llmtr: drop retired ids and correct Turkey-hosted model data (#4813)

* llmtr: correct gemma-4 context, pricing, modalities and tool calling

* llmtr: pin qwen3-6-35b tool_call to the measured value

* llmtr: correct magibu-11b-v8 pricing

* llmtr: mark medgemma-4b deprecated and correct its output cap

* llmtr: drop sincap, retired upstream on 2026-08-04

* llmtr: replace trendyol-7b with the model it now aliases

* llmtr: add trendyol-asure-12b

* llmtr: add muse-glimmer-30b-tr

* llmtr: tidy muse-glimmer-30b-tr source comment

* llmtr: point muse-glimmer-30b-tr at the Meta lab entry

* trendyol: add Asure 12B lab entry

* llmtr: point trendyol-asure-12b at the new lab entry

models/trendyol/asure-12b.toml

@@ -0,0 +1,30 @@
+# Sources:
+# https://huggingface.co/Trendyol/Trendyol-LLM-Asure-12B
+# https://huggingface.co/api/models/Trendyol/Trendyol-LLM-Asure-12B (createdAt, license, base_model)
+# https://huggingface.co/Trendyol/Trendyol-LLM-Asure-12B/raw/main/config.json (max_position_embeddings)
+# `reasoning` and `tool_call` are not stated on the model card; both were
+# measured against a host serving these weights (llmtr.com, 2026-08-16):
+# a request carrying `tools` returns no tool_calls, and no reasoning output
+# is produced.
+name = "Trendyol Asure 12B"
+description = "Turkish-language multimodal instruct model built on Gemma 3 12B for e-commerce text, chat, and image-text tasks"
+family = "gemma"
+release_date = "2026-02-19"
+last_updated = "2026-02-20"
+attachment = true
+reasoning = false
+temperature = true
+tool_call = false
+open_weights = true
+license = "Gemma"
+
+[limit]
+context = 131_072
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Trendyol/Trendyol-LLM-Asure-12B"

providers/llmtr/models/gemma-4.toml

@@ -1,21 +1,17 @@
+# Turkey-hosted deployment of Gemma 4 31B IT. Context window, output cap and
+# pricing come from LLMTR's public catalog API (https://llmtr.com/api/models,
+# accessed 2026-08-16). Tool calling and the reasoning toggle were verified
+# against POST https://llmtr.com/v1/chat/completions on the same date.
+# Toggle: reasoning = true|false
 name = "Gemma 4"
-description = "Open Gemma instruction model for efficient chat and self-hosted deployments"
-release_date = "2026-04-22"
-last_updated = "2026-04-22"
-attachment = false
-reasoning = false
-temperature = true
-tool_call = false
-open_weights = true
+base_model = "google/gemma-4-31b-it"
+reasoning_options = [{ type = "toggle" }]
 
 [cost]
-input = 5.0
-output = 10.0
+input = 2.0
+output = 5.0
+cache_read = 0.5
 
 [limit]
-context = 32_768
-output = 8_192
-
-[modalities]
-input = ["text"]
-output = ["text"]
+context = 131_072
+output = 131_072

providers/llmtr/models/magibu-11b-v8.toml

@@ -1,7 +1,11 @@
+# Pricing and limits from https://llmtr.com/api/models (accessed 2026-08-16).
+# The previous entry listed this route as free; LLMTR charges for it.
+# Tool calling verified absent on 2026-08-16: a chat/completions request with
+# `tools` is rejected with HTTP 400.
 name = "Magibu 11B v8"
-description = "General-purpose chat model for instruction following, writing, and analysis"
+description = "Turkish-language chat model for instruction following and assistant flows"
 release_date = "2026-06-05"
-last_updated = "2026-06-05"
+last_updated = "2026-08-16"
 attachment = false
 reasoning = false
 temperature = true
@@ -9,12 +13,12 @@ tool_call = false
 open_weights = false
 
 [cost]
-input = 0.0
-output = 0.0
+input = 0.1
+output = 0.5
 
 [limit]
 context = 8_192
-output = 4_096
+output = 8_192
 
 [modalities]
 input = ["text"]

providers/llmtr/models/medgemma-4b.toml

@@ -1,12 +1,15 @@
+# Limits and pricing from https://llmtr.com/api/models (accessed 2026-08-16),
+# where LLMTR marks this route as being retired.
 name = "MedGemma 4B"
-description = "Multimodal model for analyzing text, images, documents, and rich media"
+description = "Multimodal medical-domain Gemma variant for text and image analysis"
 release_date = "2026-04-26"
-last_updated = "2026-04-26"
+last_updated = "2026-08-16"
 attachment = true
 reasoning = false
 temperature = true
 tool_call = false
 open_weights = true
+status = "deprecated"
 
 [cost]
 input = 3.0
@@ -14,7 +17,7 @@ output = 5.0
 
 [limit]
 context = 8_192
-output = 4_096
+output = 8_192
 
 [modalities]
 input = ["text", "image"]

providers/llmtr/models/muse-glimmer-30b-tr.toml

@@ -0,0 +1,16 @@
+# LLMTR's Turkey-hosted deployment of Meta's Muse Glimmer 30B.
+# Pricing from https://llmtr.com/api/models (accessed 2026-08-16).
+# No reasoning control: verified 2026-08-16 that reasoning output is returned
+# unconditionally, `reasoning = true|false` changes nothing, and every
+# `reasoning_effort` value (none/minimal/low/medium/high/xhigh/max) is rejected
+# with `The "reasoning_effort" parameter is not supported by
+# "llmtr/muse-glimmer-30b-tr". This model does not expose a thinking mode.`
+# Hence always-on with no caller control, not a toggle.
+name = "Muse Glimmer 30B (TR)"
+base_model = "meta/muse-glimmer-30b"
+reasoning_options = []
+
+[cost]
+input = 2.0
+output = 5.0
+cache_read = 0.5

providers/llmtr/models/qwen3-6-35b.toml

@@ -1,7 +1,11 @@
-# LLMTR's endpoint-specific reasoning pages document Responses effort and GLM
-# toggles, but no raw Qwen3.6-35B toggle field or values; this claim is unproven.
-# https://llmtr.com/docs/gateway/reasoning-effort/ (accessed 2026-06-25)
+# LLMTR serves this route from its own Turkey-hosted deployment. Tool calling
+# does not carry over from the base entry: verified on 2026-08-16 that a
+# chat/completions request with `tools` returns HTTP 200 and no tool_calls,
+# so it is pinned false here. Limits and pricing from
+# https://llmtr.com/api/models (accessed 2026-08-16).
+# Toggle: reasoning = true|false
 base_model = "alibaba/qwen3.6-35b-a3b"
+tool_call = false
 reasoning_options = [{ type = "toggle" }]
 
 [cost]
@@ -10,3 +14,4 @@ output = 10.0
 
 [limit]
 context = 16_384
+output = 16_384

providers/llmtr/models/sincap.toml

@@ -1,21 +0,0 @@
-name = "Sincap"
-description = "General-purpose chat model for instruction following, writing, and analysis"
-release_date = "2026-05-05"
-last_updated = "2026-05-05"
-attachment = false
-reasoning = false
-temperature = true
-tool_call = false
-open_weights = false
-
-[cost]
-input = 0.0
-output = 0.0
-
-[limit]
-context = 128_000
-output = 8_192
-
-[modalities]
-input = ["text"]
-output = ["text"]

providers/llmtr/models/trendyol-7b.toml

@@ -1,22 +0,0 @@
-name = "Trendyol 7B"
-description = "Qwen instruction model for multilingual chat, reasoning, and tool use"
-family = "qwen"
-release_date = "2026-06-06"
-last_updated = "2026-06-06"
-attachment = false
-reasoning = false
-temperature = true
-tool_call = false
-open_weights = true
-
-[cost]
-input = 0.0
-output = 0.0
-
-[limit]
-context = 32_768
-output = 8_192
-
-[modalities]
-input = ["text"]
-output = ["text"]

providers/llmtr/models/trendyol-asure-12b.toml

@@ -0,0 +1,15 @@
+# Turkey-hosted deployment. Replaces the earlier `trendyol-7b` entry: that id
+# is now only a migration alias and resolves to this model.
+# Pricing and the served window from https://llmtr.com/api/models
+# (accessed 2026-08-16); this host serves a smaller window than the model's
+# native 131,072.
+base_model = "trendyol/asure-12b"
+
+[cost]
+input = 0.1
+output = 0.5
+cache_read = 0.025
+
+[limit]
+context = 40_960
+output = 40_960