DeepSeek V4.1 Flash bug drops answers, leaks reasoning into visible text
logser13 · ghdev · 2026-09-18
An OpenCode user filed an issue reporting two user-visible defects in DeepSeek V4.1 Flash (released 2026-09-10).
- Symptom A — lost answers: via OpenCode Go, assistant messages finish with finish: "stop" but tokens.output = 0, with the entire reply stored only in the reasoning part and no text part, so the UI stops after tool calls with no answer. Observed 5 times out of 1715 assistant messages (Sept 17–18), with message IDs and token logs as evidence.
- Symptom B — reasoning leaks: via the direct DeepSeek API, tokens.reasoning = 0 (reasoning not separated) and the model's English planning lines ("No new messages...", "Let me report concisely") get concatenated into the visible text before the actual answer.
The author notes model metadata declares reasoning: true with multiple effort levels, that older DeepSeek V4 models show no such behavior across thousands of messages, and requests that answers always be emitted as content with reasoning consistently separated into reasoningcontent.
More from Models
- Skeptical deep dive confirms Humanity's Last Exam errors; official o3-mini grader marked right answers wrong every time — paul_cal · 2026-09-20
- Matt Shumer asks if Jev could help with scalable oversight and alignment checks — mattshumer_ · 2026-09-20
- FrontierSWE v2 opens 24.1-point gap: Claude Fable 5.1 scores 56.29% vs GPT-5.6's 32.2% — geoffwolfe · 2026-09-20
- 22M local model beats JEV 93% vs 80% on Banking77 in 8ms on CPU — Prompt Engineering · 2026-09-20
- Jev loses to Gemini on 1,565-email classification benchmark, but dev still wants it in production — socialwithaayan · 2026-09-20
- Jev Detector scans ~10,000 words for AI slop in ~2 seconds, free with no sign-up — socialwithaayan · 2026-09-20