harn-vm 0.8.168

# DashScope (Alibaba Model Studio, international endpoint) — first-party
# Qwen routes. Pricing verified 2026-07-02 against
# https://www.alibabacloud.com/help/en/model-studio/model-pricing.
#
# Rows are keyed `dashscope/<wire-id>` so they stay collision-free with
# the same weights hosted elsewhere (e.g. the bare `qwen3.6-35b-a3b`
# llamacpp row); `wire_model` records the exact id sent on the wire.
#
# DashScope bills per request at the tier the TOTAL input size lands in
# (<=32K / <=128K / <=256K / <=1M); this catalog's single-rate-card shape
# records the base <=32K tier, so long-context requests cost more than a
# naive rate-card estimate. Implicit prompt caching bills cache hits at
# 20% of the input rate (explicit caches at 10%); cache_read below uses
# the implicit 20% figure.

[models."dashscope/qwen3-coder-plus"]
name = "Qwen3 Coder Plus (DashScope)"
provider = "dashscope"
wire_model = "qwen3-coder-plus"
served_variant = "dashscope"
api_dialect = "openai_chat"
context_window = 1048576
capabilities = ["tools", "streaming", "prompt_caching"]
pricing = { input_per_mtok = 1.00, output_per_mtok = 5.00, cache_read_per_mtok = 0.20 }
tier = "frontier"
open_weight = false
strengths = ["coding", "agentic", "long_context", "tool_use"]

[models."dashscope/qwen3-coder-next"]
name = "Qwen3 Coder Next (DashScope)"
provider = "dashscope"
wire_model = "qwen3-coder-next"
logical_model = "qwen3-coder-next"
equivalence_group = "qwen3-coder-next"
served_variant = "dashscope"
api_dialect = "openai_chat"
context_window = 262144
capabilities = ["tools", "streaming", "prompt_caching"]
pricing = { input_per_mtok = 0.30, output_per_mtok = 1.50, cache_read_per_mtok = 0.06 }
architecture = { parameter_count_b = 80.0, active_parameter_count_b = 3.0, moe = true, license = "Apache-2.0", source_url = "https://huggingface.co/Qwen/Qwen3-Coder-Next", last_verified = "2026-07-02" }
benchmarks = { swe_bench_verified = 70.6, swe_bench_pro = 44.3 }
tier = "mid"
open_weight = true
strengths = ["coding", "cheap", "long_context", "agentic", "tool_use"]

[models."dashscope/qwen3.6-35b-a3b"]
name = "Qwen3.6 35B A3B (DashScope)"
provider = "dashscope"
wire_model = "qwen3.6-35b-a3b"
logical_model = "qwen3.6-35b-a3b"
equivalence_group = "qwen3.6-35b-a3b"
served_variant = "dashscope"
api_dialect = "openai_chat"
context_window = 262144
capabilities = ["tools", "vision", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.375, output_per_mtok = 2.25, cache_read_per_mtok = 0.075 }
tier = "mid"
open_weight = true
strengths = ["coding", "agentic", "tool_use", "long_context", "cheap"]

[models."dashscope/qwen3.5-397b-a17b"]
name = "Qwen3.5 397B A17B (DashScope)"
provider = "dashscope"
wire_model = "qwen3.5-397b-a17b"
served_variant = "dashscope"
api_dialect = "openai_chat"
context_window = 262144
capabilities = ["tools", "vision", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.60, output_per_mtok = 3.60, cache_read_per_mtok = 0.12 }
architecture = { parameter_count_b = 397.0, active_parameter_count_b = 17.0, moe = true, license = "Apache-2.0", source_url = "https://huggingface.co/Qwen/Qwen3.5-397B-A17B", last_verified = "2026-07-02" }
benchmarks = { swe_bench_verified = 76.4 }
tier = "frontier"
open_weight = true
strengths = ["coding", "reasoning", "tool_use", "long_context", "vision"]

# Qwen3.7-Max — Alibaba's closed frontier tier; flat (non-tiered) $2.50 /
# $7.50 per MTok with a 1M window on DashScope.
[models."dashscope/qwen3.7-max"]
name = "Qwen3.7 Max (DashScope)"
provider = "dashscope"
wire_model = "qwen3.7-max"
logical_model = "qwen3.7-max"
equivalence_group = "qwen3.7-max"
served_variant = "dashscope"
api_dialect = "openai_chat"
context_window = 1000000
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 2.50, output_per_mtok = 7.50, cache_read_per_mtok = 0.50 }
tier = "frontier"
open_weight = false
strengths = ["coding", "agentic", "tool_use", "long_context", "reasoning"]