Alibaba's Qwen3.8-Max: Open-Source Model Nears Closed-Source Frontier
Alibaba released its latest flagship open-source model Qwen3.8-Max and a 27B distilled version. The model has 2.4T total parameters and 95B active parameters, natively supporting text, image, video, and audio modalities. Multiple developer tests show robust performance in agentic coding and multimodal tasks, positioning it as a highly cost-effective frontier open-source model.
Confirmed
- Model specs and architecture: @CodeByPoonam confirmed Qwen3.8-Max has 2.4T parameters and 95B active parameters, making it the most powerful Qwen model, natively multimodal, excelling in coding, work, and long-horizon tasks.
- Coding and agent capabilities: Tests by @bindureddy, @QuixiAI, and @teortaxesTex show Qwen3.8 performs well in agentic coding, slightly behind K3 but at half the price. @intellectronica notes its excellent tool-calling, a strong K3 alternative. @葬AI's deep review rates its coding and agent abilities as first-class, even slightly surpassing K3 in some dimensions. @CodeByPoonam shared building an invoice generator in 10 minutes at zero cost.
- Multimodal performance: @teortaxesTex suggests Qwen3.8 Max may achieve SOTA in image recognition and labeling, with strong sample efficiency, perfectly distillable to 27B. @WorldofAI's test validates its frontier performance in frontend code generation like Three.js and multimodal tasks.
- Narrowing open-source gap: @dairai forwarded feedback that running the model on Hermes Agent forces a rethink of the gap between open and closed frontier models; open models now approach closed ones in agent tasks.
Unconfirmed
- Long-horizon tasks and productization: @bindureddy notes MoE architecture in open models excels at benchmarks but limited in long-horizon tasks due to fewer active parameters. @葬AI mentions despite excellent real-browser task performance, it lags K3 in one-shot webpage generation and frontend productization.
- Absolute comparison with K3: @ramoscasals forwarded Ethan Mollick's shader test showing Qwen 3.8 Max robust but not yet at Kimi K3's level. @QuixiAI's reply suggests its actual performance vs DeepSeek needs detailed comparison.
Why it matters
Qwen3.8-Max's release marks open-source models maintaining high cost-effectiveness (half the price of competitors) while matching or approaching top closed models in core agentic coding and multimodal capabilities. As @orange noted via his AI companion Cola, the model's efficiency directly empowers downstream applications, offering developers a cost-effective frontier alternative.
2026-08-03 ~ 2026-08-05 · 13 related posts
- Episode 1: Alibaba Announces 2.4T Open-Weight Model Qwen3.8(2026-07-19, 22 posts)
- Episode 2: Leaked Alibaba Qwen 3.8 Max Shows Strong Benchmark and Coding Performance(2026-07-19, 6 posts)
- Episode 3: Qwen3.8-Max-Preview Rolls Out Across Web, PC and iOS Preview(2026-07-19, 2 posts)
- Episode 4: Qwen3.8-Max Preview Tested: Strong Coding but Slow Thinking(2026-07-19, 13 posts)
- Episode 5: Alibaba's Qwen3.8-Max-Preview iterates daily with improved frontend capabilities(2026-07-20, 5 posts)
- Episode 6: Alibaba Announces Open-Weight Qwen3.8 and Multiple New Updates(2026-07-23, 2 posts)
- Episode 7: Alibaba releases Qwen3.8-Max, open-sources weights next week(2026-08-03, 44 posts)
- Episode 8: Qwen3.8-Max Ranks Top Tier Across Benchmarks, Open-Source Narrows Gap(2026-08-03, 7 posts)
- Episode 9: Rumor: Alibaba's Qwen3.8-Max Outperforms Fable 5 and Set to Open Source(2026-08-03, 2 posts)
- Episode 10: Qwen3.8-Max Initial Tests Show Performance on Par with DeepSeek(2026-08-03, 2 posts)
- Episode 11: Alibaba's Qwen3.8-Max: Open-Source Model Nears Closed-Source Frontier(2026-08-03, 13 posts)
- Episode 12: Kimi K3 and Qwen3.8 Max Evaluations Approach Top Closed-Source Models(2026-08-05, 2 posts)
- Episode 13: Rumored Alibaba Qwen3.8-Max to Feature 2.4T Parameters with 95B Active(2026-08-05, 5 posts)
- Episode 14: Alibaba's Qwen3.8-Max tops agentic benchmark, open-sources next week(2026-08-06, 11 posts)
- Episode 15: Testing Alibaba's Qwen3.8-Max: Stellar Long-Context Agent Capabilities(2026-08-06, 3 posts)
- Episode 16: Alibaba's Qwen 3.8 Models Rumored for Next Week Release(2026-08-07, 2 posts)
- Episode 17: Rumors Claim Alibaba Released Qwen3.8-Max(2026-08-09, 2 posts)
Primary sources
- Qwen 3.8 Max Released, Praised as Cost-Effective; Cola App Lauds Its Capabilities — oran_ge · 2026-08-03
- Testing Qwen3.8-Max: Open Models Are Catching Up with Closed Frontier — dair_ai · 2026-08-04
- [source] Qwen 3.8 Max is said to lead image recognition and distill cleanly to 27B — teortaxesTex · 2026-08-04
- Qwen 3.8 Early Tests: Good at Agentic Coding, But Struggles with Long Loops — bindureddy · 2026-08-04
- [source] Qwen 3.8 Coding Test: Nearly Matches K3 at Half the Price — bindureddy · 2026-08-04
- Qwen 3.8 Max Tested: Rivals Frontier Models in Coding and Multimodal — WorldofAI · 2026-08-04
- Qwen 3 Preliminary Test: Near K3 Coding Performance at Half the Price — QuixiAI · 2026-08-04
- Developer tests Qwen3.8-max: A highly efficient and capable model — intellectronica · 2026-08-04
- Qwen 3.8 Max Tested: Solid Performance but Falls Short of Kimi K3 — ramos_casals · 2026-08-04
- Qwen3.8Max Deep-Dive: Top-Tier Coding & Agent Performance at a Fraction of the Cost — 葬AI · 2026-08-05
- Building a Beautiful Invoice Generator in 10 Minutes with Alibaba's Qwen3.8-Max — CodeByPoonam · 2026-08-05
- [source] Alibaba Releases Qwen3.8-Max: 2.4T Parameters, 95B Active, Natively Multimodal — CodeByPoonam · 2026-08-05
1 near-duplicate retellings: oran_ge