Qwen 3

1. As of May 2026 LongCat Dit 3.5B and Moss TTS 8B are the best SOTA tts models and Qwen tts is not even close. (www.reddit.com)

   8 5w qwen

   [Disclaimer: i am totally avoiding fish audio s2 pro because its not a real open-sourced model(non commercial license)] So the context is i asked many ai to give me best tts model as of now but most of it said qwen 3 tts, and voxtral etc.…

2. Two flat-fee agent endpoints, no token meter: OpenClaw chat ($7/mo, 128K ctx) + All You Can Code ($19/mo, 256K ctx). OpenAI v1. (www.reddit.com)

   +13 5w aiderclineopenclaw+1

   For anyone running agents (coding or otherwise) who'd rather pay a flat fee than meter tokens. Two tiers, both flat fee, both unlimited: OpenClaw ($7/mo) - Nemotron-3-Nano-Omni-30B-A3B - 128K context - For general-purpose agents: research,…

3. I trained TIME: short context-triggered thinking on Qwen model instead of overthinking (www.reddit.com)

   +31 5w qwen

   Started this as a personal project for my Open-WebUI setup to use. Somehow it ended up as an ACL 2026 paper.

4. MiroThinker-1.7, an open-weight deep research agent (Qwen3 MoE base) — mini is 30B/3B active, curious what tok/s people get on consumer hardware (www.reddit.com)

   +99 5w moe

   As usual, disclosure first: I'm on the team that built this. Our MiroThinker-1.7-deepresearch and 1.7-mini-deepresearch API went live, mini is a deep research agent built on Qwen3 MoE (30B total, 3B active for mini).

5. How are you all handling state for long-running agents? Stateless sandboxes are eating my evenings (www.reddit.com)

   +12 5w

   ok I want to know if I am the only one. been running a local coding agent against qwen3 coder on a 4090 box, with a remote sandbox for the actual code execution.

6. Case Study: Dogfooding a Facebook Agent Before Deploying It to a Realtor (www.reddit.com)

   +11 5w operator

   A real estate firm came to us wanting an AI agent that could run their Facebook page. Not a scheduler.

7. Came home to find Pi with Qwen3.627B had run rm -rf ..... (www.reddit.com)

   +2934 6w

   on the build cache because it had run my computer out of disk space. So I assign my coding agent (pi) a task, and then leave the house.

8. Predicting Rare LLM Failures with 30× Fewer Rollouts (www.lesswrong.com via hn)

   +2 6w qwen

   TL;DR: We estimate how often Qwen 3 4B exhibits rare harmful behaviors with 30× fewer rollouts than naive sampling, using a new method that interpolates between the model and a less-safe variant in logit space. Authors: Francisco Pernice (…

9. Qwen/WebWorld 32B/14B/8B (Qwen3 finetune) (www.reddit.com)

   +63 7w gpt-5qwen

   WebWorld is a large-scale open-web world model series for training and evaluating web agents. It is trained on 1M+ real-world web interaction trajectories via a scalable hierarchical data pipeline, supporting: Long-horizon simulation (30+…

← all threads