diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 0776f0fdfe..5fb7395c00 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -4248,6 +4248,7 @@ class LlamaCppBackend:
         frequency_penalty: Optional[float] = None,
         seed: Optional[int] = None,
         parallel_tool_calls: Optional[bool] = None,
+        typical_p: Optional[float] = None,
     ) -> Generator[str | dict, None, None]:
         """
         Send a chat completion request to llama-server and stream tokens back.
@@ -4306,6 +4307,11 @@ class LlamaCppBackend:
             payload["seed"] = seed
         if parallel_tool_calls is not None:
             payload["parallel_tool_calls"] = parallel_tool_calls
+        # Locally typical sampling. llama-server default 1.0 disables it;
+        # the field is llama.cpp-specific (no cloud provider accepts it),
+        # so we only forward it when the caller explicitly sets one.
+        if typical_p is not None:
+            payload["typical_p"] = typical_p
         payload["stream_options"] = {"include_usage": True}
 
         url = f"{self.base_url}/v1/chat/completions"
@@ -4451,6 +4457,7 @@ class LlamaCppBackend:
         frequency_penalty: Optional[float] = None,
         seed: Optional[int] = None,
         parallel_tool_calls: Optional[bool] = None,
+        typical_p: Optional[float] = None,
     ) -> Generator[dict, None, None]:
         """
         Agentic loop: let the model call tools, execute them, and continue.
@@ -4548,6 +4555,8 @@ class LlamaCppBackend:
                 payload["seed"] = seed
             if parallel_tool_calls is not None:
                 payload["parallel_tool_calls"] = parallel_tool_calls
+            if typical_p is not None:
+                payload["typical_p"] = typical_p
 
             try:
                 _auth_headers = (
@@ -5240,6 +5249,8 @@ class LlamaCppBackend:
             stream_payload["seed"] = seed
         if parallel_tool_calls is not None:
             stream_payload["parallel_tool_calls"] = parallel_tool_calls
+        if typical_p is not None:
+            stream_payload["typical_p"] = typical_p
         stream_payload["stream_options"] = {"include_usage": True}
 
         cumulative = ""
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 040bb6ce6c..b2c12b82c4 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -828,6 +828,18 @@ class ChatCompletionRequest(BaseModel):
             "default (which is `true` everywhere today)."
         ),
     )
+    typical_p: Optional[float] = Field(
+        None,
+        ge = 0.0,
+        le = 1.0,
+        description = (
+            "Locally typical sampling (llama.cpp `typ_p`). 1.0 disables. "
+            "Local llama-server only — no SaaS provider currently accepts "
+            "this field, so the frontend capability map gates it off for "
+            "every external provider and the local path forwards it on "
+            "/v1/chat/completions."
+        ),
+    )
     fast_mode: Optional[bool] = Field(
         None,
         description = (
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 72ab4afed8..5d71b446b5 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -2517,6 +2517,7 @@ async def openai_chat_completions(
                     frequency_penalty = payload.frequency_penalty,
                     seed = payload.seed,
                     parallel_tool_calls = payload.parallel_tool_calls,
+                    typical_p = payload.typical_p,
                 )
 
             _tool_sentinel = object()
@@ -2690,6 +2691,7 @@ async def openai_chat_completions(
                 frequency_penalty = payload.frequency_penalty,
                 seed = payload.seed,
                 parallel_tool_calls = payload.parallel_tool_calls,
+                typical_p = payload.typical_p,
             )
 
         _gguf_sentinel = object()
@@ -4963,6 +4965,7 @@ def _build_passthrough_payload(
     frequency_penalty = None,
     seed = None,
     parallel_tool_calls = None,
+    typical_p = None,
     tool_choice = "auto",
     response_format = None,
     chat_template_kwargs = None,
@@ -5013,6 +5016,11 @@ def _build_passthrough_payload(
         body["seed"] = seed
     if parallel_tool_calls is not None:
         body["parallel_tool_calls"] = parallel_tool_calls
+    # llama.cpp-specific locally-typical sampling (typ_p in the sampler
+    # chain). No SaaS provider accepts this; the frontend capability map
+    # gates it to local only.
+    if typical_p is not None:
+        body["typical_p"] = typical_p
     if response_format is not None:
         # llama-server applies a GBNF grammar derived from the JSON schema
         # when response_format is present. Field is documented flat at the
@@ -5447,6 +5455,7 @@ def _build_openai_passthrough_body(payload, backend_ctx = None) -> dict:
         frequency_penalty = payload.frequency_penalty,
         seed = payload.seed,
         parallel_tool_calls = payload.parallel_tool_calls,
+        typical_p = payload.typical_p,
         tool_choice = tool_choice,
         response_format = _extract_response_format(payload),
         chat_template_kwargs = tpl_kwargs,
diff --git a/studio/backend/tests/test_sampling_params_routing.py b/studio/backend/tests/test_sampling_params_routing.py
index 1727d01028..50714fcac4 100644
--- a/studio/backend/tests/test_sampling_params_routing.py
+++ b/studio/backend/tests/test_sampling_params_routing.py
@@ -1001,6 +1001,40 @@ def test_local_anthropic_passthrough_helpers_accept_parallel_tool_calls():
     assert body.get("parallel_tool_calls") is False, body
 
 
+def test_local_passthrough_forwards_typical_p_when_set():
+    """`typical_p` is a llama.cpp-specific sampler (`typ_p` in the
+    sampler chain). The local-llama-cpp passthrough payload builder must
+    forward it when set so the chat-adapter can opt in for local
+    backends without the field bleeding into external providers (whose
+    capability map gates it off)."""
+    from routes import inference as route_mod
+
+    body = route_mod._build_passthrough_payload(
+        openai_messages = [{"role": "user", "content": "hi"}],
+        openai_tools = None,
+        temperature = 0.6,
+        top_p = 0.95,
+        top_k = 20,
+        max_tokens = 64,
+        stream = True,
+        typical_p = 0.7,
+    )
+    assert body.get("typical_p") == 0.7, body
+
+    # When unset, the field is omitted entirely so llama-server falls
+    # back to its 1.0 default.
+    body2 = route_mod._build_passthrough_payload(
+        openai_messages = [{"role": "user", "content": "hi"}],
+        openai_tools = None,
+        temperature = 0.6,
+        top_p = 0.95,
+        top_k = 20,
+        max_tokens = 64,
+        stream = True,
+    )
+    assert "typical_p" not in body2, body2
+
+
 def test_anthropic_4_7_sampling_removed_regex_matches_expected_ids():
     """Pin the canonical Claude 4.7 model-id shape so the frontend
     ANTHROPIC_4_7_SAMPLING_REMOVED_REGEX in
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index c6550d96e8..52e2b87d2d 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -1587,6 +1587,15 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
               params.parallelToolCalls === false
                 ? { parallel_tool_calls: false }
                 : {}),
+              // llama.cpp `typ_p`. External providers all have
+              // capabilities.typicalP=false; only the permissive local
+              // buckets (custom/vllm/ollama/llama_cpp) opt in. `null`
+              // (unset) or `1.0` (llama-server default) is a no-op.
+              ...(externalCapabilities?.typicalP &&
+              params.typicalP !== null &&
+              params.typicalP !== 1
+                ? { typical_p: params.typicalP }
+                : {}),
               // Built-in tools: Search pill maps to provider-side
               // web_search (currently OpenAI / Anthropic / OpenRouter /
               // Kimi); Code pill maps to Anthropic's server-side
@@ -1707,6 +1716,12 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
               : {}),
             ...(params.seed !== null ? { seed: params.seed } : {}),
             ...(params.stop.length > 0 ? { stop: params.stop } : {}),
+            // llama.cpp `typ_p`. Local only — external providers gate
+            // it off via capability map. `null` (unset) or 1.0 (server
+            // default) is a no-op so we forward only meaningful values.
+            ...(params.typicalP !== null && params.typicalP !== 1
+              ? { typical_p: params.typicalP }
+              : {}),
             parallel_tool_calls: params.parallelToolCalls,
             image_base64: imageBase64,
             audio_base64: audioBase64,
diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts
index e97d419b45..5a48aa8fc6 100644
--- a/studio/frontend/src/features/chat/provider-capabilities.ts
+++ b/studio/frontend/src/features/chat/provider-capabilities.ts
@@ -13,6 +13,10 @@
  * a null capability — every knob renders for them.
  */
 
+// NB: when adding a new sampling knob, default it to `false` on every
+// SaaS provider in PROVIDER_CAPABILITIES below (only local backends
+// + the permissive {custom, vllm, ollama, llama_cpp, openrouter}
+// providers should expose llama.cpp-specific samplers).
 export interface ProviderCapabilities {
   /**
    * Temperature sampling. Reasoning-class models (OpenAI's gpt-5.x / o3 via
@@ -58,6 +62,13 @@ export interface ProviderCapabilities {
    * `disable_parallel_tool_use: true` on Anthropic (inverted).
    */
   parallelToolCalls: boolean;
+  /**
+   * llama.cpp `typ_p` (locally typical sampling). Local llama-server
+   * only — no SaaS provider currently accepts this field. Default is
+   * `false` for every external provider and `true` only for the local
+   * permissive {custom, vllm, ollama, llama_cpp} buckets.
+   */
+  typicalP: boolean;
 }
 
 /**
@@ -401,6 +412,7 @@ const OPENAI_COMPAT_BASE: ProviderCapabilities = {
   stop: true,
   serviceTier: false,
   parallelToolCalls: true,
+  typicalP: false,
 };
 
 const ALL_SUPPORTED: ProviderCapabilities = {
@@ -415,6 +427,7 @@ const ALL_SUPPORTED: ProviderCapabilities = {
   stop: true,
   serviceTier: false,
   parallelToolCalls: true,
+  typicalP: true,
 };
 
 // Reasoning-class OpenAI models served via /v1/responses fix temperature
@@ -437,6 +450,7 @@ const OPENAI_REASONING_CAPABILITIES: ProviderCapabilities = {
   stop: false,
   serviceTier: true,
   parallelToolCalls: true,
+  typicalP: false,
 };
 const OPENAI_CHAT_CAPABILITIES: ProviderCapabilities = {
   temperature: true,
@@ -452,6 +466,7 @@ const OPENAI_CHAT_CAPABILITIES: ProviderCapabilities = {
   stop: false,
   serviceTier: true,
   parallelToolCalls: true,
+  typicalP: false,
 };
 
 // Prefix list for OpenAI reasoning-class model ids. Kept in sync with
@@ -493,6 +508,23 @@ function isClaude47SamplingRemoved(modelId: string | null | undefined): boolean
   return ANTHROPIC_4_7_SAMPLING_REMOVED_REGEX.test(normalized);
 }
 
+// DeepSeek reasoning-class models silently ignore temperature, top_p,
+// presence_penalty, frequency_penalty and 400 on logprobs/top_logprobs.
+// `deepseek-reasoner` is the dedicated thinking model;
+// `deepseek-v4-flash` runs reasoning-mode under the same flag as well.
+// Match by prefix so future revisions (deepseek-reasoner-2027 etc.)
+// continue to gate correctly.
+const DEEPSEEK_REASONING_MODEL_PREFIXES = [
+  "deepseek-reasoner",
+  "deepseek-r1",
+] as const;
+
+function isDeepSeekReasoningModelId(modelId: string | null | undefined): boolean {
+  const normalized = modelId?.trim().toLowerCase() ?? "";
+  if (!normalized) return false;
+  return DEEPSEEK_REASONING_MODEL_PREFIXES.some((p) => normalized.startsWith(p));
+}
+
 const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
   // Default OpenAI bucket is reasoning-class (current registry only ships
   // gpt-5.x / o3 ids), but per-model resolution in getProviderCapabilities
@@ -520,6 +552,7 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
     stop: true,
     serviceTier: true,
     parallelToolCalls: true,
+    typicalP: false,
   },
   mistral: OPENAI_COMPAT_BASE,
   gemini: OPENAI_COMPAT_BASE,
@@ -547,8 +580,17 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
     stop: true,
     serviceTier: false,
     parallelToolCalls: false,
+    typicalP: false,
   },
   // DeepSeek deprecated presence/frequency penalty in their current docs.
+  // Chat-class defaults (deepseek-chat / deepseek-v4-flash non-thinking):
+  // accept temperature, top_p, seed, stop. Reasoning class
+  // (deepseek-reasoner / deepseek-v4-flash thinking-mode) ignores
+  // temperature, top_p, presence_penalty, frequency_penalty entirely and
+  // 400s on logprobs — see
+  // https://api-docs.deepseek.com/guides/reasoning_model. Per-model
+  // resolution in getProviderCapabilities downshifts reasoner ids onto
+  // DEEPSEEK_REASONING_CAPABILITIES.
   deepseek: {
     temperature: true,
     topP: true,
@@ -561,6 +603,7 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
     stop: true,
     serviceTier: false,
     parallelToolCalls: true,
+    typicalP: false,
   },
   qwen: OPENAI_COMPAT_BASE,
   huggingface: OPENAI_COMPAT_BASE,
@@ -591,6 +634,8 @@ const DEFAULT_EXTERNAL_CAPABILITIES = OPENAI_COMPAT_BASE;
  *     (OPENAI_REASONING_CAPABILITIES).
  *   - anthropic + claude-*-4-7: temperature/top_p/top_k stripped to
  *     match the backend 400-avoidance regex.
+ *   - deepseek + reasoning model (deepseek-reasoner / r1): hides
+ *     temperature/top_p (silently ignored upstream).
  */
 export function getProviderCapabilities(
   providerType: string | null | undefined,
@@ -604,6 +649,9 @@ export function getProviderCapabilities(
   if (providerType === "anthropic" && isClaude47SamplingRemoved(modelId)) {
     return { ...base, temperature: false, topP: false, topK: false };
   }
+  if (providerType === "deepseek" && isDeepSeekReasoningModelId(modelId)) {
+    return { ...base, temperature: false, topP: false };
+  }
   return base;
 }
 
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index 3cf4d83a48..0db72ac5c2 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -304,6 +304,14 @@ export interface OpenAIChatCompletionsRequest {
    * keeps each provider's upstream default.
    */
   parallel_tool_calls?: boolean;
+  /**
+   * llama.cpp `typ_p` (locally typical sampling). Local llama-server
+   * only — no SaaS provider currently accepts this. 1.0 disables
+   * (llama-server default). External-provider capability map already
+   * gates this off, so on the wire it only appears for local + the
+   * permissive {custom, vllm, ollama, llama_cpp} buckets.
+   */
+  typical_p?: number;
   /**
    * Anthropic fast-mode toggle. Opus 4.6 / 4.7 only; backend drops
    * silently on every other model + provider. See
diff --git a/studio/frontend/src/features/chat/types/runtime.ts b/studio/frontend/src/features/chat/types/runtime.ts
index 7fac74979b..0793573583 100644
--- a/studio/frontend/src/features/chat/types/runtime.ts
+++ b/studio/frontend/src/features/chat/types/runtime.ts
@@ -44,6 +44,12 @@ export interface InferenceParams {
    * upstream defaults across all three.
    */
   parallelToolCalls: boolean;
+  /**
+   * Locally typical sampling (llama.cpp `typ_p`). Local llama-server
+   * only — no SaaS provider currently accepts this. 1.0 disables (and
+   * is the llama-server default). `null` = unset (not forwarded).
+   */
+  typicalP: number | null;
   maxSeqLength: number;
   maxTokens: number;
   systemPrompt: string;
@@ -70,6 +76,7 @@ export const DEFAULT_INFERENCE_PARAMS: InferenceParams = {
   stop: [],
   serviceTier: null,
   parallelToolCalls: true,
+  typicalP: null,
   maxSeqLength: 4096,
   maxTokens: 8192,
   systemPrompt: "",
diff --git a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
index 1b186db727..ede79c7098 100644
--- a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
+++ b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
@@ -182,6 +182,16 @@ function sanitizeInferenceParams(
   if (typeof value.parallelToolCalls === "boolean") {
     params.parallelToolCalls = value.parallelToolCalls;
   }
+  // typicalP: nullable float (null = no typ_p on the wire, matching
+  // llama-server's default 1.0). Mirror seed's nullable-float handling.
+  if (value.typicalP === null) {
+    params.typicalP = null;
+  } else if (
+    typeof value.typicalP === "number" &&
+    Number.isFinite(value.typicalP)
+  ) {
+    params.typicalP = value.typicalP;
+  }
   // Mirror trustRemoteCode handling so the toggle survives reload
   // and the /api/chat/settings round-trip.
   if (typeof value.fastMode === "boolean") {