From 91b662c2b39377d0d5cf1c4ac21210ac0ec874f8 Mon Sep 17 00:00:00 2001
From: dongjiang1989 <dongjiang1989@126.com>
Date: Tue, 15 Sep 2026 16:09:02 +0800
Subject: [PATCH] =?UTF-8?q?feat:=20add=20iFlytek=20(=E8=AE=AF=E9=A3=9E?=
 =?UTF-8?q?=E6=98=9F=E8=BE=B0=20MaaS)=20provider?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Adds the iFlytek MaaS Inference API provider (讯飞星辰MaaS推理服务, https://maas.xfyun.cn) with 18 models — 4 first-party Spark and 14 third-party (DeepSeek, GLM, Kimi, MiniMax, Qwen).

Provider: providers/iflytek/
- API: https://maas-api.cn-huabei-1.xf-yun.com/v2 (OpenAI-compatible, text-only)
- env: IFLYTEK_API_KEY
- npm: @ai-sdk/openai-compatible

Lab models (models/iflytek/):
- spark-x2.5: 293B-A30B MoE flagship, 256K context, tool_call=true
- spark-x2-flash: lightweight fast model, 256K context, tool_call=true
- spark-x2.5-4b: 4B edge Dense, 1M context, open-weight, tool_call=true
- spark-x2.5-1.7b: 1.7B ultra-light edge, 1M context, open-weight

Provider models (18 total) — override-only after base_model merge:
- Spark (first-party): spark-x2.5, spark-x2-flash, spark-x2.5-4b, spark-x2.5-1.7b
  (tool_call overridden to false on all 4 — host lacks tool support)
- DeepSeek: deepseek-v4-pro, deepseek-v4-flash, deepseek-v3.2
- GLM/Zhipu: glm-5.2, glm-5.1, glm-5, glm-4-7-flash
- Kimi/Moonshot: kimi-k2.6, kimi-k2.5
- MiniMax: minimax-m2.5
- Qwen/Alibaba: qwen3.5-397b-a17b, qwen3.6-35b-a3b, qwen3.5-35b-a3b, qwen3-coder-next

Reasoning controls (per iFlytek host docs, aligned with lab/peer baselines):
- DeepSeek V4 Pro: toggle (thinking.type) + effort high/max
- DeepSeek V4 Flash: toggle (thinking.type) + effort low/high/max
- DeepSeek V3.2: toggle (thinking.type) only (matches OpenRouter/TokenGo/Novita)
- GLM-5.2: effort high/max (only GLM with graded effort on this host)
- GLM-5.1/GLM-5: toggle (enable_thinking) only
- GLM-4.7-Flash: toggle (enable_thinking) (matches first-party zhipuai baseline)
- Kimi K2.6/K2.5: toggle (enable_thinking) only
- MiniMax M2.5: [] always-on, no caller control (matches minimax first-party,
  opencode, alibaba-token-plan-cn, cortecs, ollama-cloud peers)
- Qwen 397B/3.6-35B/3.5-35B: toggle (enable_thinking) only — host 'OSS'
  phrase targets OpenAI OSS gpt-oss, not Qwen; no budget_tokens (host lacks
  thinking_budget); no graded effort
- Qwen3-Coder-Next: non-reasoning
- Spark-X2.5/Flash/4B/1.7B: toggle (enable_thinking) + high/max (same surface)

All toggle models have leading '# Toggle: <field> = <values>' wire comments.
No redundant restated lab values (attachment, modalities, tool_call only
overridden when they differ from the lab entry).

Context limit overrides (host < lab):
- GLM-4.7-Flash: 128K (lab 200K)
- Kimi K2.5: 128K (lab 262K)
- MiniMax M2.5: 128K (lab 204.8K)
- Qwen3.5-397B-A17B: 256K (lab 262K)
- Qwen3.6-35B-A3B / Qwen3.5-35B-A3B: 128K (lab 262K)

Multimodal override: attachment=false + text-only [modalities] on Kimi and
Qwen (base labs are multimodal; iFlytek endpoint is text-only).

Pricing: USD per million tokens from Token Plan points at 0.01 CNY/point
(1 point ≈ 0.01 CNY per Token Plan doc: 200 CNY / 20000 points for standard
members), converted at 7.25 CNY/USD. Spark-X2.5 uses list price (CNY
3.2/0.48/12); 50% promo noted separately. Spark-X2.5-4B/1.7B currently
limited-time free (cost = 0).

Schema: added 'spark' to ModelFamily enum in packages/core/src/family.ts.

Data sources (accessed 2026-09-14):
- https://www.xfyun.cn/doc/spark/推理服务-http.html
- https://www.xfyun.cn/doc/spark/TokenPlan.html
- https://maas.xfyun.cn/modelSquare

Excluded: Token Plan API and Coding Plan API (separate billing surfaces),
Spark-X2/Spark-X2-Agent (decommissioned).

bun validate passes locally.

Signed-off-by: dongjiang1989 <dongjiang1989@126.com>
---
 models/iflytek/spark-x2-flash.toml            | 19 +++++++++
 models/iflytek/spark-x2.5-1.7b.toml           | 24 +++++++++++
 models/iflytek/spark-x2.5-4b.toml             | 24 +++++++++++
 models/iflytek/spark-x2.5.toml                | 20 ++++++++++
 packages/core/src/family.ts                   |  3 ++
 providers/iflytek/logo.svg                    |  4 ++
 providers/iflytek/models/deepseek-v3.2.toml   | 23 +++++++++++
 .../iflytek/models/deepseek-v4-flash.toml     | 30 ++++++++++++++
 providers/iflytek/models/deepseek-v4-pro.toml | 30 ++++++++++++++
 providers/iflytek/models/glm-4-7-flash.toml   | 25 ++++++++++++
 providers/iflytek/models/glm-5.1.toml         | 20 ++++++++++
 providers/iflytek/models/glm-5.2.toml         | 22 ++++++++++
 providers/iflytek/models/glm-5.toml           | 20 ++++++++++
 providers/iflytek/models/kimi-k2.5.toml       | 30 ++++++++++++++
 providers/iflytek/models/kimi-k2.6.toml       | 27 +++++++++++++
 providers/iflytek/models/minimax-m2.5.toml    | 23 +++++++++++
 .../iflytek/models/qwen3-coder-next.toml      | 14 +++++++
 providers/iflytek/models/qwen3.5-35b-a3b.toml | 32 +++++++++++++++
 .../iflytek/models/qwen3.5-397b-a17b.toml     | 36 +++++++++++++++++
 providers/iflytek/models/qwen3.6-35b-a3b.toml | 32 +++++++++++++++
 providers/iflytek/models/spark-x2-flash.toml  | 26 ++++++++++++
 providers/iflytek/models/spark-x2.5-1.7b.toml | 27 +++++++++++++
 providers/iflytek/models/spark-x2.5-4b.toml   | 28 +++++++++++++
 providers/iflytek/models/spark-x2.5.toml      | 27 +++++++++++++
 providers/iflytek/provider.toml               | 40 +++++++++++++++++++
 25 files changed, 606 insertions(+)
 create mode 100644 models/iflytek/spark-x2-flash.toml
 create mode 100644 models/iflytek/spark-x2.5-1.7b.toml
 create mode 100644 models/iflytek/spark-x2.5-4b.toml
 create mode 100644 models/iflytek/spark-x2.5.toml
 create mode 100644 providers/iflytek/logo.svg
 create mode 100644 providers/iflytek/models/deepseek-v3.2.toml
 create mode 100644 providers/iflytek/models/deepseek-v4-flash.toml
 create mode 100644 providers/iflytek/models/deepseek-v4-pro.toml
 create mode 100644 providers/iflytek/models/glm-4-7-flash.toml
 create mode 100644 providers/iflytek/models/glm-5.1.toml
 create mode 100644 providers/iflytek/models/glm-5.2.toml
 create mode 100644 providers/iflytek/models/glm-5.toml
 create mode 100644 providers/iflytek/models/kimi-k2.5.toml
 create mode 100644 providers/iflytek/models/kimi-k2.6.toml
 create mode 100644 providers/iflytek/models/minimax-m2.5.toml
 create mode 100644 providers/iflytek/models/qwen3-coder-next.toml
 create mode 100644 providers/iflytek/models/qwen3.5-35b-a3b.toml
 create mode 100644 providers/iflytek/models/qwen3.5-397b-a17b.toml
 create mode 100644 providers/iflytek/models/qwen3.6-35b-a3b.toml
 create mode 100644 providers/iflytek/models/spark-x2-flash.toml
 create mode 100644 providers/iflytek/models/spark-x2.5-1.7b.toml
 create mode 100644 providers/iflytek/models/spark-x2.5-4b.toml
 create mode 100644 providers/iflytek/models/spark-x2.5.toml
 create mode 100644 providers/iflytek/provider.toml

diff --git a/models/iflytek/spark-x2-flash.toml b/models/iflytek/spark-x2-flash.toml
new file mode 100644
index 00000000000..08226ddefe4
--- /dev/null
+++ b/models/iflytek/spark-x2-flash.toml
@@ -0,0 +1,19 @@
+name = "Spark X2 Flash"
+description = "Lightweight iFlytek Spark model for fast, economical text generation"
+family = "spark"
+release_date = "2026-06-01"
+last_updated = "2026-06-01"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 262_144
+output = 65_536
+
+[modalities]
+input = ["text"]
+output = ["text"]
diff --git a/models/iflytek/spark-x2.5-1.7b.toml b/models/iflytek/spark-x2.5-1.7b.toml
new file mode 100644
index 00000000000..05c0da164ba
--- /dev/null
+++ b/models/iflytek/spark-x2.5-1.7b.toml
@@ -0,0 +1,24 @@
+name = "Spark X2.5 1.7B"
+description = "Ultra-lightweight 1.7B edge Dense model with native 1M context; designed for IoT/smart-home control, robotics, and on-device agents; 90.3% Domux control accuracy at 0.85s avg response"
+family = "spark"
+release_date = "2026-09-01"
+last_updated = "2026-09-01"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = false
+structured_output = false
+knowledge = "2026-08"
+open_weights = true
+
+[limit]
+context = 1_048_576
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/XHToken/Spark-X2.5-1.7B"
diff --git a/models/iflytek/spark-x2.5-4b.toml b/models/iflytek/spark-x2.5-4b.toml
new file mode 100644
index 00000000000..de2373c34e1
--- /dev/null
+++ b/models/iflytek/spark-x2.5-4b.toml
@@ -0,0 +1,24 @@
+name = "Spark X2.5 4B"
+description = "Compact 4B edge Dense model with native 1M context; hybrid attention (1 full + 3 sliding window layers); 200+ languages; agentic, tool-use, code, and math reasoning"
+family = "spark"
+release_date = "2026-09-01"
+last_updated = "2026-09-01"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = false
+knowledge = "2026-08"
+open_weights = true
+
+[limit]
+context = 1_048_576
+output = 65_536
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/XHToken/Spark-X2.5-4B"
diff --git a/models/iflytek/spark-x2.5.toml b/models/iflytek/spark-x2.5.toml
new file mode 100644
index 00000000000..9fbc585d4b6
--- /dev/null
+++ b/models/iflytek/spark-x2.5.toml
@@ -0,0 +1,20 @@
+name = "Spark X2.5"
+description = "iFlytek flagship 293B-A30B MoE model with 256K context, 200+ languages, strong at code and agentic tasks"
+family = "spark"
+release_date = "2026-09-07"
+last_updated = "2026-09-07"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2026-08"
+open_weights = false
+
+[limit]
+context = 262_144
+output = 65_536
+
+[modalities]
+input = ["text"]
+output = ["text"]
diff --git a/packages/core/src/family.ts b/packages/core/src/family.ts
index 9f719d80376..deb243e3120 100644
--- a/packages/core/src/family.ts
+++ b/packages/core/src/family.ts
@@ -445,6 +445,9 @@ export const ModelFamilyValues = [
 
   // Elephant
   "elephant",
+
+  // iFlytek Spark
+  "spark",
 ] as const;
 
 export const ModelFamily = z.enum(ModelFamilyValues);
diff --git a/providers/iflytek/logo.svg b/providers/iflytek/logo.svg
new file mode 100644
index 00000000000..538fed07fec
--- /dev/null
+++ b/providers/iflytek/logo.svg
@@ -0,0 +1,4 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor">
+  <path d="M12 2C6.48 2 2 6.48 2 12s4.48 10 10 10 10-4.48 10-10S17.52 2 12 2zm0 18c-4.41 0-8-3.59-8-8s3.59-8 8-8 8 3.59 8 8-3.59 8-8 8z"/>
+  <path d="M8 11h2v5H8zm3-3h2v8h-2zm3 2h2v6h-2z"/>
+</svg>
diff --git a/providers/iflytek/models/deepseek-v3.2.toml b/providers/iflytek/models/deepseek-v3.2.toml
new file mode 100644
index 00000000000..77d57eaab3e
--- /dev/null
+++ b/providers/iflytek/models/deepseek-v3.2.toml
@@ -0,0 +1,23 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# DeepSeek-V3.2: 128K context, hybrid reasoning with thinking/non-thinking modes.
+# Tool calling SUPPORTED on this host (per host docs, only V3.2 and GLM-4.7).
+# Toggle: thinking.type = enabled|disabled per host docs (matching
+# established relay baseline from OpenRouter, TokenGo, Novita for this
+# hybrid model — toggle only, no graded effort).
+# Pricing: Token Plan points 200/40/300 per 1M at 0.01 CNY/point
+# (Token Plan doc defines points as the unified meter; 1 point ≈ 0.01 CNY
+# is the published conversion for standard members at 200 CNY / 20000 points),
+# converted at 7.25 CNY/USD.
+
+base_model = "deepseek/deepseek-v3.2"
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.28
+output = 0.41
+cache_read = 0.06
diff --git a/providers/iflytek/models/deepseek-v4-flash.toml b/providers/iflytek/models/deepseek-v4-flash.toml
new file mode 100644
index 00000000000..47cdafb4298
--- /dev/null
+++ b/providers/iflytek/models/deepseek-v4-flash.toml
@@ -0,0 +1,30 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# DeepSeek-V4-Flash: 1M context, fast economical reasoning model.
+# Toggle: thinking.type = enabled|disabled (OpenAI format) per host docs;
+# effort = low|high|max per host docs (host supports thinking intensity
+# control for DeepSeek V4 via reasoning_effort or output_config.effort).
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Pricing: Token Plan points 100/20/200 per 1M, converted at 7.25 CNY/USD.
+
+base_model = "deepseek/deepseek-v4-flash"
+tool_call = false
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.14
+output = 0.28
+cache_read = 0.03
+
+[limit]
+output = 131_072
diff --git a/providers/iflytek/models/deepseek-v4-pro.toml b/providers/iflytek/models/deepseek-v4-pro.toml
new file mode 100644
index 00000000000..81e1cac8baa
--- /dev/null
+++ b/providers/iflytek/models/deepseek-v4-pro.toml
@@ -0,0 +1,30 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# DeepSeek-V4-Pro: 1M context, reasoning model.
+# Toggle: thinking.type = enabled|disabled (OpenAI format) per host docs;
+# effort = high|max per host docs (host supports thinking intensity control
+# for DeepSeek V4 via reasoning_effort or output_config.effort = high|max).
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Pricing: Token Plan points 1200/240/2400 per 1M, converted at 7.25 CNY/USD.
+
+base_model = "deepseek/deepseek-v4-pro"
+tool_call = false
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 1.66
+output = 3.31
+cache_read = 0.33
+
+[limit]
+output = 131_072
diff --git a/providers/iflytek/models/glm-4-7-flash.toml b/providers/iflytek/models/glm-4-7-flash.toml
new file mode 100644
index 00000000000..1f6ebc3e273
--- /dev/null
+++ b/providers/iflytek/models/glm-4-7-flash.toml
@@ -0,0 +1,25 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# GLM-4.7-Flash: 128K context on this host (Token Plan listing; lab is 200K).
+# Toggle: enable_thinking = true|false per host docs ("仅支持思考的模型不受
+# 此参数影响" — GLM-4.7-Flash is a thinking model, so the toggle applies).
+# Matches first-party zhipuai baseline and peers (empiriolabs, poe/novita).
+# Tool calling SUPPORTED on this host (per host docs, only V3.2 and GLM-4.7).
+# Pricing: Token Plan points 100/20/150 per 1M at 0.01 CNY/point
+# (Token Plan doc: 200 CNY / 20000 points for standard members),
+# converted at 7.25 CNY/USD.
+
+base_model = "zhipuai/glm-4.7-flash"
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.14
+output = 0.21
+cache_read = 0.03
+
+[limit]
+context = 128_000
diff --git a/providers/iflytek/models/glm-5.1.toml b/providers/iflytek/models/glm-5.1.toml
new file mode 100644
index 00000000000..a7be6884e26
--- /dev/null
+++ b/providers/iflytek/models/glm-5.1.toml
@@ -0,0 +1,20 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# GLM-5.1: 200K context on this host (matches lab).
+# Toggle: enable_thinking = true|false per host docs (matches lab/peer
+# baseline for GLM-5.1; only GLM-5.2 supports graded high/max on this host).
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Pricing: Token Plan points 800/160/2800 per 1M, converted at 7.25 CNY/USD.
+
+base_model = "zhipuai/glm-5.1"
+tool_call = false
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 1.10
+output = 3.86
+cache_read = 0.22
diff --git a/providers/iflytek/models/glm-5.2.toml b/providers/iflytek/models/glm-5.2.toml
new file mode 100644
index 00000000000..ecba9329d07
--- /dev/null
+++ b/providers/iflytek/models/glm-5.2.toml
@@ -0,0 +1,22 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# GLM-5.2: 1M context on this host (matches lab).
+# Reasoning: reasoning_effort = high|max per iFlytek host docs (only GLM-5.2
+# among GLM models supports graded effort on this host via
+# reasoning_effort or output_config.effort; GLM-5.1 and GLM-5 expose toggle
+# only).
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Pricing: Token Plan points 800/200/2800 per 1M, converted at 7.25 CNY/USD.
+
+base_model = "zhipuai/glm-5.2"
+tool_call = false
+reasoning_options = [{ type = "effort", values = ["high", "max"] }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 1.10
+output = 3.86
+cache_read = 0.28
diff --git a/providers/iflytek/models/glm-5.toml b/providers/iflytek/models/glm-5.toml
new file mode 100644
index 00000000000..836d1870c5c
--- /dev/null
+++ b/providers/iflytek/models/glm-5.toml
@@ -0,0 +1,20 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# GLM-5: 200K context on this host (matches lab).
+# Toggle: enable_thinking = true|false per host docs (matches lab/peer
+# baseline for GLM-5; only GLM-5.2 supports graded high/max on this host).
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Pricing: Token Plan points 600/120/2200 per 1M, converted at 7.25 CNY/USD.
+
+base_model = "zhipuai/glm-5"
+tool_call = false
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.83
+output = 3.03
+cache_read = 0.17
diff --git a/providers/iflytek/models/kimi-k2.5.toml b/providers/iflytek/models/kimi-k2.5.toml
new file mode 100644
index 00000000000..6c98396de18
--- /dev/null
+++ b/providers/iflytek/models/kimi-k2.5.toml
@@ -0,0 +1,30 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# Kimi-K2.5: 128K context on this host (Token Plan listing; lab 262K).
+# Toggle: enable_thinking = true|false per host docs (enable_thinking bool
+# controls thinking on/off; no per-model effort grading on this host for
+# Kimi). Tool calling NOT supported on this host (only DeepSeek V3.2 and
+# GLM-4.7). Text-only on this API; base lab's multimodal capabilities not
+# exposed here.
+# Pricing: Token Plan points 400/80/2100 per 1M, converted at 7.25 CNY/USD.
+
+base_model = "moonshotai/kimi-k2.5"
+tool_call = false
+attachment = false
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[cost]
+input = 0.55
+output = 2.90
+cache_read = 0.11
+
+[limit]
+context = 128_000
diff --git a/providers/iflytek/models/kimi-k2.6.toml b/providers/iflytek/models/kimi-k2.6.toml
new file mode 100644
index 00000000000..d08941c337b
--- /dev/null
+++ b/providers/iflytek/models/kimi-k2.6.toml
@@ -0,0 +1,27 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# Kimi-K2.6: 256K context on this host (matches lab).
+# Toggle: enable_thinking = true|false per host docs (enable_thinking bool
+# controls thinking on/off; no per-model effort grading on this host for
+# Kimi). Tool calling NOT supported on this host (only DeepSeek V3.2 and
+# GLM-4.7). Text-only on this API; base lab's multimodal capabilities not
+# exposed here.
+# Pricing: Token Plan points 650/130/2700 per 1M, converted at 7.25 CNY/USD.
+
+base_model = "moonshotai/kimi-k2.6"
+tool_call = false
+attachment = false
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[cost]
+input = 0.90
+output = 3.72
+cache_read = 0.18
diff --git a/providers/iflytek/models/minimax-m2.5.toml b/providers/iflytek/models/minimax-m2.5.toml
new file mode 100644
index 00000000000..63e7150d3c5
--- /dev/null
+++ b/providers/iflytek/models/minimax-m2.5.toml
@@ -0,0 +1,23 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# MiniMax-M2.5: 128K context on this host (Token Plan listing; lab 204.8K).
+# Reasoning: always-on with no caller control (matching peer baseline from
+# opencode — reasoning_options = []). reasoning_content returned interleaved.
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Pricing: Token Plan points 210/42/840 per 1M, converted at 7.25 CNY/USD.
+
+base_model = "minimax/MiniMax-M2.5"
+tool_call = false
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.29
+output = 1.16
+cache_read = 0.06
+
+[limit]
+context = 128_000
diff --git a/providers/iflytek/models/qwen3-coder-next.toml b/providers/iflytek/models/qwen3-coder-next.toml
new file mode 100644
index 00000000000..8ee2d2ccb03
--- /dev/null
+++ b/providers/iflytek/models/qwen3-coder-next.toml
@@ -0,0 +1,14 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# Qwen3-Coder-Next-FP8: 256K context, code-specialized model.
+# Non-reasoning per base lab model.
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Pricing: Token Plan points 250/50/1000 per 1M, converted at 7.25 CNY/USD.
+
+base_model = "alibaba/qwen3-coder-next"
+tool_call = false
+
+[cost]
+input = 0.34
+output = 1.38
+cache_read = 0.07
diff --git a/providers/iflytek/models/qwen3.5-35b-a3b.toml b/providers/iflytek/models/qwen3.5-35b-a3b.toml
new file mode 100644
index 00000000000..2d1379af736
--- /dev/null
+++ b/providers/iflytek/models/qwen3.5-35b-a3b.toml
@@ -0,0 +1,32 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# Qwen3.5-35B-A3B: 128K context on this host (Token Plan; lab 262_144).
+# Toggle: enable_thinking = true|false per host docs. No budget_tokens
+# (host does not expose thinking_budget). No graded effort (host "OSS"
+# phrase targets OpenAI OSS gpt-oss family, not Qwen; matching
+# toggle-only peer baseline).
+# Tool calling NOT supported on this host. Text-only on this API; base
+# lab is multimodal, attachment/modalities overridden.
+# Pricing: Token Plan points 40/8/170 per 1M at 0.01 CNY/point,
+# converted at 7.25 CNY/USD.
+
+base_model = "alibaba/qwen3.5-35b-a3b"
+tool_call = false
+attachment = false
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[cost]
+input = 0.06
+output = 0.23
+cache_read = 0.01
+
+[limit]
+context = 128_000
diff --git a/providers/iflytek/models/qwen3.5-397b-a17b.toml b/providers/iflytek/models/qwen3.5-397b-a17b.toml
new file mode 100644
index 00000000000..a7720e6ade3
--- /dev/null
+++ b/providers/iflytek/models/qwen3.5-397b-a17b.toml
@@ -0,0 +1,36 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# Qwen3.5-397B-A17B: 256K context on this host (Token Plan; lab 262_144).
+# Toggle: enable_thinking = true|false per host docs. No budget_tokens
+# (host does not expose thinking_budget; DashScope-specific). No graded
+# effort (host "仅针对OpenAI开源的OSS模型" phrase targets OpenAI OSS
+# gpt-oss family, not Qwen; matching toggle-only peer baseline from
+# TokenGo and similar relays).
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Text-only on this API; base lab is multimodal, attachment/modalities
+# overridden.
+# Pricing: Token Plan points 120/24/720 per 1M at 0.01 CNY/point
+# (Token Plan doc defines points as the unified meter; 1 point ≈ 0.01 CNY
+# is the published conversion for standard members at 200 CNY / 20000 points),
+# converted at 7.25 CNY/USD.
+
+base_model = "alibaba/qwen3.5-397b-a17b"
+tool_call = false
+attachment = false
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[cost]
+input = 0.17
+output = 0.99
+cache_read = 0.03
+
+[limit]
+context = 256_000
diff --git a/providers/iflytek/models/qwen3.6-35b-a3b.toml b/providers/iflytek/models/qwen3.6-35b-a3b.toml
new file mode 100644
index 00000000000..278855ef7fa
--- /dev/null
+++ b/providers/iflytek/models/qwen3.6-35b-a3b.toml
@@ -0,0 +1,32 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# Qwen3.6-35B-A3B: 128K context on this host (Token Plan; lab 262_144).
+# Toggle: enable_thinking = true|false per host docs. No budget_tokens
+# (host does not expose thinking_budget). No graded effort (host "OSS"
+# phrase targets OpenAI OSS gpt-oss family, not Qwen; matching
+# toggle-only peer baseline).
+# Tool calling NOT supported on this host. Text-only on this API; base
+# lab is multimodal, attachment/modalities overridden.
+# Pricing: Token Plan points 100/20/360 per 1M at 0.01 CNY/point,
+# converted at 7.25 CNY/USD.
+
+base_model = "alibaba/qwen3.6-35b-a3b"
+tool_call = false
+attachment = false
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[cost]
+input = 0.14
+output = 0.50
+cache_read = 0.03
+
+[limit]
+context = 128_000
diff --git a/providers/iflytek/models/spark-x2-flash.toml b/providers/iflytek/models/spark-x2-flash.toml
new file mode 100644
index 00000000000..269ac84fac4
--- /dev/null
+++ b/providers/iflytek/models/spark-x2-flash.toml
@@ -0,0 +1,26 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# Spark-X2-Flash: 256K context, lightweight first-party iFlytek model.
+# Toggle: enable_thinking = true|false
+# Effort: reasoning_effort = high|max
+# Thinking content returned in message.reasoning_content (interleaved).
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Pricing: estimated from Token Plan point ratios (100/20/200 points per 1M)
+# at 0.01 CNY/point (Token Plan doc: 200 CNY / 20000 points for standard
+# members), converted at 7.25 CNY/USD.
+
+base_model = "iflytek/spark-x2-flash"
+tool_call = false
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["high", "max"] },
+]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.14
+output = 0.28
+cache_read = 0.03
diff --git a/providers/iflytek/models/spark-x2.5-1.7b.toml b/providers/iflytek/models/spark-x2.5-1.7b.toml
new file mode 100644
index 00000000000..4c0c0f718cc
--- /dev/null
+++ b/providers/iflytek/models/spark-x2.5-1.7b.toml
@@ -0,0 +1,27 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://maas.xfyun.cn/modelSquare
+# Spark-X2.5-1.7B: 1M context, open-weight 1.7B ultra-light edge model.
+# Toggle: enable_thinking = true|false per host docs; effort = high|max
+# matching the documented Spark reasoning intensity control (same surface
+# as the flagship Spark-X2.5 / Spark-X2-Flash on this host).
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Pricing: listed as 限时免费 (limited-time free) on MaaS model square;
+# cost is 0 while the free promotion runs.
+
+base_model = "iflytek/spark-x2.5-1.7b"
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0
+output = 0
+cache_read = 0
diff --git a/providers/iflytek/models/spark-x2.5-4b.toml b/providers/iflytek/models/spark-x2.5-4b.toml
new file mode 100644
index 00000000000..772f3a2dd88
--- /dev/null
+++ b/providers/iflytek/models/spark-x2.5-4b.toml
@@ -0,0 +1,28 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://maas.xfyun.cn/modelSquare
+# Spark-X2.5-4B: 1M context, open-weight 4B edge Dense model.
+# Toggle: enable_thinking = true|false per host docs; effort = high|max
+# matching the documented Spark reasoning intensity control (same surface
+# as the flagship Spark-X2.5 / Spark-X2-Flash on this host).
+# Tool calling NOT supported on this host (only DeepSeek V3.2 and GLM-4.7).
+# Pricing: listed as 限时免费 (limited-time free) on MaaS model square;
+# cost is 0 while the free promotion runs.
+
+base_model = "iflytek/spark-x2.5-4b"
+tool_call = false
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0
+output = 0
+cache_read = 0
diff --git a/providers/iflytek/models/spark-x2.5.toml b/providers/iflytek/models/spark-x2.5.toml
new file mode 100644
index 00000000000..ec589ad9d0f
--- /dev/null
+++ b/providers/iflytek/models/spark-x2.5.toml
@@ -0,0 +1,27 @@
+# Sources (accessed 2026-09-14):
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# Spark-X2.5: 256K context, first-party iFlytek flagship (293B-A30B MoE).
+# Toggle: enable_thinking = true|false
+# Effort: reasoning_effort = high|max
+# Thinking content returned in message.reasoning_content (interleaved).
+# Pricing: list price CNY 3.2 input / 0.48 cache_read / 12 output per 1M
+# tokens → USD 0.44 / 0.07 / 1.65 at 7.25 CNY/USD. Note: a limited-time
+# 50% promotion (as of 2026-09) halves these to CNY 1.6/0.24/6; cost above
+# reflects the list price since the promotion end date is unpublished.
+# Tool calling not supported per host docs.
+
+base_model = "iflytek/spark-x2.5"
+tool_call = false
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["high", "max"] },
+]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.44
+output = 1.65
+cache_read = 0.07
diff --git a/providers/iflytek/provider.toml b/providers/iflytek/provider.toml
new file mode 100644
index 00000000000..44c0609f4b3
--- /dev/null
+++ b/providers/iflytek/provider.toml
@@ -0,0 +1,40 @@
+name = "iFlytek"
+npm = "@ai-sdk/openai-compatible"
+env = ["IFLYTEK_API_KEY"]
+# iFlytek MaaS Inference API (讯飞星辰MaaS推理服务).
+# OpenAI-compatible chat completions at /v2/chat/completions.
+# Text-only endpoint — base lab multimodal capabilities (image/video/audio
+# input on Kimi, Qwen) are not exposed on this API.
+#
+# Reasoning controls per model family (from host docs, accessed 2026-09-14):
+# - DeepSeek V4 Pro: thinking.type toggle (enabled|disabled) + effort
+#   high|max (host-specific intensity control via reasoning_effort or
+#   output_config.effort).
+# - DeepSeek V4 Flash: thinking.type toggle + effort low|high|max.
+# - DeepSeek V3.2: thinking.type toggle only (hybrid model, matches
+#   established relay baseline from OpenRouter/TokenGo/Novita).
+# - GLM-5.2: effort high|max only (only GLM on this host with graded effort).
+# - GLM-5.1, GLM-5: enable_thinking toggle only (no graded effort).
+# - GLM-4.7-Flash: enable_thinking toggle (matches first-party zhipuai
+#   baseline and empiriolabs/poe peers; host docs: enable_thinking bool
+#   applies to thinking models, GLM-4.7-Flash is a thinking model).
+# - Kimi K2.5/K2.6: enable_thinking toggle only.
+# - MiniMax M2.5: always-on reasoning, no caller control ([]).
+# - Qwen 397B/3.6-35B/3.5-35B: enable_thinking toggle only (host "OSS"
+#   phrase targets OpenAI OSS gpt-oss, not Qwen; no budget_tokens, no
+#   graded effort).
+# - Qwen3-Coder-Next: non-reasoning.
+# - Spark-X2.5/X2-Flash (first-party): enable_thinking toggle + effort high|max.
+# - Spark-X2.5-4B/1.7B (open-weight): enable_thinking toggle + effort high|max
+#   (same Spark surface as flagship).
+#
+# All reasoning models return message.reasoning_content (interleaved).
+# Tool calling: only DeepSeek V3.2 and GLM-4.7 series per host docs.
+# JSON mode: response_format = {"type": "json_object"} on DeepSeek, GLM, Kimi,
+# Qwen series; set search_disable=true recommended under JSON mode.
+#
+# Sources:
+# https://www.xfyun.cn/doc/spark/推理服务-http.html
+# https://www.xfyun.cn/doc/spark/TokenPlan.html
+doc = "https://www.xfyun.cn/doc/spark/%E6%8E%A8%E7%90%86%E6%9C%8D%E5%8A%A1-http.html"
+api = "https://maas-api.cn-huabei-1.xf-yun.com/v2"