From bcf7b286ca233a9e79603c35b2b3bee227e68729 Mon Sep 17 00:00:00 2001
From: MicroHEROX <140800253+MicroHEROX@users.noreply.github.com>
Date: Fri, 4 Sep 2026 00:20:28 +0800
Subject: [PATCH] fix(amd): Qwen3.8-Flash-Next accepts image input on the AMD
 host

---
 providers/amd/models/Qwen3.8-Flash-Next.toml | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/providers/amd/models/Qwen3.8-Flash-Next.toml b/providers/amd/models/Qwen3.8-Flash-Next.toml
index 290a819bfdd..3f2fb729032 100644
--- a/providers/amd/models/Qwen3.8-Flash-Next.toml
+++ b/providers/amd/models/Qwen3.8-Flash-Next.toml
@@ -1,16 +1,17 @@
-# Sources (accessed 2026-08-28):
+# Sources (accessed 2026-09-03):
 # https://developer.amd.com.cn/radeon/tokenfactory
 # Pricing from the Token Factory models API: input $0.15, output $0.47,
 # cache read $0.016 per MTok.
-# AMD hosts a text-only vllm-router deployment (input/output modalities are
-# text only), so the multimodal base entry is overridden below.
+# The AMD deployment accepts image input (GET /v1/models reports text+image
+# modalities) and image understanding was verified live: a rendered-digit
+# test image was read back correctly. The multimodal base entry is therefore
+# overridden to text+image (no video on this host).
 # Reasoning control verified live against POST /v1/chat/completions:
 # reasoning_effort is a validated enum low|medium|xhigh with xhigh the default
 # (values outside the enum return 400; no none, no toggle). No
 # reasoning_content is ever returned (stream + non-stream), so there is no
 # toggle and no interleaved channel on this host.
 base_model = "alibaba/qwen3.8-flash-next"
-attachment = false
 reasoning_options = [
   { type = "effort", values = ["low", "medium", "xhigh"] },
 ]
@@ -21,5 +22,5 @@ output = 0.47
 cache_read = 0.016
 
 [modalities]
-input = ["text"]
+input = ["text", "image"]
 output = ["text"]