feat(melious): add Melious provider with the GLM, DeepSeek and Kimi families

providers/melious/models/deepseek-r1-0528.toml

@@ -0,0 +1,31 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 0.60 in / EUR 2.40
+# out / EUR 0.16 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/deepseek-r1-0528 (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget. Measured against this
+# deployment on 2026-09-14, it has no effect on this model: reasoning_effort="none"
+# still returns reasoning (reasoning_content 2_786/3_010 chars across repeat samples),
+# and low/medium/high produce no ordered difference. reasoning_options is therefore []
+# -- this model reasons, with no caller control on this host. That is an affirmative
+# measurement, not an untested default.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+# limit.context: the catalog reports 164_000 for this deployment, not the lab's
+# 128_000.
+base_model = "deepseek/deepseek-r1"
+reasoning_options = []
+
+[cost]
+input = 0.69552
+output = 2.78208
+cache_read = 0.185472
+
+[limit]
+context = 164_000
+
+[interleaved]
+field = "reasoning_content"

providers/melious/models/deepseek-v3.2.toml

@@ -0,0 +1,33 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 0.30 in / EUR 0.50
+# out / EUR 0.08 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/deepseek-v3.2 (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget, so no toggle or
+# budget_tokens. Measured against this deployment on 2026-09-14:
+# reasoning_effort="none" is a real off state (usage.reasoning_tokens = 0 and no
+# reasoning_content), while low/medium/high all return reasoning with no ordered
+# difference in depth, so they are not published as separate grades. The host also
+# accepts undocumented values without a 4xx, so acceptance is not evidence. Because
+# off is reachable through the effort field this is an effort list containing none,
+# not a toggle.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+# limit.context: the catalog reports 164_000 for this deployment, not the lab's
+# 128_000.
+base_model = "deepseek/deepseek-v3.2"
+reasoning_options = [{ type = "effort", values = ["none", "high"] }]
+
+[cost]
+input = 0.34776
+output = 0.5796
+cache_read = 0.092736
+
+[limit]
+context = 164_000
+
+[interleaved]
+field = "reasoning_content"

providers/melious/models/deepseek-v4-flash-0731.toml

@@ -0,0 +1,28 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 0.10 in / EUR 0.25
+# out / EUR 0.02 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/deepseek-v4-flash-0731 (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget, so no toggle or
+# budget_tokens. Measured against this deployment on 2026-09-14:
+# reasoning_effort="none" is a real off state (usage.reasoning_tokens = 0 and no
+# reasoning_content), while low/medium/high all return reasoning with no ordered
+# difference in depth, so they are not published as separate grades. The host also
+# accepts undocumented values without a 4xx, so acceptance is not evidence. Because
+# off is reachable through the effort field this is an effort list containing none,
+# not a toggle.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+base_model = "deepseek/deepseek-v4-flash-0731"
+reasoning_options = [{ type = "effort", values = ["none", "high"] }]
+
+[cost]
+input = 0.11592
+output = 0.2898
+cache_read = 0.023184
+
+[interleaved]
+field = "reasoning_content"

providers/melious/models/deepseek-v4-pro-0813.toml

@@ -0,0 +1,28 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 1.00 in / EUR 3.00
+# out / EUR 0.10 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/deepseek-v4-pro-0813 (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget, so no toggle or
+# budget_tokens. Measured against this deployment on 2026-09-14:
+# reasoning_effort="none" is a real off state (usage.reasoning_tokens = 0 and no
+# reasoning_content), while low/medium/high all return reasoning with no ordered
+# difference in depth, so they are not published as separate grades. The host also
+# accepts undocumented values without a 4xx, so acceptance is not evidence. Because
+# off is reachable through the effort field this is an effort list containing none,
+# not a toggle.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+base_model = "deepseek/deepseek-v4-pro-0813"
+reasoning_options = [{ type = "effort", values = ["none", "high"] }]
+
+[cost]
+input = 1.1592
+output = 3.4776
+cache_read = 0.11592
+
+[interleaved]
+field = "reasoning_content"

providers/melious/models/deepseek-v4-pro.toml

@@ -0,0 +1,28 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 1.60 in / EUR 3.20
+# out / EUR 0.40 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/deepseek-v4-pro (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget, so no toggle or
+# budget_tokens. Measured against this deployment on 2026-09-14:
+# reasoning_effort="none" is a real off state (usage.reasoning_tokens = 0 and no
+# reasoning_content), while low/medium/high all return reasoning with no ordered
+# difference in depth, so they are not published as separate grades. The host also
+# accepts undocumented values without a 4xx, so acceptance is not evidence. Because
+# off is reachable through the effort field this is an effort list containing none,
+# not a toggle.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+base_model = "deepseek/deepseek-v4-pro"
+reasoning_options = [{ type = "effort", values = ["none", "high"] }]
+
+[cost]
+input = 1.85472
+output = 3.70944
+cache_read = 0.46368
+
+[interleaved]
+field = "reasoning_content"

providers/melious/models/deepseek-v4.1-flash.toml

@@ -0,0 +1,28 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 0.20 in / EUR 1.00
+# out / EUR 0.01 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/deepseek-v4.1-flash (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget, so no toggle or
+# budget_tokens. Measured against this deployment on 2026-09-14:
+# reasoning_effort="none" is a real off state (usage.reasoning_tokens = 0 and no
+# reasoning_content), while low/medium/high all return reasoning with no ordered
+# difference in depth, so they are not published as separate grades. The host also
+# accepts undocumented values without a 4xx, so acceptance is not evidence. Because
+# off is reachable through the effort field this is an effort list containing none,
+# not a toggle.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+base_model = "deepseek/deepseek-v4.1-flash"
+reasoning_options = [{ type = "effort", values = ["none", "high"] }]
+
+[cost]
+input = 0.23184
+output = 1.1592
+cache_read = 0.011592
+
+[interleaved]
+field = "reasoning_content"

providers/melious/models/glm-5.1.toml

@@ -0,0 +1,33 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 1.30 in / EUR 4.00
+# out / EUR 0.32 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/glm-5.1 (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget, so no toggle or
+# budget_tokens. Measured against this deployment on 2026-09-14:
+# reasoning_effort="none" is a real off state (usage.reasoning_tokens = 0 and no
+# reasoning_content), while low/medium/high all return reasoning with no ordered
+# difference in depth, so they are not published as separate grades. The host also
+# accepts undocumented values without a 4xx, so acceptance is not evidence. Because
+# off is reachable through the effort field this is an effort list containing none,
+# not a toggle.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+# limit.context: the catalog reports 203_000 for this deployment, not the lab's
+# 200_000.
+base_model = "zhipuai/glm-5.1"
+reasoning_options = [{ type = "effort", values = ["none", "high"] }]
+
+[cost]
+input = 1.50696
+output = 4.6368
+cache_read = 0.370944
+
+[limit]
+context = 203_000
+
+[interleaved]
+field = "reasoning_content"

providers/melious/models/glm-5.2.toml

@@ -0,0 +1,28 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 1.00 in / EUR 4.00
+# out / EUR 0.25 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/glm-5.2 (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget, so no toggle or
+# budget_tokens. Measured against this deployment on 2026-09-14:
+# reasoning_effort="none" is a real off state (usage.reasoning_tokens = 0 and no
+# reasoning_content), while low/medium/high all return reasoning with no ordered
+# difference in depth, so they are not published as separate grades. The host also
+# accepts undocumented values without a 4xx, so acceptance is not evidence. Because
+# off is reachable through the effort field this is an effort list containing none,
+# not a toggle.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+base_model = "zhipuai/glm-5.2"
+reasoning_options = [{ type = "effort", values = ["none", "high"] }]
+
+[cost]
+input = 1.1592
+output = 4.6368
+cache_read = 0.2898
+
+[interleaved]
+field = "reasoning_content"

providers/melious/models/glm-5.3-flash.toml

@@ -0,0 +1,33 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 0.10 in / EUR 0.40
+# out / EUR 0.02 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/glm-5.3-flash (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget, so no toggle or
+# budget_tokens. Measured against this deployment on 2026-09-14:
+# reasoning_effort="none" is a real off state (usage.reasoning_tokens = 0 and no
+# reasoning_content), while low/medium/high all return reasoning with no ordered
+# difference in depth, so they are not published as separate grades. The host also
+# accepts undocumented values without a 4xx, so acceptance is not evidence. Because
+# off is reachable through the effort field this is an effort list containing none,
+# not a toggle.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+# modalities.input: this host routes ["text", "image", "video"], narrower than the
+# lab's ["text", "image", "video", "pdf"].
+base_model = "zhipuai/glm-5.3-flash"
+reasoning_options = [{ type = "effort", values = ["none", "high"] }]
+
+[cost]
+input = 0.11592
+output = 0.46368
+cache_read = 0.023184
+
+[modalities]
+input = ["text", "image", "video"]
+
+[interleaved]
+field = "reasoning_content"

providers/melious/models/glm-5.3.toml

@@ -0,0 +1,28 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 1.00 in / EUR 3.00
+# out / EUR 0.20 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/glm-5.3 (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget, so no toggle or
+# budget_tokens. Measured against this deployment on 2026-09-14:
+# reasoning_effort="none" is a real off state (usage.reasoning_tokens = 0 and no
+# reasoning_content), while low/medium/high all return reasoning with no ordered
+# difference in depth, so they are not published as separate grades. The host also
+# accepts undocumented values without a 4xx, so acceptance is not evidence. Because
+# off is reachable through the effort field this is an effort list containing none,
+# not a toggle.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+base_model = "zhipuai/glm-5.3"
+reasoning_options = [{ type = "effort", values = ["none", "high"] }]
+
+[cost]
+input = 1.1592
+output = 3.4776
+cache_read = 0.23184
+
+[interleaved]
+field = "reasoning_content"

providers/melious/models/glm-5.toml

@@ -0,0 +1,31 @@
+# Cost: Melious list price for the default `balanced` route -- EUR 0.95 in / EUR 2.90
+# out / EUR 0.23 cache read per 1M tokens -- converted at 1.1592 USD/EUR (ECB euro
+# reference rate, 2026-09-11).
+# Sources: GET https://api.melious.ai/v1/models?include_meta=true and
+# https://melious.ai/hub/models/glm-5 (both accessed 2026-09-14)
+# FX: https://www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/eurofxref-graph-usd.en.html
+# Effort: reasoning_effort is the only reasoning control this host exposes, and it
+# forwards no lab-native thinking field and no reasoning budget. Measured against this
+# deployment on 2026-09-14, it has no effect on this model: reasoning_effort="none"
+# still returns reasoning (reasoning_content 2_444/2_239/686/3_907 chars across repeat
+# samples), and low/medium/high produce no ordered difference. reasoning_options is
+# therefore [] -- this model reasons, with no caller control on this host. That is an
+# affirmative measurement, not an untested default.
+# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14)
+# Reasoning text is returned in choices[].message.reasoning_content. cache_read:
+# usage.cached_tokens are billed at Melious's discounted prefix-cache rate.
+# limit.context: the catalog reports 203_000 for this deployment, not the lab's
+# 204_800.
+base_model = "zhipuai/glm-5"
+reasoning_options = []
+
+[cost]
+input = 1.10124
+output = 3.36168
+cache_read = 0.266616
+
+[limit]
+context = 203_000
+
+[interleaved]
+field = "reasoning_content"