@turmind/halo 1.5.11 → 1.6.1

package/templates/models/aws-bedrock-openai.yaml

@@ -1,10 +1,10 @@
 id: aws-bedrock-openai
 runtime: bedrock-mantle
-revision: 2026100501
-displayName: AWS Bedrock OpenAI-compatible (GPT-6.1 / 6 / 5.6 · Grok 4.6 · Kimi K3)
-displayName_zh: AWS Bedrock OpenAI 兼容接口（GPT-6.1 / 6 / 5.6 · Grok 4.6 · Kimi K3）
-description: Frontier models on Amazon Bedrock via the bedrock-runtime endpoint's OpenAI Responses API (/openai/v1) — OpenAI GPT-6.1 Sol, GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna), xAI Grok 4.6, Moonshot Kimi K3. Cross-Region inference profiles (global.*) only — every region serves every model, no region split. No server-side web_search (that is Mantle-only).
-description_zh: 通过 bedrock-runtime 端点的 OpenAI Responses API（/openai/v1）调用 Bedrock 上的前沿模型——OpenAI GPT-6.1 Sol、GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）、xAI Grok 4.6、月之暗面 Kimi K3。仅支持跨区域推理 profile（global.*）——每个区域都能调所有模型，无区域分裂。不支持服务端 web_search（那是 Mantle 独有）。
+revision: 2026101005
+displayName: AWS Bedrock OpenAI-compatible (GPT-6.1 / 6 / 5.6 · Grok 4.7 / 4.6 · Kimi K3 · GLM 5.3)
+displayName_zh: AWS Bedrock OpenAI 兼容接口（GPT-6.1 / 6 / 5.6 · Grok 4.7 / 4.6 · Kimi K3 · GLM 5.3）
+description: Frontier models on Amazon Bedrock via the bedrock-runtime endpoint's OpenAI Responses API (/openai/v1) — OpenAI GPT-6.1 Sol, GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna), xAI Grok 4.7 / 4.6, Moonshot Kimi K3, Z.AI GLM 5.3. Cross-Region inference profiles (global.*) only — every region serves every model, no region split. No server-side web_search (that is Mantle-only).
+description_zh: 通过 bedrock-runtime 端点的 OpenAI Responses API（/openai/v1）调用 Bedrock 上的前沿模型——OpenAI GPT-6.1 Sol、GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）、xAI Grok 4.7 / 4.6、月之暗面 Kimi K3、智谱 GLM 5.3。仅支持跨区域推理 profile（global.*）——每个区域都能调所有模型，无区域分裂。不支持服务端 web_search（那是 Mantle 独有）。
 # Why a second Bedrock/OpenAI provider next to aws-bedrock-mantle: this is the
 # bedrock-runtime host (https://bedrock-runtime.<region>.amazonaws.com/openai/v1),
 # which AWS now recommends for new applications; bedrock-mantle is the legacy
@@ -17,7 +17,8 @@ description_zh: 通过 bedrock-runtime 端点的 OpenAI Responses API（/openai/
 #   - live-probed 2026-09-24 with SigV4: all nine global.* ids below return
 #     200 (gpt-6.1-sol probed 2026-09-30) in us-east-1 / us-west-2 /
 #     eu-west-1 / ap-northeast-1 alike, so the endpoint is purely a latency /
-#     data-residency pick — unlike Mantle's per-region GPT matrix.
+#     data-residency pick — unlike Mantle's per-region GPT matrix. Grok 4.7
+#     and GLM 5.3 (probed 2026-10-08) are bedrock-runtime-only, same rule.
 #   - requests are always synchronous and there are NO server-side tools:
 #     `tools:[{type:'web_search'}]` → 400 "web search is not supported for
 #     this request" (verified). MantleAgent never sends those, but the
@@ -60,8 +61,9 @@ secrets:
 models:
   # OpenAI GPT-6.1 / GPT-6 / GPT-5.6 — the same seven models as
   # aws-bedrock-mantle.yaml (see that file for the per-model background:
-  # 1,050,000 / 128,000 for GPT-6, 272,000 / 64,000 for 5.6, 2x input price
-  # past 272K). Live-probed 2026-09-24 on this host with invalid-value
+  # 1,050,000 context for GPT-6 and 5.6 per the Bedrock model cards, 128,000 /
+  # 64,000 max output, 2x input price past 272K — hence contextWindow 272000
+  # as the new-agent default, maxContextWindow the real limit). Live-probed 2026-09-24 on this host with invalid-value
   # requests: every one (except GPT-6.1 Sol, below) lists reasoning.effort
   # none/low/medium/high/xhigh/max in its own validation error and accepts
   # text.verbosity low/medium/high — identical surface, so presets are copied
@@ -71,6 +73,8 @@ models:
     displayName: GPT-6 Sol
     maxOutputTokens: 128000
     contextWindow: 272000
+    maxContextWindow: 1050000
+    compressAt: 0.9
     capabilities:
       image: true
       video: false
@@ -82,7 +86,7 @@ models:
       thinking:
         mode: adaptive
         defaultEnabled: true
-        default: medium
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -90,7 +94,7 @@ models:
           - { value: xhigh, label: XHigh }
           - { value: max, label: Max }
       verbosity:
-        default: low
+        default: medium
         presets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -100,11 +104,13 @@ models:
   # with SigV4: 200 from all four endpoint presets like the other global.* ids.
   # reasoning.effort accepts low/medium/high/xhigh/max — `none` 400s, unlike
   # GPT-6 Sol, so no "none" preset; text.verbosity low/medium/high; image input
-  # OK; 1,050,000 / 128,000 per the OpenAI model page.
+  # OK; 1M context / 131,072 output per the Bedrock model card.
   - id: global.openai.gpt-6.1-sol
     displayName: GPT-6.1 Sol
-    maxOutputTokens: 128000
+    maxOutputTokens: 131072
     contextWindow: 272000
+    maxContextWindow: 1000000
+    compressAt: 0.9
     capabilities:
       image: true
       video: false
@@ -116,7 +122,7 @@ models:
       thinking:
         mode: adaptive
         defaultEnabled: true
-        default: medium
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -124,7 +130,7 @@ models:
           - { value: xhigh, label: XHigh }
           - { value: max, label: Max }
       verbosity:
-        default: low
+        default: medium
         presets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -134,6 +140,8 @@ models:
     displayName: GPT-6 Astra
     maxOutputTokens: 128000
     contextWindow: 272000
+    maxContextWindow: 1050000
+    compressAt: 0.9
     capabilities:
       image: true
       video: false
@@ -145,7 +153,7 @@ models:
       thinking:
         mode: adaptive
         defaultEnabled: true
-        default: medium
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -153,7 +161,7 @@ models:
           - { value: xhigh, label: XHigh }
           - { value: max, label: Max }
       verbosity:
-        default: low
+        default: medium
         presets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -163,6 +171,8 @@ models:
     displayName: GPT-6 Luna
     maxOutputTokens: 128000
     contextWindow: 272000
+    maxContextWindow: 1050000
+    compressAt: 0.9
     capabilities:
       image: true
       video: false
@@ -174,7 +184,7 @@ models:
       thinking:
         mode: adaptive
         defaultEnabled: true
-        default: medium
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -182,7 +192,7 @@ models:
           - { value: xhigh, label: XHigh }
           - { value: max, label: Max }
       verbosity:
-        default: low
+        default: medium
         presets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -192,6 +202,8 @@ models:
     displayName: GPT-5.6 Sol
     maxOutputTokens: 64000
     contextWindow: 272000
+    maxContextWindow: 1050000
+    compressAt: 0.9
     capabilities:
       image: true
       video: false
@@ -203,7 +215,7 @@ models:
       thinking:
         mode: adaptive
         defaultEnabled: true
-        default: medium
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -211,7 +223,7 @@ models:
           - { value: xhigh, label: XHigh }
           - { value: max, label: Max }
       verbosity:
-        default: low
+        default: medium
         presets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -221,6 +233,8 @@ models:
     displayName: GPT-5.6 Terra
     maxOutputTokens: 64000
     contextWindow: 272000
+    maxContextWindow: 1050000
+    compressAt: 0.9
     capabilities:
       image: true
       video: false
@@ -232,7 +246,7 @@ models:
       thinking:
         mode: adaptive
         defaultEnabled: true
-        default: medium
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -240,7 +254,7 @@ models:
           - { value: xhigh, label: XHigh }
           - { value: max, label: Max }
       verbosity:
-        default: low
+        default: medium
         presets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -250,6 +264,8 @@ models:
     displayName: GPT-5.6 Luna
     maxOutputTokens: 64000
     contextWindow: 272000
+    maxContextWindow: 1050000
+    compressAt: 0.9
     capabilities:
       image: true
       video: false
@@ -261,7 +277,7 @@ models:
       thinking:
         mode: adaptive
         defaultEnabled: true
-        default: medium
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -269,15 +285,62 @@ models:
           - { value: xhigh, label: XHigh }
           - { value: max, label: Max }
       verbosity:
-        default: low
+        default: medium
+        presets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
+  # Grok 4.7 (Bedrock launch 2026-09-28; global. / us. profiles only, no
+  # in-Region id) — self-verification on long tasks, better use of its 500K
+  # context. bedrock-runtime only: the model card lists no bedrock-mantle and
+  # Mantle 404s it in us-east-1 / us-east-2 / us-west-2 (live-probed
+  # 2026-10-08), so it isn't in aws-bedrock-mantle.yaml. Max output isn't
+  # disclosed (the endpoint accepts any max_output_tokens) — 64000 carried
+  # over from Grok 4.6. Live-probed 2026-10-08: reasoning.effort accepts ONLY
+  # low/medium/high/xhigh (`none` and `max` → 400 "unsupported"), same as
+  # 4.6 — so no Max preset; the card's server default is high, the form
+  # default stays medium like 4.6. text.verbosity low/medium/high, image
+  # input OK ("Red"), function_call OK, implicit caching OK (run 2-3
+  # cached_tokens 8448 of 8531). Reachable from all four presets. Launch-week
+  # capacity is flaky: many calls 500/503, or stream an `error` +
+  # `response.failed` frame (server_error / rate_limit_exceeded) right after
+  # response.in_progress, and succeed on a resend. Global CRIS
+  # pricing $2.00 / $6.00 per 1M (cache read $0.50); US Geo $2.20 / $6.60
+  # (cache read $0.55).
+  - id: global.xai.grok-4.7
+    displayName: Grok 4.7
+    maxOutputTokens: 64000
+    contextWindow: 272000
+    maxContextWindow: 500000
+    compressAt: 0.9
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: xhigh
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
+      verbosity:
+        default: medium
         presets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
           - { value: high, label: High }
 
   # Grok 4.6 (Bedrock launch 2026-08-18). 500K context per the Bedrock model
-  # card and xAI's model page (input beyond 200K bills at 2x — cap
-  # context.maxTokens in agent.yaml unless that's intended). Max output isn't
+  # card and xAI's model page (input beyond 200K bills at 2x — lower
+  # context.maxTokens in agent.yaml if that matters). Max output isn't
   # disclosed by either (the endpoint accepts any max_output_tokens without
   # validation) — 64000 is the same conservative carry-over used for GPT-5.6.
   # Live-probed 2026-09-24: reasoning.effort accepts ONLY low/medium/high/xhigh
@@ -288,6 +351,8 @@ models:
     displayName: Grok 4.6
     maxOutputTokens: 64000
     contextWindow: 272000
+    maxContextWindow: 500000
+    compressAt: 0.9
     capabilities:
       image: true
       video: false
@@ -299,14 +364,14 @@ models:
       thinking:
         mode: adaptive
         defaultEnabled: true
-        default: medium
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
           - { value: high, label: High }
           - { value: xhigh, label: XHigh }
       verbosity:
-        default: low
+        default: medium
         presets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -325,6 +390,8 @@ models:
     displayName: Kimi K3
     maxOutputTokens: 131072
     contextWindow: 272000
+    maxContextWindow: 1000000
+    compressAt: 0.9
     capabilities:
       image: true
       video: false
@@ -336,7 +403,57 @@ models:
       thinking:
         mode: adaptive
         defaultEnabled: true
+        default: xhigh
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
+          - { value: max, label: Max }
+      verbosity:
         default: medium
+        presets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
+  # GLM 5.3 on Bedrock (launch 2026-10-05; global. / us. profiles only, no
+  # in-Region id) — Z.AI's 744B-total / ~40B-active MoE for agentic coding.
+  # The native API is in zhipu.yaml; this is the Bedrock route. Serving
+  # surface, live-probed 2026-10-08: bedrock-runtime /openai/v1 Responses
+  # (200 from all four presets, streaming + function_call OK) — Mantle 404s
+  # it in every region (the card lists no bedrock-mantle either), Converse /
+  # Invoke / chat-completions exist too but MantleAgent needs only Responses.
+  # reasoning.effort: the validation error lists none/low/medium/high/xhigh/
+  # max and all six return 200, reasoning_tokens rising with effort on a
+  # primality prompt (0 / 398 / 458 / 454 / 634 / 792) — unlike the native
+  # API's low/high/max. `none` gets no preset (MantleAgent maps it to high);
+  # Thinking off omits the field and the model still thinks. text.verbosity
+  # low/medium/high. Text-only: an input_image part isn't rejected but the
+  # model says it can't see it, so image: false. Card: 1M context, 128K max
+  # output (the endpoint doesn't validate max_output_tokens). Implicit
+  # caching OK (run 1 cache_write_tokens 9023, runs 2-3 cached_tokens 9023).
+  # The card marks access as limited to eligible customers (this account
+  # has it). Pricing not yet in the AWS Price List API — see the Bedrock
+  # pricing page.
+  - id: global.zai.glm-5.3
+    displayName: GLM 5.3
+    maxOutputTokens: 128000
+    contextWindow: 272000
+    maxContextWindow: 1000000
+    compressAt: 0.9
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -344,7 +461,7 @@ models:
           - { value: xhigh, label: XHigh }
           - { value: max, label: Max }
       verbosity:
-        default: low
+        default: medium
         presets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }

package/templates/models/deepseek.yaml

@@ -1,6 +1,6 @@
 id: deepseek
 runtime: deepseek
-revision: 2026100501
+revision: 2026101002
 displayName: DeepSeek
 displayName_zh: DeepSeek
 description: DeepSeek V4 via OpenAI-compatible chat completions API
@@ -30,6 +30,8 @@ models:
     displayName: DeepSeek V4 Pro
     maxOutputTokens: 65536
     contextWindow: 272000
+    maxContextWindow: 1000000
+    compressAt: 0.9
     capabilities:
       image: false
       video: false
@@ -40,7 +42,7 @@ models:
           - { value: 5m, label: 5min }
       thinking:
         defaultEnabled: true
-        default: medium
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
@@ -56,6 +58,8 @@ models:
     displayName: DeepSeek Flash
     maxOutputTokens: 65536
     contextWindow: 272000
+    maxContextWindow: 1000000
+    compressAt: 0.9
     capabilities:
       image: true
       video: false
@@ -66,7 +70,7 @@ models:
           - { value: 5m, label: 5min }
       thinking:
         defaultEnabled: true
-        default: medium
+        default: xhigh
         effortPresets:
           - { value: low, label: Low }
           - { value: medium, label: Medium }

package/templates/models/doubao.yaml

@@ -1,12 +1,12 @@
 id: doubao
 runtime: doubao
-revision: 2026100501
+revision: 2026101005
 displayName: Doubao (Volcengine)
 displayName_zh: 豆包（火山方舟）
-description: ByteDance Doubao Seed-2.0 family on Volcengine Ark — pro / lite / mini / code, OpenAI-compatible
-description_zh: 字节豆包 Seed-2.0 系列（火山方舟），OpenAI 兼容协议
+description: ByteDance Doubao Seed-2.1 / 2.0 family on Volcengine Ark — pro / lite / mini / code, OpenAI-compatible
+description_zh: 字节豆包 Seed-2.1 / 2.0 系列（火山方舟），OpenAI 兼容协议
 defaultEndpoint: https://ark.cn-beijing.volces.com/api/v3
-defaultModelId: doubao-seed-2-0-pro-260215
+defaultModelId: doubao-seed-2-1-pro-260628
 
 secrets:
   - key: api_key
@@ -32,19 +32,34 @@ secrets:
 #         N+1 misses, turn N+2 hits again) — not stable steady-state.
 #     Treat the cache as opportunistic; UI badges are accurate when shown.
 #   - Tool calls return standard OpenAI tool_calls shape.
-#   - Image: pro/lite/mini all accepted image_url blocks; in our smoke
-#     test all four answered "Black" for what was actually a red 64×64 PNG,
-#     so reliability is uncertain. We default image:false; user can flip
-#     after their own verification.
+#   - Image: the 2.0 models accepted image_url blocks, but all four answered
+#     "Black" for a red 64×64 PNG, so they stay image:false (halo then
+#     drops user images with a notice and doesn't offer view_image). Only
+#     2.1-pro-260628 is image:true — doubao-agent sends images as content
+#     parts (halo 1.6.1+; earlier versions dropped them).
 #   - Doubao 2.0 docs aren't fully published; model id strings sourced
 #     from the Volcengine Ark console (the activated-model list).
+# 2026-10-10 census: limits are the Ark `GET /api/v3/models` token_limits
+# (== the official model list, docs/82379/1330310): 2.1-pro-260628 and the
+# 2.0 pro / code 260215 → 262144 ctx; 2.0 lite / mini 260428 → 262144 ctx but
+# 229376 max input, so contextWindow = 229376; every 2.x model takes up to
+# 131072 output (max_tokens 131073 → 400) and Ark trims the reply to the
+# remaining quota instead of erroring. 2.1-pro-260628 live-probed: thinking
+# enabled/disabled, tool calls, image. Not carried: 2.1-pro / lite-260915 and
+# doubao-seed-evolving (1M ctx) return ModelNotOpen on our account — add them
+# after a live probe on an account that has them activated; 2.1-turbo-260628
+# likewise. Ark's newer `reasoning_effort` knob (deep-thinking doc) is not
+# wired — doubao-agent still sends only thinking:{type}.
 models:
-  - id: doubao-seed-2-0-pro-260215
-    displayName: Doubao Seed 2.0 Pro
-    maxOutputTokens: 16384
-    contextWindow: 256000
+  - id: doubao-seed-2-1-pro-260628
+    displayName: Doubao Seed 2.1 Pro
+    maxOutputTokens: 131072
+    contextWindow: 262144
+    maxContextWindow: 262144
+    compressAt: 0.9
     capabilities:
-      image: false
+      # Live probe 2026-10-10: named the red 64×64 PNG correctly.
+      image: true
       video: false
       audio: false
       promptCaching:
@@ -69,8 +84,10 @@ models:
 
   - id: doubao-seed-2-0-lite-260428
     displayName: Doubao Seed 2.0 Lite
-    maxOutputTokens: 16384
-    contextWindow: 256000
+    maxOutputTokens: 131072
+    contextWindow: 229376
+    maxContextWindow: 229376
+    compressAt: 0.9
     capabilities:
       image: false
       video: false
@@ -94,8 +111,10 @@ models:
 
   - id: doubao-seed-2-0-mini-260428
     displayName: Doubao Seed 2.0 Mini
-    maxOutputTokens: 16384
-    contextWindow: 256000
+    maxOutputTokens: 131072
+    contextWindow: 229376
+    maxContextWindow: 229376
+    compressAt: 0.9
     capabilities:
       image: false
       video: false
@@ -117,10 +136,43 @@ models:
         effortPresets:
           - { value: medium, label: On }
 
+  # --- deprecated: retires 2026-11-24 14:00 CST (https://www.volcengine.com/docs/82379/1350667) — remove after cutoff ---
+  - id: doubao-seed-2-0-pro-260215
+    displayName: Doubao Seed 2.0 Pro (deprecated 2026-11-24)
+    maxOutputTokens: 131072
+    contextWindow: 262144
+    maxContextWindow: 262144
+    compressAt: 0.9
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        # Doubao caches automatically (no cache_control parameter needed
+        # in the request). The badge is just a visibility hint — UI shows
+        # the read/write counts so users see when cache is helping.
+        # Doc says default TTL is the platform default; observed empirically
+        # to behave like ~5min ephemeral. Cache only kicks in for prompts
+        # above ~4096 tokens; first 1-3 turns of a fresh prompt commonly
+        # miss while the cache is warming up.
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        defaultEnabled: true
+        default: medium
+        # effort label is Halo-side only — Doubao only sees enabled/disabled.
+        # We still expose Off/On as toggles by mapping both options to the
+        # same flag (keeps the form consistent across providers).
+        effortPresets:
+          - { value: medium, label: On }
+
   - id: doubao-seed-2-0-code-preview-260215
-    displayName: Doubao Seed 2.0 Code Preview
-    maxOutputTokens: 16384
-    contextWindow: 256000
+    displayName: Doubao Seed 2.0 Code Preview (deprecated 2026-11-24)
+    maxOutputTokens: 131072
+    contextWindow: 262144
+    maxContextWindow: 262144
+    compressAt: 0.9
     capabilities:
       image: false
       video: false

package/templates/models/hunyuan.yaml

@@ -1,17 +1,17 @@
 id: hunyuan
 runtime: hunyuan
-revision: 2026100501
+revision: 2026101003
 displayName: Hunyuan (Tencent)
 displayName_zh: 混元（腾讯）
-description: Tencent Hunyuan Hy3 preview — 295B/21B MoE, agent-oriented with no_think / think_low / think_high modes
-description_zh: 腾讯混元 Hy3 preview，295B/21B 激活 MoE，面向 Agent 工作负载，支持 no_think / think_low / think_high 三档思考
+description: Tencent Hunyuan Hy4 preview / Hy3 on TokenHub — agent-oriented MoE with no_think / think_low / think_high modes
+description_zh: 腾讯混元 Hy4 preview / Hy3（TokenHub），面向 Agent 的 MoE 模型，支持 no_think / think_low / think_high 三档思考
 defaultEndpoint: https://tokenhub.tencentmaas.com/v1
-defaultModelId: hy3-preview
+defaultModelId: hy3
 
 secrets:
   - key: api_key
-    description: Tencent Hunyuan Hy3 API Key. Format starts with sk-.
-    description_zh: 腾讯混元 Hy3 API Key，sk- 开头
+    description: Tencent Hunyuan (TokenHub) API Key. Format starts with sk-.
+    description_zh: 腾讯混元（TokenHub）API Key，sk- 开头
     default: <<HUNYUAN_API_KEY>>
     secret: true
 
@@ -22,20 +22,37 @@ secrets:
 #     system prompt across turns shows usage.prompt_tokens_details
 #     .cached_tokens > 0 from turn 2 onward.
 #   - Thinking modes per Hy3 product naming → `reasoning_effort`:
-#       no_think    → omit reasoning_effort (zero reasoning_tokens)
+#       no_think    → reasoning_effort: 'none' (zero reasoning_tokens)
 #       think_low   → reasoning_effort: 'low'
 #       think_high  → reasoning_effort: 'high'
-#     hunyuan-agent.ts maps thinking.enabled=false → no_think (no field),
+#     hunyuan-agent.ts maps thinking.enabled=false → 'none',
 #     thinking.enabled=true + effort=low|medium|high → reasoning_effort.
 #   - Image input: gateway accepts image_url blocks but model says "I don't
 #     see an image" — image:false.
-#   - 256K context per product description (not enforced here; model just
-#     accepts long inputs).
+# 2026-10-10 census (TokenHub `GET /v1/models` + live probes):
+#   - Lineup: Hy4 preview (2026-08-28, 770B/49B MoE, 1M context — official
+#     release https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/,
+#     HF config max_position_embeddings 1048576) and Hy3 GA (2026-07-06,
+#     "building on Hy3 preview", 256K — https://www.tencent.com/en-us/tencent-hunyuan-officially-releases-hy3-advancing-agent-capabilities-and-deeper-product-integration/,
+#     HF config 262144). Both listed `online`; hy3-preview is no longer listed
+#     but still answers — no retirement date published, kept at the tail.
+#   - hy3 / hy4-preview think by DEFAULT when reasoning_effort is omitted
+#     (reasoning_tokens > 0; hy3-preview still doesn't), so "off" must send
+#     reasoning_effort:'none' — verified 0 reasoning tokens on all three.
+#   - Caching re-verified on hy3 / hy4-preview (turn 2 cached_tokens ≈ prompt).
+#   - Tool calls verified on hy3 / hy4-preview. Image: hy3 / hy4-preview both
+#     answered wrongly about a solid red PNG — keep image:false.
+#   - Max output: no official per-model number found (search-only claims of
+#     64K, unconfirmed: https://www.tencentcloud.com/techpedia/147980); the
+#     gateway accepted max_tokens 2,000,000 without error, so it isn't a hard
+#     check either — maxOutputTokens stays at the conservative 16384.
 models:
-  - id: hy3-preview
-    displayName: Hunyuan Hy3 Preview
+  - id: hy3
+    displayName: Hunyuan Hy3
     maxOutputTokens: 16384
-    contextWindow: 256000
+    contextWindow: 262144
+    maxContextWindow: 262144
+    compressAt: 0.9
     capabilities:
       image: false
       video: false
@@ -53,7 +70,50 @@ models:
         # we expose Off / Low / High since "Off" is the natural opposite of
         # an enable toggle.
         defaultEnabled: true
-        default: low
+        default: high
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: high, label: High }
+
+  - id: hy4-preview
+    displayName: Hunyuan Hy4 Preview
+    maxOutputTokens: 16384
+    contextWindow: 272000
+    maxContextWindow: 1048576
+    compressAt: 0.9
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        defaultEnabled: true
+        default: high
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: high, label: High }
+
+  # --- superseded by hy3 (GA); dropped from /v1/models, no retirement date published ---
+  - id: hy3-preview
+    displayName: Hunyuan Hy3 Preview
+    maxOutputTokens: 16384
+    contextWindow: 262144
+    maxContextWindow: 262144
+    compressAt: 0.9
+    capabilities:
+      image: false
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        defaultEnabled: true
+        default: high
         effortPresets:
           - { value: low, label: Low }
           - { value: high, label: High }