Reported by @debikurnia (2026-09-22).
Model: mimo-v2.6-flash (also mimo-v2.6-pro); mimo-v2.5 is unaffected.
API: OpenAI-compatible POST /v1/chat/completions, images as image_url data URLs, temperature=0.
Gateways: OpenCode Go (https://opencode.ai/zen/go/v1) and OpenRouter (xiaomi/mimo-v2.6-flash) — identical failure on both.
In ONE conversation, send one new image per turn. With 1–4 images the model reads correctly; from the 5th image on, it answers with the content of an earlier image (HTTP 200, no error), while its reasoning_content claims it is reading "the latest attached image".
| Images in conversation | Expected | Actual |
|---|---|---|
3 (A,E,F) |
PASS | PASS |
4 (A,E,F,H) |
PASS | PASS |
5 (A,E,F,H,K) |
K | H (4th image) |
8 (A,E,F,H,K,L,T,X) |
all correct | A,E,F,H,H,H,T,T — deterministic, 3/3 runs (opencode-go), 2/2 (OpenRouter) |
| Control | Result |
|---|---|
| Fresh routing session every turn | PASS |
| History kept but prior images replaced by text | PASS |
mimo-v2.5 (8 images) |
PASS |
kimi-k3 (8 images, non-MiMo) |
PASS |
mimo-v2.6-pro |
FAIL |
Requires Node ≥ 18. No dependencies — the script generates its own images.
OPENCODE_API_KEY=... node repro.mjs --provider opencode-go --rounds 3
OPENROUTER_API_KEY=... node repro.mjs --provider openrouterObserved usage: prompt_tokens grows normally (~270 per image) while prompt_tokens_details.cached_tokens appears/grows around the first misread.
repro.mjs— standalone reproducerevidence-01…evidence-08— raw per-turncontent,reasoning_content,usage
Full raw logs also show the verbatim model reasoning at the first bad turn (expected K, answered H):
"The user wants me to identify a single capital letter from the latest attached image. I will look at the most recent image provided in the conversation. The image displays a black capital letter "H" on a white background. The prompt asks for the single letter only. I will reply with "H"."