Aivan Monceller

Created a video how to train Flux LoRAs using my convenience script utilizing ai-toolkit on @quickpodio youtu.be/0I06JHyvuwQ

@ron_joshi Can it clone voice?

Introducing Kitten TTS V0.8: open-source TTS that fits in 25MB. Three variants: 80M | 40M | 14M (<25MB) Highly expressive. Runs on CPU. Built for edge. No GPU? No problem. Ship voice anywhere. Check it out:

@gabriel1 Only if I want to read the code.

only bottleneck is consuming code, so make sure to tell codex that you want just that: "write extremely easy to consume code, optimize for how easy the code is to read. make the code skimmable. avoid cleverness. use early returns."

@Dimillian Definitely my cup of tea ..

Codex is one-shotting a full Animal Crossing PC port to a macOS port, and from 32 to 64 bit

Stack Overflow just turned their users into a free labeling farm. That new "Challenges" tab isn't for fun. It’s a solution to the LLM data contamination problem. By creating "fresh" puzzles, they generate data that no AI has seen yet. They aren't a Q&A site anymore. They are a data moat.

@alexellisuk @chriswinfield @UnslothAI What changed since our last conversation? , it just works right?

@chriswinfield @UnslothAI Nothing special in this case. It looks like @UnslothAI did all the work in the GGUF file. The unlock for me was getting off Ollama and moving to llama.cpp directly. Newer builds and more flexible tuning.

I did a thing.. and now local Claude feels _really_ fast. (Qwen3.5-35B-A3B Q4 M from @UnslothAI - Superterm for the terminal/AI manager)

why moving checkpoints to s3 is a bottleneck: it treats every 10GB save as a 100% new file. @huggingface’s new Storage Buckets use Xet deduplication to solve the training to inference loop: Training: it only uploads the 5% weight changes (content defined chunking). Inference: localized CDN cache lets you stream models directly to GPU nodes without the pull-then-load lag. no git overhead. no full-file re-syncs. just deltas.

@LeeLeepenkman @karpathy @Yuchenj_UW Do you keep up with upstream?

@karpathy @Yuchenj_UW good ad for my codex fork btw - plz try github.com/lee101/codex-i… see what you think :D a lot of this was a issue with prompting inside codex itself

GPT-5.4 xhigh seems bad at following instructions. Last night I launched two AI research agents running @karpathy’s autoresearch. Claude Opus 4.6 (high): > ran for 12+ hours, 118 experiments done, still running GPT-5.4 xhigh: > stopped after 6 experiments > blamed me for “manually interrupting” it > I interrogated it > It admitted it made a mistake and stopped the loop itself, despite an explicit LOOP FOREVER instruction in the md file. 💀

@fahdmirza @zhijianliu_ @QuixiAI Please release 27B and support GGUF , will test in a heartbeat

youtu.be/J3EdGgez1-I?si… @zhijianliu_ @QuixiAI

💥 We ran Qwen3.5 9B at INT4 with ParoQuant — and it barely lost anything 🌀 ♠ A brand new ICLR 2026 quantization method that finally solves reasoning accuracy at 4-bit 🔬 🔹 4x smaller than FP16 — runs locally on consumer hardware 🔹 AIME-24 score drops only 2.3 points vs FP16 — AWQ drops 13 🔹 Smarter than AWQ, faster than QTIP — best of both worlds 🔹 Single fused CUDA kernel — less than 10% overhead 🔹 Works with vLLM and Open WebUI out of the box 🔹 Tested live with a reasoning prompt to see if the quality holds 🔹 The quantization method that reasoning models actually needed 🔥 Watch the full video below 👇

@Ex0byt Will this work on an Nvidia GPU?

3/9/26 Experiment Updates: — Qwen 3.5 INT4 · Raw Safetensors · WebGPU · M4 Max · Pure WGSL 220 tok/s sustained on 0.8B — flat regardless of context window size. Full family: 0.8B: ~220tok/s · 2B: 163 tok/s· 4B: 74 tok/s· 9B: 55 tok/s · 27B: tok/s 17 tok/s (49% of theoretical max) Three underutilization enhancements that got us here: → DeltaNet WG 256→128 — 50% threads were idle → zero wasted SIMD slots → MLP gate+SiLU 4T split — 192→768 workgroups → 4× memory pipeline saturation → Major breaktorugh: FlashDecode splits 32→64 — tok/s now stays flat as context size grows!

@ToNYD2WiLD @DonvitoAI @ComfyUI What do you mean by using WAN to run LTX. The WAN I know is the WAN video models. Is there an inference engine called WAN? is this Wan2GP?

@aivandroid @DonvitoAI @ComfyUI I’m using WAN to run LTX though but I’ll try comfy to see if it works now it was giving issues

Running LTX 2.3 locally in @ComfyUI RTX 3090

@alexellisuk @stevibe Yes I did not use a chat template file, I followed the official docs of unsloth. They did not mention the need to customize these parameters. So I thought you intentionally needed that for some reason. It compacted multiple times , so it did reach 200k context several times.

@aivandroid @stevibe Yes this seems to be a persistent issue. I noticed that you didn't use --jinja or --chat-template-file? How much context did you get up to?

The RTX 3090 is a 5-year-old GPU and it still runs a 27B model at 20 tok/s I tested Qwen3.5:27b across 3 generations of NVIDIA: 5090 → ~60 tok/s 4090 → ~40 tok/s 3090 → ~20 tok/s Perfectly linear scaling. Double the generation, double the speed.

@alexellisuk @stevibe I wonder why, it doesn't have a problem writing and updating files for me.

@aivandroid @stevibe Third time trying to get it to write out its analysis ~ 34k context into a markdown file. Runs intensely for > 2mins every time and never writes a file or any text to disk.

@alexellisuk @stevibe I've been using it continously yesterday over the course of 3 hours on claude code , improving the same single codebase, on a single 3090 , it's slower when it compacts but never really halted

@aivandroid @stevibe Thanks glad I could help. Still a bit bumpy.. I got a great deep dive after almost 5m of processing - it produced beautiful and concise output But then I said "now write that response to ARCH.md" and it took several more minutes then did nothing at all Like I've seen before..

@CoinhubD @ostrisai How do you teach it overtime?

@ostrisai Kimi on a run pod and just teach it over time?

Codex, Claude Code, or something else? What is the current best for a big repo, this week.

Try this 262k context on 3090 this time, I learned from you. This is very useable nn claude code llama-server -hf unsloth/Qwen3.5-27B-GGUF:Q4_K_M -np 1 -ngl 99 -c 262144 -fa on --cache-type-k q4_0 --cache-type-v q4_0 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 --presence_penalty 0.0

Got the same error again with: CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=80 ⎿ API Error: 400 {"error":{"code":400,"message":"request (66049 tokens) exceeds the available context size (66048 tokens), try increasing it","type":"exceed_context_size_error","n_prompt_tokens":66049,"n_ctx":66048}}

@ToNYD2WiLD @DonvitoAI @ComfyUI WAN times are not the same as LTX times. WAN is slower

@DonvitoAI @ComfyUI im doing something wrong then, im doing WAN and it is taking 30mins

This is output of Qwen3.5-27B-GGUF by @UnslothAI Using Claude Code and the frontend skill This is promising!!! Coding just using a local model in a consumer gpu RTX 3090 Use this settings to run llama-server -hf unsloth/Qwen3.5-27B-GGUF:Q4_K_M -ngl 99 -c 262144 -fa on --cache-type-k q4_0 --cache-type-v q4_0 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 Thanks @aivandroid

@thsottiaux make it configurable to the point that there is no reason to fork it for most use cases.

With GPT-5.4 out. What should Codex ship or improve next?

Can someone from @premium X look at this.

The Codex era of coding is here. I’m usually too cheap for paid plans. But I just bought 3 Codex Pro subs for my friends. Why? Because explaining the ROI is impossible. You have to actually use it to believe it.