Qwen 3

1. Which kind of base/fine-tunes have you done? And which data did you use? (www.reddit.com)

   +1 7w fine-tuningqwen

2. TPU v7x Ironwood vs Nvidia B200 (www.reddit.com)

   +2 7w moe

   Google published Ironwood inference benchmarks in their AI-Hypercomputer/tpu-recipes repo. Nvidia has InferenceMAX numbers for B200.

3. Tokens per second - RTX 5000 Ada generation (www.reddit.com)

   +14 7w ollama

   Hi everyone, I am testing the LocalLLaMA. I have a laptop with an RTX 5000 Ada generation, with Ollama and Open Webui.

4. Deepseek-r1 thinks for 30 minutes? (www.reddit.com)

   9 7w deepseek

   I was trying to ask a question about coding using DeepSeek-R1-0528-Qwen3-8B-Q4_K_M, and the thinking took 30 minutes??? https://preview.redd.it/kex3fgg4lgvg1.png?width=277&format=png&auto=webp&s=5f7e7cdc8502b935ea8b8fb83e0e4af60c3c4533 I h…

5. Show HN: Flint – A 30B model fine-tuned for less repetition (springboards.ai via hn)

   +5 7w mmlu

   As frontier LLMs have very little output diversity even for open ended queries. We built Flint to see if we could reverse this.

6. Compile English function descriptions into 22MB neural programs that run locally via llama.cpp (www.reddit.com)

   +157 7w llama

   We built a system where a neural compiler takes a plain-English function description and produces a "neural program" (a combination of a continuous LoRA adapter and a discrete pseudo-program). At inference time, these adapt a fixed interpr…

7. Qwen 3 Coder Next has a bug! Help Test? (www.reddit.com)

   31 7w qwen

   Hey y'all. So I've stumbled upon a really specific and esoteric "bug" where an llm can't comprehend a URL in like, 90% of scenarios.

8. inference on the Qwen3 -Coder-480B-A35B-Instruct with 4xH200 (www.reddit.com)

   7 7w

   Hello guys, I want to do the inference on the Qwen3 -Coder-480B-A35B-Instruct. I have a 4xH200 machine.

9. openrouter/elephant-alpha is 99% Chinese, likely Qwen 3 Nex (www.reddit.com)

   4 8w qwen

   openrouter/elephant-alpha is 99% Chinese, likely Qwen 3 Next. промт "Напиши сложный алгоритм на Python для анализа временных рядов, используя методы из китайских научных работ по эконометрике.

10. Built a Japanese ASR benchmark because existing ones can't measure quality differences properly (www.reddit.com)

    +9 8w fine-tuning

    Was fine-tuning a Japanese ASR model (based on Qwen3-ASR) to handle technical terminology better. The model clearly improved — "Next.js" comes out as "Next.js" instead of "ネクストジェイズ", punctuation works, etc.

11. ClaudeCodeCLI vs OpenCode vs Cline vs QwenCode (www.reddit.com)

    +24 8w clinemcp

    ClaudeCodeCLI vs OpenCode vs Cline vs QwenCode Local coding LLM - Qwen3-coder-next-80b-nvfp4 Wich "tool" do you can recommend for it, and with "Skills/Plugins/MCP's"?

← all threads