qwen3.8-flash is treated as text-only; alibabaStandard preset carries no modality metadata

[efwds]

Summary

qwen3.8-flash is classified as text-only by the client modality auto-detect, even though the
ModelStudio endpoint accepts image and video input. The practical effect: inline media never
reaches the model as pixels — read_file on an image/PDF is silently routed through the vision
bridge (a second, separately billed model call) and I receive a lossy text transcription instead.

Two distinct problems are involved:

1. MODALITY_PATTERNS in packages/core/src/core/modalityDefaults.ts has no rule for the
   qwen3.8-flash / qwen3.8-plus families (only qwen3.8-max), so they fall through to the
   catch-all [/^qwen/, {}].
2. The alibabaStandard provider preset carries no modalities metadata at all, so Standard
   API-key users have no path to correct capability data other than hand-editing settings.json.

Environment

• Qwen Code 0.22.2 (npm latest at time of writing — not a stale-install issue)
• Windows 11, security.auth.selectedType = "openai"
• Provider: alibabaStandard ("Standard API Key"), baseUrl = https://dashscope.aliyuncs.com/compatible-mode/v1
• Model: qwen3.8-flash (entered manually via /model; not present in the preset list)

Authoritative capability data

qianwen models info qwen3.8-flash --format json (CLI session auth, 2026-08-27) returns:

"modality": { "input": ["image", "text", "video"], "output": ["text"] },
"context": { "context_window": 1000000, "max_input": 991808, "max_output": 131072 }

So the model does accept image and video input. pdf and audio are not accepted.

Reproduction

1. Select qwen3.8-flash on the alibabaStandard provider and restart the session.
2. read_file any PNG.

Observed: the tool result is not an image but a bridge transcription block headed
[Untrusted machine transcription of 1 image(s) by qwen3.6-plus ... The image cannot be read by any tool]. The assistant receives second-hand text and an extra billed call is incurred per image.

Expected: the image is attached natively ([media attached in following user message]) and the
selected model sees the pixels.

Confirmed by experiment with a synthetic image containing the string ZEBRA 7749: before the
workaround only the transcription was available; after the workaround the model reported genuine
pixel-level detail (subpixel red/cyan fringing on the glyphs) that the transcription never conveyed.

Root cause

1. The auto-detect table is fail-closed and lags model releases (by design)

defaultModalities() documents the intent: "Unknown models default to text-only (empty object) to
avoid sending unsupported media types that would cause unrecoverable API errors." The table was
introduced by fix: add modality defaults to prevent API errors when reading PDFs (2026-02-27) and
has since grown purely by reactive per-model PRs:

Date	Commit
2026-04-08	add qwen3.6-plus
2026-05-15	add Qwen3.6-35B quant variants
2026-06-02	add MiniMax-M3
2026-06-08	add qwen3.7-plus multimodal
2026-06-20	add missing Token Plan models
2026-07-22	fix(core): add image modality support for qwen3.8-max and kimi-k3 models (#7491)
2026-08-04	resolve DashScope thinking-knob conflicts by family

The last edit to this file is 2026-08-04, whereas qwen3.8-flash's catalog entry was updated
2026-08-25. The 3.8 series entered the table incrementally (max only), and isTieredEffortWireModel()
shows the same shape — it also matches only qwen3.8-max. So flash/plus simply have not been
added yet. main still has no rule for them as of today.

2. Capability knowledge is duplicated across three lists that drift independently

List	Location	Carries modalities?	Covers 3.8 series
Regex fallback	core/src/core/modalityDefaults.ts	yes	max only
VISION_MODEL_PREFIX_PATTERNS	DashScopeOpenAICompatibleProvider	yes	no — stops at qwen3.7-plus
Provider presets	providers/presets/*.ts	depends on the preset	see below

3. alibabaStandard declares no modality metadata, unlike the plan presets

providers/presets/alibaba-token-plan.ts does carry it, e.g.:

{ id: "qwen3.8-max", contextWindowSize: 1e6, enableThinking: true,
  thinkingMandatory: true, modalities: { image: true, video: true } },

providers/presets/alibaba-standard.ts does not — its whole model list is six entries with only
contextWindowSize / enableThinking, no modalities, no 3.8-series model, and no
supportsModelDiscovery (both plan presets set that flag; Standard does not):

models: [
  { id: "qwen3.6-plus", contextWindowSize: 1e6, enableThinking: true },
  { id: "qwen3.7-plus", contextWindowSize: 1e6, enableThinking: true },
  { id: "qwen3.7-max",  contextWindowSize: 1e6, enableThinking: true },
  { id: "glm-5.1",      contextWindowSize: 202752, enableThinking: true },
  { id: "deepseek-v4-pro", contextWindowSize: 1e6, enableThinking: true },
  { id: "deepseek-v4-flash", contextWindowSize: 1e6 },
],

Consequently, on the Standard path even preset-known models get no capability metadata and fall
through to the regex table. And because qwen3.8-flash is typed in manually, buildModelConfigs()
takes the inputs.modelIds branch for an id absent from the preset spec map, so the emitted
generationConfig comes only from whatever the user filled in the advanced steps — which is nothing
by default.

Workaround (and why it is fragile)

Declaring the modalities explicitly on the provider entry fixes it, and the correct persisted key is
modalities — not multimodal, which is only the interactive /model wizard's input shape and
is silently ignored when written into settings.json (this cost me one wasted edit; a schema
warning for unknown generationConfig keys would help):

{
  "id": "qwen3.8-flash",
  "baseUrl": "https://dashscope.aliyuncs.com/compatible-mode/v1",
  "envKey": "DASHSCOPE_API_KEY",
  "generationConfig": {
    "modalities": { "image": true, "video": true },
    "contextWindowSize": 1000000
  }
}

However this setting is not sticky within a session: shouldUpdateModelDerivedDefault() treats
kind === "modelProviders" as overridable, so applyRawModelDerivedDefaults() re-runs
defaultModalities() on a raw /model switch and silently reverts image support to {} until the
process restarts. Arguably a separate bug: a value the user explicitly pinned in settings.json
should not be treated as a computed default.

Suggested fixes

1. Add qwen3.8-flash ({ image: true, video: true }) and a qwen3.8-plus rule to
   MODALITY_PATTERNS, before the /^qwen/ catch-all.
2. Give the alibabaStandard preset the same modalities metadata the Token Plan preset already has,
   and consider supportsModelDiscovery for it.
3. Longer term, derive per-model modality from an authoritative catalog instead of hand-copied lists —
   the data already exists server-side (modality.input), and the three client-side lists above are
   demonstrably out of sync with each other.
4. Treat user-authored settings.json generationConfig.modalities as a pinned source so raw model
   switches do not silently discard it.
5. Warn (or fail) on unknown keys inside generationConfig, to catch the multimodal/modalities
   mix-up.

Related

• fix(providers): sync Token Plan model list with the Bailian catalog #8651 — fix(providers): sync Token Plan model list with the Bailian catalog (open). Its wording
  describes exactly this failure mode for qwen3.8-max: "anyone reaching the model without this
  preset entry silently loses video".
• bug(auth): Bailian Personal Token Plan models are out of sync and image/video generation fails #8432 — Token Plan model list out of sync; image/video generation fails.
• #7491 — added image modality for qwen3.8-max only.

I found no existing report for qwen3.8-flash / qwen3.8-plus specifically.

[qwen-code-ci-bot]

✅ Qwen Triage finished — view run. See the stage comments in this thread for the result.

✅ Qwen Triage 已完成 —— 查看运行。结果见本线程中的各阶段评论。

[qwen-code-ci-bot]

Triage

• Type: bug — observed behavior with a solid reproduction (catalog evidence, before/after experiment)
• Labels: type/bug, category/core, scope/core, priority/P2
• Next: confirm root cause in current source, then append findings

qwen3.8-flash / qwen3.8-plus hit the text-only catch-all in the client modality table, so inline images are routed through the vision bridge instead of reaching the model as pixels (lossy transcription + an extra billed call per image). The alibabaStandard preset also carries no modality metadata at all. No duplicate found — #8651 / #8432 / #7491 are related but not this case.

中文说明

• 类型：bug —— 已观测到、复现充分（有官方模型能力数据与 before/after 实验）
• 标签：type/bug、category/core、scope/core、priority/P2
• 下一步：在当前源码中确认根因，并追加结论

qwen3.8-flash / qwen3.8-plus 命中客户端模态表的 text-only 兜底规则，导致内联图片被转给 vision bridge，而不是以像素形式送给模型（有损转写，且每张图片额外计费一次）。alibabaStandard preset 也完全没有携带模态元数据。未发现重复 —— #8651 / #8432 / #7491 相关但不是同一问题。

Verification

Reproduction not executed (unattended CI run — static review only). Root cause confirmed by reading current main:

1. packages/core/src/core/modalityDefaults.ts — MODALITY_PATTERNS has rules for qwen3.5-plus / qwen3.6-plus / qwen3.7-plus (image + video) and qwen3.8-max (image only). qwen3.8-flash / qwen3.8-plus fall through to the [/^qwen/, {}] catch-all → text-only. isTieredEffortWireModel() shows the same max-only shape, as the report says.
2. packages/core/src/providers/presets/alibaba-standard.ts — entries carry only contextWindowSize / enableThinking; no modalities, no 3.8-series model, no supportsModelDiscovery (the Token Plan presets carry all three).
3. DashScopeOpenAICompatibleProvider.VISION_MODEL_PREFIX_PATTERNS stops at qwen3.7-plus — the third independently-drifting list, exactly as described.
4. Stickiness sub-bug confirmed in code: modelsConfig.ts → shouldUpdateModelDerivedDefault() treats kind === 'modelProviders' as overridable, so a raw /model switch re-runs defaultModalities() and silently discards modalities the user pinned in settings.json.
5. Impact chain verified: fileUtils.ts gates native image attachment on modalities.image; when false, vision-bridge-service.ts produces the [Untrusted machine transcription …] block the reporter observed.

Fix direction

• Add qwen3.8-flash / qwen3.8-plus → { image: true, video: true } to MODALITY_PATTERNS before the /^qwen/ catch-all (the reporter's catalog dump shows image+video input, no pdf/audio).
• Bring the alibabaStandard preset to parity with the Token Plan presets (modality metadata, 3.8-series entries, discovery flag).
• Suggestions Are you interested in AI Terminal? #4 (pinned settings.json modalities reverted on model switch) and TypeError in Authentication Selection Interface #5 (warn on unknown generationConfig keys) are independent defects — worth splitting into separate issues rather than bundling here.
• Related drift in the same table: qwen3.8-max is image-only in the regex table but image + video in the Token Plan preset; open PR fix(providers): sync Token Plan model list with the Bailian catalog #8651 covers the preset side.

中文说明（验证结果）

未在 CI 中实际执行复现（无人值守运行，仅静态审查）。根因已在当前 main 源码中确认：

1. modalityDefaults.ts 的 MODALITY_PATTERNS 只有 3.5/3.6/3.7-plus（image+video）和 qwen3.8-max（仅 image）规则；qwen3.8-flash / qwen3.8-plus 落入 [/^qwen/, {}] 兜底 → text-only。isTieredEffortWireModel() 同样只匹配 max。
2. alibaba-standard.ts preset 的模型项只有 contextWindowSize / enableThinking，没有 modalities、没有 3.8 系列、也没有 supportsModelDiscovery（Token Plan preset 三者都有）。
3. DashScopeOpenAICompatibleProvider.VISION_MODEL_PREFIX_PATTERNS 到 qwen3.7-plus 为止 —— 确实是第三份各自漂移的清单。
4. 会话内不生效的子 bug 在代码层面成立：shouldUpdateModelDerivedDefault() 把 kind === 'modelProviders' 视为可覆盖，raw /model 切换会重跑 defaultModalities()，悄悄丢弃用户在 settings.json 里显式固定的模态。
5. 影响链已核实：fileUtils.ts 以 modalities.image 决定是否原生命附图；为 false 时走 vision-bridge-service.ts，产生报告者看到的转写块。

修复方向

• 在 /^qwen/ 兜底之前为 qwen3.8-flash / qwen3.8-plus 增加 { image: true, video: true } 规则（能力数据显示输入为 image+video，不含 pdf/audio）。
• 让 alibabaStandard preset 与 Token Plan preset 对齐（模态元数据、3.8 系列条目、模型发现开关）。
• 建议 Are you interested in AI Terminal? #4（固定模态被模型切换覆盖）与 TypeError in Authentication Selection Interface #5（generationConfig 未知键告警）是相互独立的问题，建议拆分为单独 issue，不与本问题捆绑。
• 同表中的相关漂移：qwen3.8-max 在正则表中仅有 image，而 Token Plan preset 为 image+video；open PR fix(providers): sync Token Plan model list with the Bailian catalog #8651 覆盖 preset 一侧。

--- Qwen Code

[yiliang114]

Reproduced: defaultModalities('qwen3.8-flash') returns {} (text-only) on main because MODALITY_PATTERNS has no entry for qwen3.8-flash/qwen3.8-plus. Fix PR: #10195