feat(compute): add DeepSeek V4 variants and Qwen3.8 Flash (#82)

- add DeepSeek V4 Flash Vision Exp provider and model specifications - add DeepSeek V4 Flash 0731 and Qwen3.8 Flash model specifications - add regression coverage and bump preset data version
2026-09-05 19:13:33 +08:00 · 2026-08-27 17:43:36 +08:00
parent c46cf07d2f
commit d0e9bf5085
5 changed files with 275 additions and 8 deletions
--- a/tests/validate.test.mjs
+++ b/tests/validate.test.mjs
@@ -105,7 +105,7 @@ describe('真实数据全量校验', () => {
        if (spec.routing) routed.push(spec)
      }
    }
-    assert.equal(routed.length, 43)
+    assert.equal(routed.length, 46)
    assert.equal(routed.every((spec) => spec.routing.reasoning.supportedModes.includes(spec.routing.reasoning.defaultMode)), true)
    assert.equal(routed.every((spec) => Array.isArray(spec.spec.capabilities)), true)
  })
@@ -234,6 +234,37 @@ describe('真实数据全量校验', () => {
    }
  })

+  it('DeepSeek V4 Flash Vision Exp 应提供独立的视觉规格', () => {
+    const provider = JSON.parse(readFileSync(join(ROOT, 'compute', 'providers', 'deepseek.json'), 'utf8'))
+    const specFile = JSON.parse(readFileSync(join(ROOT, 'compute', 'model-specs', 'deepseek.json'), 'utf8'))
+    const modelId = 'deepseek-v4-flash-vision-exp'
+    const model = provider.models.find((item) => item.modelName === modelId)
+
+    assert.ok(provider.services.includes('vision'))
+    assert.ok(model, `provider 缺少 ${modelId}`)
+    assert.equal(model.contextWindow, 1000000)
+    assert.equal(model.maxOutputTokens, 384000)
+    assert.ok(model.capabilities.includes('vision'))
+    assert.ok(model.serviceType.includes('vision'))
+    assert.equal(model.extra.supportsThinking, true)
+    assert.equal(model.extra.thinkingDefault, true)
+
+    const specs = specFile.specs.filter((item) => item.id === modelId)
+    assert.equal(specs.length, 1, `model-specs 中 ${modelId} 应且仅应有一条规格`)
+    assert.deepEqual(specs[0].match.exact, [modelId])
+    assert.equal('patterns' in specs[0].match, false)
+    assert.equal(specs[0].family, modelId)
+    assert.equal(specs[0].spec.contextWindow, 1000000)
+    assert.equal(specs[0].spec.maxOutputTokens, 384000)
+    assert.equal(specs[0].spec.supportsReasoning, true)
+    assert.ok(specs[0].spec.capabilities.includes('vision'))
+    assert.ok(specs[0].spec.serviceType.includes('vision'))
+
+    const flash = specFile.specs.find((item) => item.id === 'deepseek-v4-flash')
+    assert.deepEqual(flash.match.exact, ['deepseek-v4-flash'])
+    assert.equal('patterns' in flash.match, false)
+  })
+
  it('MiMo V2.5 ASR 应有独立精确规格，避免回落到 MiMo V2.5 family', () => {
    const specFile = JSON.parse(readFileSync(join(ROOT, 'compute', 'model-specs', 'xiaomi.json'), 'utf8'))
    const specs = specFile.specs.filter((item) => item.id === 'mimo-v2.5-asr')
@@ -324,6 +355,29 @@ describe('真实数据全量校验', () => {
    assert.ok(specs[0].spec.capabilities.includes('vision'))
  })

+  it('Qwen3.8 Flash 应提供完整的多模态推理规格', () => {
+    const specFile = JSON.parse(readFileSync(join(ROOT, 'compute', 'model-specs', 'qwen.json'), 'utf8'))
+    const modelId = 'qwen3.8-flash'
+    const specs = specFile.specs.filter((item) => item.id === modelId)
+
+    assert.equal(specs.length, 1, `model-specs 中 ${modelId} 应且仅应有一条规格`)
+    const modelSpec = specs[0]
+    assert.deepEqual(modelSpec.match.exact, [modelId])
+    assert.equal('patterns' in modelSpec.match, false)
+    assert.equal(modelSpec.family, modelId)
+    assert.equal(modelSpec.spec.contextWindow, 1000000)
+    assert.equal(modelSpec.spec.maxOutputTokens, 131072)
+    assert.equal(modelSpec.spec.supportsReasoning, true)
+    assert.equal(modelSpec.spec.releasedAt, '2026-08-26')
+    assert.ok(modelSpec.spec.capabilities.includes('vision'))
+    assert.ok(modelSpec.spec.capabilities.includes('video_understanding'))
+    assert.ok(modelSpec.spec.capabilities.includes('fast'))
+    assert.deepEqual(modelSpec.spec.serviceType, ['chat', 'reasoning', 'vision'])
+    assert.equal(modelSpec.spec.extra.thinkingMaxTokens, 262144)
+    assert.equal(modelSpec.routing.tier, 'lightweight')
+    assert.equal(modelSpec.routing.reasoning.defaultMode, 'xhigh')
+  })
+
  it('基础模型与后缀版本应使用独立精确规格，避免互相误匹配', () => {
    const qwen = JSON.parse(readFileSync(join(ROOT, 'compute', 'model-specs', 'qwen.json'), 'utf8'))
    const deepseek = JSON.parse(readFileSync(join(ROOT, 'compute', 'model-specs', 'deepseek.json'), 'utf8'))
@@ -333,12 +387,16 @@ describe('真实数据全量校验', () => {
    const qwenPreview = specFor(qwen, 'qwen3.8-max-preview')
    const deepseekBase = specFor(deepseek, 'deepseek-v4-pro')
    const deepseek0813 = specFor(deepseek, 'deepseek-v4-pro-0813')
+    const deepseekFlash = specFor(deepseek, 'deepseek-v4-flash')
+    const deepseekFlash0731 = specFor(deepseek, 'deepseek-v4-flash-0731')

    for (const [id, modelSpec] of [
      ['qwen3.8-max', qwenBase],
      ['qwen3.8-max-preview', qwenPreview],
      ['deepseek-v4-pro', deepseekBase],
      ['deepseek-v4-pro-0813', deepseek0813],
+      ['deepseek-v4-flash', deepseekFlash],
+      ['deepseek-v4-flash-0731', deepseekFlash0731],
    ]) {
      assert.ok(modelSpec, `model-specs 缺少 ${id}`)
      assert.deepEqual(modelSpec.match.exact, [id])
@@ -347,10 +405,14 @@ describe('真实数据全量校验', () => {

    assert.equal('patterns' in deepseekBase.match, false)
    assert.equal('patterns' in deepseek0813.match, false)
+    assert.equal('patterns' in deepseekFlash.match, false)
+    assert.equal('patterns' in deepseekFlash0731.match, false)
    assert.equal(qwenBase.spec.contextWindow, qwenPreview.spec.contextWindow)
    assert.equal(qwenBase.spec.maxOutputTokens, qwenPreview.spec.maxOutputTokens)
    assert.equal(deepseekBase.spec.contextWindow, deepseek0813.spec.contextWindow)
    assert.equal(deepseekBase.spec.maxOutputTokens, deepseek0813.spec.maxOutputTokens)
+    assert.equal(deepseekFlash.spec.contextWindow, deepseekFlash0731.spec.contextWindow)
+    assert.equal(deepseekFlash.spec.maxOutputTokens, deepseekFlash0731.spec.maxOutputTokens)
  })

  it('所有 Provider 应按供应商归属计价，模型来源不覆盖供应商币种', () => {
--- a/compute/model-specs/deepseek.json
+++ b/compute/model-specs/deepseek.json
@@ -162,13 +162,10 @@
    {
      "id": "deepseek-v4-flash",
      "displayName": "DeepSeek V4 Flash",
-      "family": "deepseek-v4",
+      "family": "deepseek-v4-flash",
      "match": {
        "exact": [
          "deepseek-v4-flash"
-        ],
-        "patterns": [
-          "deepseek-v4-flash*"
        ]
      },
      "spec": {
@@ -211,6 +208,111 @@
          "defaultMode": "high"
        }
      }
+    },
+    {
+      "id": "deepseek-v4-flash-0731",
+      "displayName": "DeepSeek V4 Flash 0731",
+      "family": "deepseek-v4-flash-0731",
+      "match": {
+        "exact": [
+          "deepseek-v4-flash-0731"
+        ]
+      },
+      "spec": {
+        "contextWindow": 1000000,
+        "maxOutputTokens": 384000,
+        "capabilities": [
+          "chat",
+          "code",
+          "reasoning",
+          "deep_thinking",
+          "multilingual",
+          "tool_use"
+        ],
+        "serviceType": [
+          "chat",
+          "reasoning"
+        ],
+        "defaultTemperature": 1,
+        "supportsReasoning": true,
+        "extra": {
+          "supportsThinking": true,
+          "thinkingDefault": true,
+          "reasoningEffort": [
+            "high",
+            "max"
+          ]
+        },
+        "description": "DeepSeek V4 Flash 0731 固定版本，100 万上下文"
+      },
+      "routing": {
+        "tier": "lightweight",
+        "routingPriority": 35,
+        "eligibleForAgent": true,
+        "defaultReference": false,
+        "reasoning": {
+          "supportedModes": [
+            "auto",
+            "high",
+            "max"
+          ],
+          "defaultMode": "high"
+        }
+      }
+    },
+    {
+      "id": "deepseek-v4-flash-vision-exp",
+      "displayName": "DeepSeek V4 Flash Vision Exp",
+      "family": "deepseek-v4-flash-vision-exp",
+      "match": {
+        "exact": [
+          "deepseek-v4-flash-vision-exp"
+        ]
+      },
+      "spec": {
+        "contextWindow": 1000000,
+        "maxOutputTokens": 384000,
+        "capabilities": [
+          "chat",
+          "code",
+          "reasoning",
+          "deep_thinking",
+          "vision",
+          "multilingual",
+          "tool_use"
+        ],
+        "serviceType": [
+          "chat",
+          "reasoning",
+          "vision"
+        ],
+        "defaultTemperature": 1,
+        "supportsReasoning": true,
+        "releasedAt": "2026-08-21",
+        "extra": {
+          "supportsThinking": true,
+          "thinkingDefault": true,
+          "reasoningEffort": [
+            "high",
+            "max"
+          ]
+        },
+        "description": "DeepSeek V4 Flash 实验性视觉版本，支持文本与图像输入，文本能力与 V4 Flash 一致"
+      },
+      "routing": {
+        "tier": "lightweight",
+        "routingPriority": 35,
+        "eligibleForAgent": true,
+        "defaultReference": false,
+        "reasoning": {
+          "supportedModes": [
+            "auto",
+            "high",
+            "max"
+          ],
+          "defaultMode": "high"
+        }
+      }
    }
  ]
 }
--- a/compute/model-specs/qwen.json
+++ b/compute/model-specs/qwen.json
@@ -50,6 +50,66 @@
        }
      }
    },
+    {
+      "id": "qwen3.8-flash",
+      "displayName": "Qwen3.8 Flash",
+      "family": "qwen3.8-flash",
+      "match": {
+        "exact": [
+          "qwen3.8-flash"
+        ]
+      },
+      "spec": {
+        "contextWindow": 1000000,
+        "maxOutputTokens": 131072,
+        "capabilities": [
+          "chat",
+          "reasoning",
+          "deep_thinking",
+          "code",
+          "multilingual",
+          "tool_use",
+          "long_context",
+          "agent",
+          "vision",
+          "video_understanding",
+          "fast"
+        ],
+        "serviceType": [
+          "chat",
+          "reasoning",
+          "vision"
+        ],
+        "defaultTemperature": 0.6,
+        "defaultTopP": 0.95,
+        "supportsReasoning": true,
+        "extra": {
+          "thinkingDefault": true,
+          "thinkingMaxTokens": 262144
+        },
+        "releasedAt": "2026-08-26",
+        "description": "通义千问 Qwen3.8 Flash，高性价比多模态推理模型，支持 100 万上下文及文本、图像、视频输入"
+      },
+      "routing": {
+        "tier": "lightweight",
+        "routingPriority": 40,
+        "eligibleForAgent": true,
+        "defaultReference": false,
+        "reasoning": {
+          "supportedModes": [
+            "auto",
+            "off",
+            "minimal",
+            "low",
+            "medium",
+            "high",
+            "xhigh",
+            "max"
+          ],
+          "defaultMode": "xhigh"
+        }
+      }
+    },
    {
      "id": "qwen3.8-max-preview",
      "displayName": "Qwen3.8 Max Preview",
--- a/compute/providers/deepseek.json
+++ b/compute/providers/deepseek.json
@@ -11,13 +11,56 @@
  "priceCurrency": "CNY",
  "services": [
    "chat",
-    "reasoning"
+    "reasoning",
+    "vision"
  ],
  "tombstones": [
    "deepseek-chat",
    "deepseek-reasoner"
  ],
  "models": [
+    {
+      "modelName": "deepseek-v4-flash-vision-exp",
+      "displayName": "DeepSeek V4 Flash Vision Exp",
+      "serviceType": [
+        "chat",
+        "reasoning",
+        "vision"
+      ],
+      "description": "DeepSeek V4 Flash 实验性视觉版本，支持文本与图像输入，文本能力与 V4 Flash 一致，1M 上下文，并发上限 2500",
+      "contextWindow": 1000000,
+      "maxOutputTokens": 384000,
+      "capabilities": [
+        "chat",
+        "code",
+        "reasoning",
+        "deep_thinking",
+        "vision",
+        "long_context",
+        "multilingual",
+        "tool_use"
+      ],
+      "inputPrice": 1.008,
+      "outputPrice": 2.016,
+      "defaultTemperature": 1,
+      "defaultTopP": 1,
+      "extra": {
+        "cacheHitPrice": 0.02016,
+        "concurrencyLimit": 2500,
+        "supportsThinking": true,
+        "thinkingDefault": true,
+        "reasoningEffort": [
+          "high",
+          "max"
+        ],
+        "features": {
+          "jsonOutput": true,
+          "toolCalls": true,
+          "chatPrefixCompletion": "beta",
+          "fimCompletion": false
+        }
+      }
+    },
    {
      "modelName": "deepseek-v4-flash",
      "displayName": "DeepSeek V4 Flash",
--- a/manifest.json
+++ b/manifest.json
@@ -1,6 +1,6 @@
 {
  "version": "1.0.0",
-  "presetDataVersion": 95,
-  "updatedAt": "2026-08-26",
+  "presetDataVersion": 98,
+  "updatedAt": "2026-08-27",
  "description": "DesireCore 官方配置中心"
 }