DeepSeek V4 Flash Vision Exp

Dataset caveat first: of the 25 top-upvoted HackerNews comments shown, roughly half belong to a thread about a DIFFERENT product (a 28MB 'Cactus Needle' micro-LLM for tool calls / Home Assistant), so signal specific to DeepSeek V4 Flash Vision Exp comes from only ~8 comments. Those on-topic reports split sharply. Positive: one heavy user running it via OpenRouter as a daily driver for a week reports 'never experienced incoherent rubbish even at 500k+ contexts' and tool-calling reliability 'better than Opus 5 in the Claude Code harness.' Vision is seen as filling a real gap: 'DS being unable to precisely view Playwright screenshots is the only thing I really miss from Sonnet. This is promising.' Negative: another user 'wasting hours' trying to get V4 Flash (with Qwen 3.8 27B as vision agent) to read sheet music for a Terminal Bench 3 task reports 'none of it was working' — then self-diagnoses that he was 'testing the harness' on a task he would never actually use. A clock-reading test failed for one user (answered 5:10) yet succeeded 9 out of 10 fresh sessions for another ('Lesson learned to double check what I read on the internet'), suggesting stochastic rather than deterministic vision quality. Users also flag mechanics: images are tokenized by dimensions, auto-resized (small images upscaled toward ~384x384, larger ones downscaled), and billed together with text tokens. A quoted investor-meeting transcript confirms multimodal was planned natively for 'V4 and subsequent iterations.' Off-topic philosophical comments (surveillance, AGI) carry no product signal.

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