brakes on the spreading of psychosis.

[Dimitris Papailiopoulos @DimitrisPapail]

https://t.co/A6AtCXG3Jp

[Qwen @Alibaba_Qwen]

Qwen4 is coming.

Deeper knowledge. Richer world understanding. Stronger coding and agentic capabilities.

Built for real-world tasks — from complex coding and knowledge work to front-end builds and professional software workflows.

Qwen4-Max is now in preview on Alibaba’s Token Plan. You can also get early access to the new model on QwenWork, Qoder, Qwen APP, and QwenStudio.

Put it to the test on a real problem. Show us what you build. 👀

Token Plan: https://t.co/LOB2Jdt3Sk

[Alex @DevOuterReaches]

having an agent set up a new windows box is actually tolerable now.

tell it "make it not suck" and go drink coffee, come back to every single microsoft app missing, no notifications, no defender, no firewall, 500+ entries in hosts file. just the way i like it.

[Alex @DevOuterReaches]

@ryanvogel i have to look at this garbage way too often. how is it 2026 and they still haven't figured out to fix their godawful login page?

[Alex @DevOuterReaches]

@yenwod_ hft over http, pretty funny

[Alex @DevOuterReaches]

@UK_Daniel_Card wouldn't most attackers just have it learn from failures and eventually develop better attack methodologies?

throw it at a defended network you control, create a feedback loop, and it will eventually stop being so noisy and shitty