diff --git a/providers/melious/logo.svg b/providers/melious/logo.svg new file mode 100644 index 00000000000..a2ff326807a --- /dev/null +++ b/providers/melious/logo.svg @@ -0,0 +1,3 @@ + + + diff --git a/providers/melious/models/deepseek-r1-0528.toml b/providers/melious/models/deepseek-r1-0528.toml new file mode 100644 index 00000000000..78d2c96463c --- /dev/null +++ b/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" diff --git a/providers/melious/models/deepseek-v3.2.toml b/providers/melious/models/deepseek-v3.2.toml new file mode 100644 index 00000000000..87e4436bc09 --- /dev/null +++ b/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" diff --git a/providers/melious/models/deepseek-v4-flash-0731.toml b/providers/melious/models/deepseek-v4-flash-0731.toml new file mode 100644 index 00000000000..6f5c2d8fa23 --- /dev/null +++ b/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" diff --git a/providers/melious/models/deepseek-v4-pro-0813.toml b/providers/melious/models/deepseek-v4-pro-0813.toml new file mode 100644 index 00000000000..0cdb0c82645 --- /dev/null +++ b/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" diff --git a/providers/melious/models/deepseek-v4-pro.toml b/providers/melious/models/deepseek-v4-pro.toml new file mode 100644 index 00000000000..ca2515e3ebf --- /dev/null +++ b/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" diff --git a/providers/melious/models/deepseek-v4.1-flash.toml b/providers/melious/models/deepseek-v4.1-flash.toml new file mode 100644 index 00000000000..679d9879d8b --- /dev/null +++ b/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" diff --git a/providers/melious/models/glm-5.1.toml b/providers/melious/models/glm-5.1.toml new file mode 100644 index 00000000000..63766b09ff8 --- /dev/null +++ b/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" diff --git a/providers/melious/models/glm-5.2.toml b/providers/melious/models/glm-5.2.toml new file mode 100644 index 00000000000..8ca4dbe4c3e --- /dev/null +++ b/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" diff --git a/providers/melious/models/glm-5.3-flash.toml b/providers/melious/models/glm-5.3-flash.toml new file mode 100644 index 00000000000..d25a740b3f3 --- /dev/null +++ b/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" diff --git a/providers/melious/models/glm-5.3.toml b/providers/melious/models/glm-5.3.toml new file mode 100644 index 00000000000..03eac7197c0 --- /dev/null +++ b/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" diff --git a/providers/melious/models/glm-5.toml b/providers/melious/models/glm-5.toml new file mode 100644 index 00000000000..19ac9f0e728 --- /dev/null +++ b/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" diff --git a/providers/melious/models/kimi-k2.5.toml b/providers/melious/models/kimi-k2.5.toml new file mode 100644 index 00000000000..fabc024c1a0 --- /dev/null +++ b/providers/melious/models/kimi-k2.5.toml @@ -0,0 +1,38 @@ +# Cost: Melious list price for the default `balanced` route -- EUR 0.50 in / EUR 2.55 +# out / EUR 0.12 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/kimi-k2.5 (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 1_177/1_714/1_100 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 262_000 for this deployment, not the lab's +# 262_144. limit.output: max_output_tokens is null for this model, so output is capped +# at the served context rather than inheriting the lab's larger 262_144. +# modalities.input: this host routes ["text", "image"], narrower than the lab's +# ["text", "image", "video"]. +base_model = "moonshotai/kimi-k2.5" +reasoning_options = [] + +[cost] +input = 0.5796 +output = 2.95596 +cache_read = 0.139104 + +[limit] +context = 262_000 +output = 262_000 + +[modalities] +input = ["text", "image"] + +[interleaved] +field = "reasoning_content" diff --git a/providers/melious/models/kimi-k2.6.toml b/providers/melious/models/kimi-k2.6.toml new file mode 100644 index 00000000000..4784a6bc4c8 --- /dev/null +++ b/providers/melious/models/kimi-k2.6.toml @@ -0,0 +1,35 @@ +# Cost: Melious list price for the default `balanced` route -- EUR 0.70 in / EUR 3.50 +# 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/kimi-k2.6 (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 256_000 for this deployment, not the lab's +# 262_144. limit.output: max_output_tokens is null for this model, so output is capped +# at the served context rather than inheriting the lab's larger 262_144. +base_model = "moonshotai/kimi-k2.6" +reasoning_options = [{ type = "effort", values = ["none", "high"] }] + +[cost] +input = 0.81144 +output = 4.0572 +cache_read = 0.266616 + +[limit] +context = 256_000 +output = 256_000 + +[interleaved] +field = "reasoning_content" diff --git a/providers/melious/models/kimi-k2.7-code.toml b/providers/melious/models/kimi-k2.7-code.toml new file mode 100644 index 00000000000..a38cc7a360b --- /dev/null +++ b/providers/melious/models/kimi-k2.7-code.toml @@ -0,0 +1,28 @@ +# Cost: Melious list price for the default `balanced` route -- EUR 0.70 in / EUR 3.00 +# out / EUR 0.19 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/kimi-k2.7-code (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 = "moonshotai/kimi-k2.7-code" +reasoning_options = [{ type = "effort", values = ["none", "high"] }] + +[cost] +input = 0.81144 +output = 3.4776 +cache_read = 0.220248 + +[interleaved] +field = "reasoning_content" diff --git a/providers/melious/models/kimi-k3.toml b/providers/melious/models/kimi-k3.toml new file mode 100644 index 00000000000..6f9d4f8021b --- /dev/null +++ b/providers/melious/models/kimi-k3.toml @@ -0,0 +1,33 @@ +# Cost: Melious list price for the default `balanced` route -- EUR 2.75 in / EUR 13.75 +# out / EUR 0.68 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/kimi-k3 (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 1_000_000 for this deployment, not the lab's +# 1_048_576. +base_model = "moonshotai/kimi-k3" +reasoning_options = [{ type = "effort", values = ["none", "high"] }] + +[cost] +input = 3.1878 +output = 15.939 +cache_read = 0.788256 + +[limit] +context = 1_000_000 + +[interleaved] +field = "reasoning_content" diff --git a/providers/melious/provider.toml b/providers/melious/provider.toml new file mode 100644 index 00000000000..db50ccc00f3 --- /dev/null +++ b/providers/melious/provider.toml @@ -0,0 +1,33 @@ +# Melious AI GmbH relays open-weight models built by other labs on European +# infrastructure. It is not the lab for any model it serves, so every model entry +# here uses base_model. Raw HTTP is POST /v1/chat/completions; the OpenAI SDK works +# by pointing base_url at the API. +# Reasoning: the only control on this surface is `reasoning_effort`. No lab-native +# field is forwarded -- no thinking object, no enable_thinking, no reasoning budget -- +# so no entry here uses `toggle` or `budget_tokens`. The host also accepts any effort +# string without a 4xx, including undocumented ones, so acceptance is not evidence. +# Every model here was measured against this deployment on 2026-09-14 instead. Two +# behaviours were found, and neither matches the vendor docs' low|medium|high ladder: +# on 12 models reasoning_effort="none" is a real off state (usage.reasoning_tokens = 0 +# and no reasoning_content) while low, medium and high all reason with no ordered +# difference in depth -- those are authored as effort ["none", "high"], an effort list +# containing none rather than a toggle, because off is reachable through this field. +# On glm-5, deepseek-r1-0528 and kimi-k2.5, "none" does not suppress reasoning and no +# level grades, so those carry reasoning_options = [] -- an affirmative measurement of +# no caller control, not an untested default. The catalog's own reasoning_type is not +# a reliable guide: glm-5.3-flash is typed always-on yet honours "none", and glm-5 is +# typed hybrid yet ignores it. Reasoning text is returned in +# choices[].message.reasoning_content. +# https://melious.ai/docs/reference/chat-completions (accessed 2026-09-14) +# Routing: a flavour is selected with a model-ID suffix (`:balanced`, `:speed`, +# `:price`, `:eco`, `:batch`). A bare chat model ID resolves to `balanced`, which is +# what the ids and prices here describe; only bare ids are published. A `preset` +# field biases routing, not reasoning depth. +# https://melious.ai/docs/concepts/routing (accessed 2026-09-14) +# Caching: usage.cached_tokens are prompt tokens served from the prefix cache and +# are billed at a cheaper cache-read rate, recorded as cost.cache_read. +name = "Melious" +npm = "@ai-sdk/openai-compatible" +api = "https://api.melious.ai/v1" +env = ["MELIOUS_API_KEY"] +doc = "https://melious.ai/docs/reference/models"