diff --git a/models/alibaba/qwen3-14b.toml b/models/alibaba/qwen3-14b.toml
new file mode 100644
index 0000000000..eb8611c23d
--- /dev/null
+++ b/models/alibaba/qwen3-14b.toml
@@ -0,0 +1,27 @@
+name = "Qwen3 14B"
+description = "Dense open Qwen model with hybrid thinking for chat, reasoning, and coding"
+family = "qwen"
+release_date = "2025-04"
+last_updated = "2025-04"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-04"
+open_weights = true
+
+[limit]
+context = 131_072
+# 8_192 is what the first-party alibaba and alibaba-cn entries use, and what
+# OpenRouter publishes. Providers that serve a larger ceiling override it —
+# Infersia serves 16_384 and declares that in its own entry.
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3-14B"
diff --git a/models/alibaba/qwen3-8b.toml b/models/alibaba/qwen3-8b.toml
new file mode 100644
index 0000000000..85f9e14a4d
--- /dev/null
+++ b/models/alibaba/qwen3-8b.toml
@@ -0,0 +1,26 @@
+name = "Qwen3 8B"
+description = "Compact dense Qwen model with hybrid thinking for agents and high-volume tasks"
+family = "qwen"
+release_date = "2025-04"
+last_updated = "2025-04"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-04"
+open_weights = true
+
+[limit]
+context = 131_072
+# 8_192 is what the first-party alibaba and alibaba-cn entries use, and what
+# OpenRouter publishes. Providers that serve a larger ceiling override it.
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3-8B"
diff --git a/providers/infersia/logo.svg b/providers/infersia/logo.svg
new file mode 100644
index 0000000000..15a2c0ead0
--- /dev/null
+++ b/providers/infersia/logo.svg
@@ -0,0 +1,29 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor">
+  <!--
+    Three server racks of ascending height, built from stacked units, the
+    tallest carrying a tittle so the mark reads as a lowercase "i".
+
+    Unlike the favicon variant of this mark, the faded racks use opacity on
+    currentColor rather than pre-flattened indigo, per the models.dev logo
+    guidelines: the logo inherits the surrounding text colour and so works in
+    both light and dark themes. The vents are gaps rather than drawn lines, so
+    they take the page background instead of a hardcoded white.
+  -->
+  <g opacity="0.35">
+    <rect x="2.5" y="19.4" width="5" height="2.6" rx="1.2"/>
+    <rect x="2.5" y="15.8" width="5" height="2.6" rx="1.2"/>
+  </g>
+
+  <g opacity="0.6">
+    <rect x="9.5" y="19.4" width="5" height="2.6" rx="1.2"/>
+    <rect x="9.5" y="15.8" width="5" height="2.6" rx="1.2"/>
+    <rect x="9.5" y="12.2" width="5" height="2.6" rx="1.2"/>
+  </g>
+
+  <rect x="16.5" y="19.4" width="5" height="2.6" rx="1.2"/>
+  <rect x="16.5" y="15.8" width="5" height="2.6" rx="1.2"/>
+  <rect x="16.5" y="12.2" width="5" height="2.6" rx="1.2"/>
+  <rect x="16.5" y="8.6" width="5" height="2.6" rx="1.2"/>
+
+  <rect x="16.5" y="1.2" width="5" height="4" rx="2"/>
+</svg>
diff --git a/providers/infersia/models/deepseek/deepseek-v4-flash-0731.toml b/providers/infersia/models/deepseek/deepseek-v4-flash-0731.toml
new file mode 100644
index 0000000000..2322697223
--- /dev/null
+++ b/providers/infersia/models/deepseek/deepseek-v4-flash-0731.toml
@@ -0,0 +1,55 @@
+# Infersia serving details for DeepSeek V4 Flash 0731.
+#
+# Sources:
+#   Pricing   https://infersia.com/models/deepseek/deepseek-v4-flash-0731
+#   Limits    https://api.infersia.com/v1/models (context_length, max_completion_tokens)
+#
+# Served as DeepSeek's own MXFP4 release — not a quantisation we produced. The
+# KV cache is fp8 in DeepSeek's sparse-MLA layout, which is what makes the full
+# 1M window practical to serve rather than nominal.
+#
+# The directory is `deepseek/` because a provider file's path is that provider's
+# own API id and we serve this as `deepseek/deepseek-v4-flash-0731`.
+#
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default false)
+# Effort: top-level reasoning_effort = "max" (deepest; "low"/"high" are
+# accepted by the engine but do not change depth, so they are not listed)
+
+base_model = "deepseek/deepseek-v4-flash-0731"
+interleaved = true
+
+# Verified against the live endpoint, not transcribed from docs. Thinking
+# defaults OFF (280 completion tokens on a short prompt); the toggle turns it
+# on, and `reasoning_effort = "max"` injects the deepest deliberation prompt
+# (7,972 tokens on the same short prompt). Off is the toggle, not an effort
+# level, so `none` is not listed; DeepSeek's documented low/high are accepted
+# by the engine as no-ops on this stack, and a level that changes nothing is
+# not a level.
+#
+# Deliberation is returned in the message field `reasoning` — vLLM's name, not
+# DeepSeek's `reasoning_content`. That is why `interleaved` is the bare flag
+# above rather than a field declaration: the schema can only name
+# reasoning_content or reasoning_details, and claiming either would point
+# clients at an empty key. Returned on both the streaming and non-streaming
+# paths, billed as completion tokens.
+[[reasoning_options]]
+type = "toggle" # API: {"chat_template_kwargs": {"enable_thinking": true}}
+
+[[reasoning_options]]
+type = "effort" # API: top-level {"reasoning_effort": "max"}
+values = ["max"]
+
+# Both differ from the canonical entry, in opposite directions: 1_048_576 is
+# the model's full window and what we serve, against a rounded 1_000_000, while
+# completions are capped at 32_768 rather than 384_000.
+[limit]
+context = 1_048_576
+output = 32_768
+
+[provider]
+shape = "completions"
+
+[cost]
+input = 0.14
+output = 0.28
+cache_read = 0.035
diff --git a/providers/infersia/models/qwen/qwen3-14b.toml b/providers/infersia/models/qwen/qwen3-14b.toml
new file mode 100644
index 0000000000..1da7092c73
--- /dev/null
+++ b/providers/infersia/models/qwen/qwen3-14b.toml
@@ -0,0 +1,30 @@
+# Infersia serving details for Qwen3 14B.
+#
+# Sources:
+#   Pricing   https://infersia.com/models/qwen/qwen3-14b
+#   Limits    https://api.infersia.com/v1/models (context_length, max_completion_tokens)
+#
+# Served AWQ int4 with an fp8 KV cache. The 131_072 window is reached with YaRN
+# (factor 4.0 over the model's native 32_768) and is verified in service.
+#
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default false)
+
+base_model = "alibaba/qwen3-14b"
+
+# Qwen3 hybrid thinking. Infersia defaults it off, so an ordinary chat turn does
+# not spend 500-1500 unrequested reasoning tokens; send the kwarg to turn it on.
+[[reasoning_options]]
+type = "toggle" # API: {"chat_template_kwargs": {"enable_thinking": true}}
+
+# Only the output ceiling differs from the base entry; context and modalities
+# are inherited unchanged.
+[limit]
+output = 16_384
+
+[provider]
+shape = "completions"
+
+[cost]
+input = 0.09
+output = 0.22
+cache_read = 0.0225
diff --git a/providers/infersia/models/qwen/qwen3-8b.toml b/providers/infersia/models/qwen/qwen3-8b.toml
new file mode 100644
index 0000000000..257740fe4e
--- /dev/null
+++ b/providers/infersia/models/qwen/qwen3-8b.toml
@@ -0,0 +1,29 @@
+# Infersia serving details for Qwen3 8B.
+#
+# Sources:
+#   Pricing   https://infersia.com/models/qwen/qwen3-8b
+#   Limits    https://api.infersia.com/v1/models (context_length, max_completion_tokens)
+#
+# Served AWQ int4 with an fp8 KV cache, at the model's native 32_768 window
+# rather than the YaRN-extended 131_072 the base entry describes.
+#
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default false)
+
+base_model = "alibaba/qwen3-8b"
+
+# Qwen3 hybrid thinking. Infersia defaults it off, so an ordinary chat turn does
+# not spend 500-1500 unrequested reasoning tokens; send the kwarg to turn it on.
+[[reasoning_options]]
+type = "toggle" # API: {"chat_template_kwargs": {"enable_thinking": true}}
+
+# Only the context differs from the base entry — native window, no YaRN.
+[limit]
+context = 32_768
+
+[provider]
+shape = "completions"
+
+[cost]
+input = 0.05
+output = 0.15
+cache_read = 0.0125
diff --git a/providers/infersia/models/qwen/qwen3.6-35b-a3b.toml b/providers/infersia/models/qwen/qwen3.6-35b-a3b.toml
new file mode 100644
index 0000000000..21e537f5d9
--- /dev/null
+++ b/providers/infersia/models/qwen/qwen3.6-35b-a3b.toml
@@ -0,0 +1,36 @@
+# Infersia serving details for Qwen3.6 35B-A3B.
+#
+# Sources:
+#   Pricing   https://infersia.com/models/qwen/qwen3.6-35b-a3b
+#   Limits    https://api.infersia.com/v1/models (context_length, max_completion_tokens)
+#
+# Served AWQ int4 with an fp8 KV cache. The 262K context is the model's native
+# window, not a YaRN extension.
+#
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default false)
+
+base_model = "alibaba/qwen3.6-35b-a3b"
+
+# Qwen3 hybrid thinking. Infersia defaults it off, so an ordinary chat turn does
+# not spend 500-1500 unrequested reasoning tokens; send the kwarg to turn it on.
+[[reasoning_options]]
+type = "toggle" # API: {"chat_template_kwargs": {"enable_thinking": true}}
+
+# Overrides the base entry's 65_536 output ceiling — Infersia caps completions
+# at 32_768 tokens. Context is inherited unchanged.
+[limit]
+output = 32_768
+
+# Text and image only. The base entry also lists video and audio: the weights
+# support both, this endpoint accepts neither.
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[provider]
+shape = "completions"
+
+[cost]
+input = 0.10
+output = 0.90
+cache_read = 0.025
diff --git a/providers/infersia/provider.toml b/providers/infersia/provider.toml
new file mode 100644
index 0000000000..cc6c2472e9
--- /dev/null
+++ b/providers/infersia/provider.toml
@@ -0,0 +1,7 @@
+name = "Infersia"
+env = ["INFERSIA_API_KEY"]
+npm = "@ai-sdk/openai-compatible"
+api = "https://api.infersia.com/v1"
+# The catalogue rather than /docs: the schema asks for the page "where models
+# are listed", and /docs is the API reference — it never names a model.
+doc = "https://infersia.com/models"