Qwen3.8 Omni Flash

23 comments (14 HackerNews, 8 ProductHunt, 1 YouTube). Direct product signal is thinner than the count suggests — much of the HN thread debates China's energy grid and model-selection tooling, not the model itself. What users actually report: (1) Reasoning quirks — a +346 HN commenter finds the new architecture's reasoning behaves 'really weirdly compared to any model I've ever seen,' leaking internal monologue between tool calls ('The user's message is just system instructions setup with no actual task'); another HN user saw similar subagent-response bleeding on a different stealth model and links a LessWrong reward-hacking explainer as a possible cause. (2) Capability vs hardware — an HN user running it locally calls it 'very capable yes but very slow': the 27B is 'relatively easy to run,' but the 125B needs ~128GB of DDR5 (~$3,000 setup) to exceed 40 t/s, with DDR4 capped at 20 t/s and an RTX 4090 topping out near 30 t/s; Apple Silicon is called 'a good option.' (3) Cost data — HN cites ArtificialAnalysis: Qwen3.8 Max (0902) at $2/$6 per 1M tokens but $5.41 per task at intelligence index 45, vs GPT-6 Astra xhigh at $2.31/task with index 53; the same commenter notes Qwen 3.8 Omni Flash 'isn't yet' on the index. (4) Ecosystem wins — one HN user replaced a discontinued Gemini 2.5 Flash Lite production image pipeline with a fine-tuned Qwen VL 3B in ~20 minutes via Claude Code + HF CLI. (5) Access friction — a ProductHunt commenter asks whether Qwen has API channels besides Alibaba Cloud, because opening Alibaba Cloud 'makes me feel discouraged.' (6) ProductHunt sentiment is enthusiastic but low-substance: 'incredibly impressive... smooth, human-like responses,' Hybrid Thinking Mode called a potential 'game-changer,' the 4B dense model reportedly close to much larger previous-gen models, and repeated 'big win for open source' framing. Several HN users also express general hesitancy about China-hosted inference (nighttime solar availability, coal expansion), and one says he simply stays on a cheaper incumbent (GLM 5.3 flash) rather than churn through the 'vast sea of models.'

No aggregate ratings were found for this product during the last harvest.