Qwen3.8 Flash

49 comments across HackerNews, Lemmy and ProductHunt, heavily skewed toward local-LLM developers; several high-upvote HN threads are only tangential (LLM economics, inference pricing, epistemology). The dominant hands-on complaint is 'overthinking': one HN user reports qwen3.8:27b 'overthinks to the magnitude of ~10x the tokens vs a ~4x faster gemma4:26b-a3b' and 'will churn over a prompt often for 5-10 minutes while gemma4 regularly finishes the same prompt in under 20 seconds' on general-purpose use. Another user is more forgiving, saying 3.8's verbosity 'really depends on the task' - verbose on open-ended prompts, 'as succinct as Muse Glimmer when it has a clear path forward.' Lemmy threads focus on the Flash-Next variant (125B total / 6B active MoE, plus 51B n-gram embedding and 4B MTP): users estimate it 'could run in Q4 on 96GB RAM and 16GB VRAM,' report benchmark scores 'beating 27b and DeepSeek Flash,' but flag ecosystem friction - llama.cpp main-branch support absent as of 2026-08-26 (PR pending; Unsloth's fork works), and Unsloth's GGUF page warns 'you will need at least 75 GB of RAM or unified memory,' with the caveat that '1-bit isn't really 1-bit at all' due to the architecture, which conversely 'allows the model to retain more of its original accuracy.' One Lemmy user with 64GB RAM + 32GB VRAM calls the large size 'a bummer' on Intel graphics requiring mirrored VRAM; another notes low active params mean 'you don't need much VRAM to get decent speeds' via --n-cpu-moe. ProductHunt commentary is mostly descriptive of the Qwen3 family (dense 0.6B-32B, MoE 30B/235B, Hybrid Thinking Mode) plus one explicit access complaint: 'Every time I want to use it, opening Alibaba Cloud just makes me feel discouraged.' A separate HN user praises Qwen-35BA3B (different variant) for surfacing institutional knowledge from siloed docs.

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