Episode Details

What a WEEK! Qwen-mass in July. Folks, AI doesn't seem to be wanting to slow down, especially Open Source! This week we see yet another jump on SWE-bench verified (3rd week in a row?) this time from our friends at Alibaba Qwen.

Was a pleasure of mine to host Junyang Lin from the team at Alibaba to come and chat with us about their incredible release with, with not 1 but three new models!

Then, we had a great chat with Joseph Nelson from Roboflow, who not only dropped additional SOTA models, but was also in Washington at the annocement of the new AI Action plan from the WhiteHouse.

Great conversations this week, as always, TL;DR in the end, tune in!

Open Source AI - QwenMass in July

This week, the open-source world belonged to our friends at Alibaba Qwen. They didn't just release one model; they went on an absolute tear, dropping bomb after bomb on the community and resetting the state-of-the-art multiple times.

A "Small" Update with Massive Impact: Qwen3-235B-A22B-Instruct-2507

Alibaba called this a minor refresh of their 235B parameter mixture-of-experts.

Sure—if you consider +13 points on GPQA, 256K context window minor. The 2507 drops hybrid thinking. Instead, Qwen now ships separate instruct and chain-of-thought models, avoiding token bloat when you just want a quick answer. Benchmarks? 81 % MMLU-Redux, 70 % LiveCodeBench, new SOTA on BFCL function-calling. All with 22 B active params.

Our friend of the pod, and head of development at Alibaba Qwen, Junyang Lin, join the pod, and talked to us about their decision to uncouple this model from the hybrid reasoner Qwen3.

"After talking with the community and thinking it through," he said, "we decided to stop using hybrid thinking mode. Instead, we'll train instruct and thinking models separately so we can get the best quality possible."

The community felt the hybrid model sometimes had conflicts and didn't always perform at its best. So, Qwen delivered a pure non-reasoning instruct model, and the results are staggering. Even without explicit reasoning, it's crushing benchmarks. Wolfram tested it on his MMLU-Pro benchmark and it got the top score of all open-weights models he's ever tested. Nisten saw the same thing on medical benchmarks, where it scored the highest on MedMCQA. This thing is a beast, getting a massive 77.5 on GPQA (up from 62.9) and 51.8 on LiveCodeBench (up from 32). This is a huge leap forward, and it proves that a powerful, well-trained instruct model can still push the boundaries of reasoning.

The New (open) King of Code: Qwen3-Coder-480B (X, Try It, HF)

Just as we were catching our breath, they dropped the main event: Qwen3-Coder. This is a 480-billion-parameter coding-specific behemoth (35B active) trained on a staggering 7.5 trillion tokens, with a 70% code ratio, that gets a new SOTA on SWE-bench verified with 69.6% (just a week after Kimi got SOTA with 65% and 2 weeks after Devstral's SOTA of 53% 😮)

To get this model to SOTA, Junyang explained they used reinforcement learning with over 20,000 parallel sandbox environments. This allows the model to interact with the environment, write code, see the output, get the reward, and learn from it in a continuous loop. The results speak for themselves.

With long context abilities 256K with up to 1M extended with YaRN, this coding beast tops the charts, and is achieving Sonnet level performance for significantly less cost!

Both models supported day-1 on W&B Inference (

Listen Now