From 7f8ce0db4e383074547da27a14393ee270d37b40 Mon Sep 17 00:00:00 2001 From: Li Tai Chen Date: Wed, 5 Aug 2026 14:16:38 -0400 Subject: [PATCH 1/4] feat(models): add fireworks deepseek-v4-pro mapping Adds Fireworks as a provider for DeepSeek V4 Pro (accounts/fireworks/models/deepseek-v4-pro), 1.74 / 3.48 per 1M tokens with 0.145 cached input, up to 1M context. Co-Authored-By: Claude --- packages/models/src/models/deepseek.ts | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) diff --git a/packages/models/src/models/deepseek.ts b/packages/models/src/models/deepseek.ts index afc45c0d16..d3cc8cc1f8 100644 --- a/packages/models/src/models/deepseek.ts +++ b/packages/models/src/models/deepseek.ts @@ -408,6 +408,25 @@ export const deepseekModels = [ tools: true, jsonOutput: true, }, + { + providerId: "fireworks", + externalId: "accounts/fireworks/models/deepseek-v4-pro", + inputPrice: "1.74e-6", + cachedInputPrice: "0.145e-6", + outputPrice: "3.48e-6", + requestPrice: "0", + serviceTiers: ["priority"], + serviceTierMultipliers: { priority: 1.5 }, + contextSize: 1048576, + maxOutput: 393216, + streaming: true, + reasoning: true, + reasoningEfforts: ["none", "low", "medium", "high", "xhigh", "max"], + vision: false, + tools: true, + jsonOutput: true, + jsonOutputSchema: true, + }, ], }, { From adb923c109147b60caf3612890627b0d6b960c3a Mon Sep 17 00:00:00 2001 From: Li Tai Chen Date: Wed, 5 Aug 2026 14:18:38 -0400 Subject: [PATCH 2/4] feat(models): add together-ai deepseek-v4-flash mapping Adds Together AI as a provider for DeepSeek V4 Flash (deepseek-ai/DeepSeek-V4-Flash-0731), 0.14 / 0.28 per 1M tokens with 0.03 cached input. Co-Authored-By: Claude --- packages/models/src/models/deepseek.ts | 18 ++++++++++++++++++ 1 file changed, 18 insertions(+) diff --git a/packages/models/src/models/deepseek.ts b/packages/models/src/models/deepseek.ts index d3cc8cc1f8..116a3259eb 100644 --- a/packages/models/src/models/deepseek.ts +++ b/packages/models/src/models/deepseek.ts @@ -594,6 +594,24 @@ export const deepseekModels = [ jsonOutput: true, jsonOutputSchema: true, }, + { + providerId: "together-ai", + externalId: "deepseek-ai/DeepSeek-V4-Flash-0731", + inputPrice: "0.14e-6", + cachedInputPrice: "0.03e-6", + outputPrice: "0.28e-6", + requestPrice: "0", + contextSize: 163840, + maxOutput: 163840, + streaming: true, + reasoning: true, + reasoningEfforts: ["none", "xhigh", "max"], + requiresDisableThinkingParam: true, + reasoningOutput: "omit", + vision: false, + tools: false, + jsonOutput: false, + }, ], }, ] as const satisfies ModelDefinition[]; From 0c64953c7373487124a38284c95ec85d43e5c83c Mon Sep 17 00:00:00 2001 From: smakosh Date: Sat, 8 Aug 2026 14:01:49 +0200 Subject: [PATCH 3/4] docs(models): explain fireworks priority multiplier Claude-Session: https://claude.ai/code/session_013DqjyWB5mkpqYzEavYCvHz --- packages/models/src/models/deepseek.ts | 2 ++ 1 file changed, 2 insertions(+) diff --git a/packages/models/src/models/deepseek.ts b/packages/models/src/models/deepseek.ts index f1a4adb075..5d936e4ae1 100644 --- a/packages/models/src/models/deepseek.ts +++ b/packages/models/src/models/deepseek.ts @@ -425,6 +425,8 @@ export const deepseekModels = [ cachedInputPrice: "0.145e-6", outputPrice: "3.48e-6", requestPrice: "0", + // Fireworks prices DeepSeek's Priority tier at 1.5x standard rather + // than the 1.25x that applies to the rest of its catalogue. serviceTiers: ["priority"], serviceTierMultipliers: { priority: 1.5 }, contextSize: 1048576, From ed7c8f942a0b8cd8e6c44909be8895dc409b0a92 Mon Sep 17 00:00:00 2001 From: Claude Date: Sat, 8 Aug 2026 13:57:11 +0000 Subject: [PATCH 4/4] fix(models): correct together v4 flash caps MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Together's DeepSeek V4 Flash deployment was declared tools: false and jsonOutput: false based on the model card listing Chat + Reasoning only. The live deployment returns well-formed tool_calls with finish_reason "tool_calls", and honours both response_format json_object and json_schema — matching its together-ai V4 Pro sibling, which already declares all three. Under-declaring routed tool and JSON traffic away from the provider and hid those paths from e2e entirely. Also restore the WHY comments on both trimmed reasoningEfforts lists, which were copied from their siblings without the explanations. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01GNFpeYxUh39cCdbp7Ttjud --- packages/models/src/models/deepseek.ts | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/packages/models/src/models/deepseek.ts b/packages/models/src/models/deepseek.ts index 5d936e4ae1..7a6245d552 100644 --- a/packages/models/src/models/deepseek.ts +++ b/packages/models/src/models/deepseek.ts @@ -433,6 +433,8 @@ export const deepseekModels = [ maxOutput: 393216, streaming: true, reasoning: true, + // Fireworks rejects "minimal" for this model; the rest of the enum + // (plus "none" for non-thinking mode) is accepted. reasoningEfforts: ["none", "low", "medium", "high", "xhigh", "max"], vision: false, tools: true, @@ -622,12 +624,18 @@ export const deepseekModels = [ maxOutput: 163840, streaming: true, reasoning: true, + // Together's deployment accepts any reasoning_effort string without + // validating it, and only the top tiers measurably change behaviour: + // xhigh and max roughly double the reasoning tokens, while + // low/medium/high land on the provider default. `none` is honoured + // through the `thinking` switch, not through reasoning_effort. reasoningEfforts: ["none", "xhigh", "max"], requiresDisableThinkingParam: true, reasoningOutput: "omit", vision: false, - tools: false, - jsonOutput: false, + tools: true, + jsonOutput: true, + jsonOutputSchema: true, }, { // RanoAI serves DeepSeek V4 Flash on Furiosa RNGD NPUs. The NPU