From 00dca982d4b8c2a4145aaea01eae0e1a7103b08a Mon Sep 17 00:00:00 2001 From: Pessimist228 Date: Sun, 20 Sep 2026 13:10:31 +0200 Subject: [PATCH 1/2] llmtech: canonical model id is nvidia/Qwen3.8-27B-NVFP4 The served build is NVIDIA's own NVFP4 checkpoint, not the community one. The previous id stays accepted by the API as a synonym, so nothing breaks for anyone who has not switched yet. --- .../models/nvidia/Qwen3.8-27B-NVFP4.toml | 22 +++++++++++++++++++ 1 file changed, 22 insertions(+) create mode 100644 providers/llmtech/models/nvidia/Qwen3.8-27B-NVFP4.toml diff --git a/providers/llmtech/models/nvidia/Qwen3.8-27B-NVFP4.toml b/providers/llmtech/models/nvidia/Qwen3.8-27B-NVFP4.toml new file mode 100644 index 00000000000..ac3503d02a3 --- /dev/null +++ b/providers/llmtech/models/nvidia/Qwen3.8-27B-NVFP4.toml @@ -0,0 +1,22 @@ +# Reasoning wire (OpenAI-compatible chat completions at https://api.llmtech.eu/v1): +# Toggle: chat_template_kwargs.enable_thinking = true | false. +# Effort: chat_template_kwargs.reasoning_effort = "low" | "medium" | "xhigh". +# Non-streaming responses carry reasoning text in the `reasoning` field; +# streaming deltas use `reasoning_content`. + +base_model = "alibaba/qwen3.8-27b" + +[[reasoning_options]] +type = "toggle" + +[[reasoning_options]] +type = "effort" +values = ["low", "medium", "xhigh"] + +[cost] +input = 0.25 +output = 2.09 +cache_read = 0.04 + +[modalities] +input = ["text", "image"] From 5575ebb65f55aaebeb4a92c3fdd223d858d10bbe Mon Sep 17 00:00:00 2001 From: Pessimist228 Date: Sun, 20 Sep 2026 13:10:33 +0200 Subject: [PATCH 2/2] llmtech: drop the old unsloth/ model path --- .../models/unsloth/Qwen3.8-27B-NVFP4.toml | 22 ------------------- 1 file changed, 22 deletions(-) delete mode 100644 providers/llmtech/models/unsloth/Qwen3.8-27B-NVFP4.toml diff --git a/providers/llmtech/models/unsloth/Qwen3.8-27B-NVFP4.toml b/providers/llmtech/models/unsloth/Qwen3.8-27B-NVFP4.toml deleted file mode 100644 index ac3503d02a3..00000000000 --- a/providers/llmtech/models/unsloth/Qwen3.8-27B-NVFP4.toml +++ /dev/null @@ -1,22 +0,0 @@ -# Reasoning wire (OpenAI-compatible chat completions at https://api.llmtech.eu/v1): -# Toggle: chat_template_kwargs.enable_thinking = true | false. -# Effort: chat_template_kwargs.reasoning_effort = "low" | "medium" | "xhigh". -# Non-streaming responses carry reasoning text in the `reasoning` field; -# streaming deltas use `reasoning_content`. - -base_model = "alibaba/qwen3.8-27b" - -[[reasoning_options]] -type = "toggle" - -[[reasoning_options]] -type = "effort" -values = ["low", "medium", "xhigh"] - -[cost] -input = 0.25 -output = 2.09 -cache_read = 0.04 - -[modalities] -input = ["text", "image"]