diff --git a/providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash-fast.toml b/providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash-fast.toml new file mode 100644 index 0000000000..6798d718e9 --- /dev/null +++ b/providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash-fast.toml @@ -0,0 +1,58 @@ +# Sources (checked 2026-11-16): +# - Public aliases, admission budget, output cap, tools, effort and request fields: +# https://oathbreaker.ai/docs +# - USD per million tokens: https://oathbreaker.ai/pricing +# - Public availability and current limits/prices: https://oathbreaker.ai/llms.txt +# Oathbreaker serves an abliterated variant of DeepSeek V4.1 Flash 1017 through its +# own gateway; it does not train the underlying DeepSeek model. +# The exact served derivative is not published as downloadable Oathbreaker weights. +# Lab weights and benchmarks are never inherited by provider entries; the canonical +# model link identifies the DeepSeek V4.1 Flash 1017 lineage, not equivalent results. +# release_date is when this alias first appeared in Oathbreaker's public llms.txt +# (observed 2026-11-16), not the base model launch. +# Limits: a request is admitted when UTF-8 input bytes, message and tool framing and +# the requested output tokens fit in a 131,072-unit budget, and the request body is at +# most 262,144 bytes (limit.request_bytes). The budget counts bytes, not tokens, so +# limit.context is a conservative token estimate: 131,072 / ~4 bytes per token of typical +# English text = 32,768. Code or non-English text can use more bytes per token. +# Effort: reasoning_effort = low|high|max. If the field is omitted, the gateway sends no +# effort parameter and the upstream model default applies; no default is documented. +# "none" is not an accepted value, and no thinking toggle field is accepted. +# max_tokens and max_completion_tokens are supported (default 8,192); output includes reasoning. +# Cache reads are implicit and billed only when reported by upstream usage. +# Chat Completions only; no image input, Responses, Messages or strict tool schemas. +# This alias serves the same model, limits and prices as oathbreaker-deepseek-v4.1-flash +# on a separate serving route. +schema = 2 +base_model = "deepseek/deepseek-v4.1-flash-1017" +name = "Oathbreaker DeepSeek V4.1 Flash Abliterated Fast" +description = "DeepSeek V4.1 Flash 1017 Abliterated on a separate Oathbreaker serving route, through its text-only OpenAI-compatible Chat Completions API" +open_weights = false +release_date = "2026-11-16" + +[modalities] +input = ["text"] +output = ["text"] + +[capabilities] +tools = { supported = true, choice = ["auto", "none", "required", "tool"] } +reasoning = { supported = true, effort = ["low", "high", "max"] } +structured_output = false +top_k = false +cache = { implicit = true } +stream_usage = true + +[limit] +context = 32_768 +output = 32_768 +request_bytes = 262_144 + +[api.chat-completions] +compatibility = { reasoning_field = "reasoning_content" } + +[cost.text.input] +price = 0.50 +cache = { read = 0.05 } + +[cost.text.output] +price = 1.50 diff --git a/providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash.toml b/providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash.toml new file mode 100644 index 0000000000..6203e16463 --- /dev/null +++ b/providers/oathbreaker/models/oathbreaker-deepseek-v4.1-flash.toml @@ -0,0 +1,56 @@ +# Sources (checked 2026-11-16): +# - Public aliases, admission budget, output cap, tools, effort and request fields: +# https://oathbreaker.ai/docs +# - USD per million tokens: https://oathbreaker.ai/pricing +# - Public availability and current limits/prices: https://oathbreaker.ai/llms.txt +# Oathbreaker serves an abliterated variant of DeepSeek V4.1 Flash 1017 through its +# own gateway; it does not train the underlying DeepSeek model. +# The exact served derivative is not published as downloadable Oathbreaker weights. +# Lab weights and benchmarks are never inherited by provider entries; the canonical +# model link identifies the DeepSeek V4.1 Flash 1017 lineage, not equivalent results. +# release_date is when this alias first appeared in Oathbreaker's public llms.txt +# (observed 2026-11-16), not the base model launch. +# Limits: a request is admitted when UTF-8 input bytes, message and tool framing and +# the requested output tokens fit in a 131,072-unit budget, and the request body is at +# most 262,144 bytes (limit.request_bytes). The budget counts bytes, not tokens, so +# limit.context is a conservative token estimate: 131,072 / ~4 bytes per token of typical +# English text = 32,768. Code or non-English text can use more bytes per token. +# Effort: reasoning_effort = low|high|max. If the field is omitted, the gateway sends no +# effort parameter and the upstream model default applies; no default is documented. +# "none" is not an accepted value, and no thinking toggle field is accepted. +# max_tokens and max_completion_tokens are supported (default 8,192); output includes reasoning. +# Cache reads are implicit and billed only when reported by upstream usage. +# Chat Completions only; no image input, Responses, Messages or strict tool schemas. +schema = 2 +base_model = "deepseek/deepseek-v4.1-flash-1017" +name = "Oathbreaker DeepSeek V4.1 Flash Abliterated" +description = "DeepSeek V4.1 Flash 1017 Abliterated served through Oathbreaker's text-only OpenAI-compatible Chat Completions API" +open_weights = false +release_date = "2026-11-16" + +[modalities] +input = ["text"] +output = ["text"] + +[capabilities] +tools = { supported = true, choice = ["auto", "none", "required", "tool"] } +reasoning = { supported = true, effort = ["low", "high", "max"] } +structured_output = false +top_k = false +cache = { implicit = true } +stream_usage = true + +[limit] +context = 32_768 +output = 32_768 +request_bytes = 262_144 + +[api.chat-completions] +compatibility = { reasoning_field = "reasoning_content" } + +[cost.text.input] +price = 0.50 +cache = { read = 0.05 } + +[cost.text.output] +price = 1.50