Community Tests Defend Qwen 3.8 27B's Heavy Reasoning Token Use

Qwen 3.8 27B was accused of "overthinking" for generating a huge number of reasoning tokens (measured at over 16K in some tests), but on 08-17 several community members pushed back with hands-on tests, arguing that this deep thinking is the necessary cost of performance approaching much larger models—while other evaluations point out its weaknesses in logic and visual reasoning.

Confirmed

• @sukazu, after hands-on comparisons with Chinese models like GLM 5.3 and DeepSeek V4 Flash/Pro, concluded that Qwen 3.8 27B isn't truly "overthinking," even though it uses more reasoning tokens than version 3.6.
• @maxwell321 ran the model locally on 3×3090 + Tesla P40 (Unsloth UD-Q8KXL quantization) and concluded that extensive "thinking" digs out fine-grained details from training knowledge, bringing performance close to Claude Sonnet.
• @AltruisticHeat9531 argued that 16K+ reasoning tokens are a necessary sacrifice to rival trillion-parameter models, citing Karpathy's "tokens to think" theory along with SWE-Rebench comparison data.
• @ResearchCrafty1804 noted that compute usage in closed-source models (e.g., Opus 5, GPT-5.6 Sol) is opaque—vendors can absorb costs server-side and look efficient—whereas open-source models' reasoning is measurable, so directly comparing reported token counts is unfair.
• @Afinetheorem's review called it one of the best local LLMs runnable on high-end consumer GPUs, with excellent coding benchmarks, but among the worst tested for logic and visual reasoning, far behind SOTA models (46/58).

Why it matters

• The discussion highlights the asymmetry in evaluating open- vs. closed-source models: open-source reasoning costs can be directly measured and criticized, while closed-source models can hide compute costs behind their services.
• The debate over whether more reasoning tokens equals inefficiency offers guidance for local deployment users: for code-heavy tasks, Qwen 3.8 27B is cost-effective; for logic and visual tasks, other models are better choices.

2026-08-17 ~ 2026-08-17 · 5 related posts

• [source] Local test: Qwen 3.8 27B's overthinking brings it near Sonnet-level performance — maxwell321 · 2026-08-17
• Why Qwen 3.8 27B Isn't Overthinking: Compared with GLM and DeepSeek — sukazu · 2026-08-17
• [source] Qwen3.8-27B Defended: More Reasoning Tokens Isn't 'Overthinking' — ResearchCrafty1804 · 2026-08-17
• [source] Qwen 3.8 27b shines in code but fails logic tests, highlighting local LLM trade-offs — Afinetheorem · 2026-08-17
• Qwen 3.8's High Token Cost is a Fair Trade for Performance — Altruistic_Heat_9531 · 2026-08-17