feat(providers): add iFlytek MaaS (China) provider

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"]

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"

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"

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"]

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);

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

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

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

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

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

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

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

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

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

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

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