diff --git a/libdd-trace-utils/benches/deserialization_v05.rs b/libdd-trace-utils/benches/deserialization_v05.rs
new file mode 100644
index 0000000000..03fe37bc21
--- /dev/null
+++ b/libdd-trace-utils/benches/deserialization_v05.rs
@@ -0,0 +1,235 @@
+// Copyright 2024-Present Datadog, Inc. https://www.datadoghq.com/
+// SPDX-License-Identifier: Apache-2.0
+
+//! Microbenchmarks for the V05 msgpack decode path
+//! (`libdd_trace_utils::msgpack_decoder::v05`).
+//!
+//! Unlike V04, the V05 wire format encodes a shared string dictionary up front and every span
+//! references its string fields (service, name, resource, type, and every meta/metrics key/value)
+//! by index into that dictionary. Decoding therefore performs a dictionary lookup-and-clone for
+//! each string slot, which is real per-span work on the agent ingestion hot path and is otherwise
+//! uncovered by benchmarks.
+//!
+//! We vary two axes that drive that cost:
+//!   * the number of spans (work scales with span count), and
+//!   * the string-sharing ratio: a "high sharing" payload reuses a small dictionary across all
+//!     spans (the common case for tracer traffic, where service/name/tag keys repeat), while a "low
+//!     sharing" payload gives each span its own unique strings (a larger dictionary and worse cache
+//!     behavior).
+
+use std::alloc::System;
+use std::collections::HashMap;
+
+use criterion::{black_box, criterion_group, BenchmarkId, Criterion, Throughput};
+use libdd_common::bench_utils::{memory_allocated_measurement, AllocatedBytesMeasurement};
+use libdd_trace_utils::msgpack_decoder;
+
+/// A V05 span is a fixed 12-element tuple. String fields (service, name, resource, type, and the
+/// meta/metrics keys/values) are `u32` indices into the shared dictionary.
+type V05Span = (
+    u32,               // service (dict index)
+    u32,               // name (dict index)
+    u32,               // resource (dict index)
+    u64,               // trace_id
+    u64,               // span_id
+    u64,               // parent_id
+    i64,               // start
+    i64,               // duration
+    i32,               // error
+    HashMap<u32, u32>, // meta (dict index -> dict index)
+    HashMap<u32, f64>, // metrics (dict index -> value)
+    u32,               // type (dict index)
+);
+
+type V05Payload = (Vec<String>, Vec<Vec<V05Span>>);
+
+/// Number of meta tags per span. Picked to resemble a typical instrumented span (service/runtime
+/// metadata, a couple of resource attributes, thread info, etc.).
+const META_TAGS_PER_SPAN: usize = 8;
+/// Number of metrics per span (sampling priority, a couple of measured values).
+const METRICS_PER_SPAN: usize = 3;
+
+/// Builds a representative V05 payload.
+///
+/// `unique_per_span` controls the string-sharing ratio:
+///   * `false` (high sharing): all spans draw their strings from a single small shared dictionary,
+///     mirroring real tracer traffic where service names and tag keys repeat heavily.
+///   * `true` (low sharing): each span contributes its own unique strings, producing a large
+///     dictionary with little reuse.
+///
+/// Data is fully deterministic.
+fn build_v05_payload(num_traces: usize, spans_per_trace: usize, unique_per_span: bool) -> Vec<u8> {
+    let mut dict = Vec::new();
+    let intern = |s: String, dict: &mut Vec<String>| -> u32 {
+        let idx = dict.len() as u32;
+        dict.push(s);
+        idx
+    };
+
+    // Shared dictionary entries reused by every span in the "high sharing" scenario.
+    let shared_service = intern("test-service".to_string(), &mut dict);
+    let shared_name = intern("test-service.handler".to_string(), &mut dict);
+    let shared_resource = intern("GET /api/v1/resource".to_string(), &mut dict);
+    let shared_type = intern("web".to_string(), &mut dict);
+
+    // Shared meta keys/values (typical tag keys that repeat across all spans).
+    let shared_meta: Vec<(u32, u32)> = [
+        ("env", "production"),
+        ("version", "1.2.3"),
+        ("runtime-id", "f0e1d2c3-b4a5-6789-0abc-def012345678"),
+        ("language", "rust"),
+        ("component", "http"),
+        ("span.kind", "server"),
+        ("http.method", "GET"),
+        ("http.status_code", "200"),
+    ]
+    .iter()
+    .take(META_TAGS_PER_SPAN)
+    .map(|(k, v)| {
+        (
+            intern((*k).to_string(), &mut dict),
+            intern((*v).to_string(), &mut dict),
+        )
+    })
+    .collect();
+
+    let shared_metric_keys: Vec<_> = ["_sampling_priority_v1", "_dd.measured", "_dd.top_level"]
+        .iter()
+        .take(METRICS_PER_SPAN)
+        .map(|k| intern((*k).to_string(), &mut dict))
+        .collect();
+
+    let mut traces = Vec::with_capacity(num_traces);
+
+    for trace_idx in 0..num_traces {
+        let mut spans = Vec::with_capacity(spans_per_trace);
+        let root_span_id = 100_000_000_000 + trace_idx as u64;
+
+        for span_idx in 0..spans_per_trace {
+            let span_id = root_span_id + span_idx as u64 + 1;
+            let parent_id = if span_idx == 0 { 0 } else { root_span_id };
+
+            let (service, name, resource, ty, meta, metric_keys) = if unique_per_span {
+                // Low sharing: every span interns its own unique strings.
+                let service = intern(format!("service-{trace_idx}-{span_idx}"), &mut dict);
+                let name = intern(format!("op-{trace_idx}-{span_idx}"), &mut dict);
+                let resource = intern(format!("GET /api/{trace_idx}/{span_idx}"), &mut dict);
+                let ty = intern(format!("type-{}", span_idx % 4), &mut dict);
+
+                let meta: Vec<(u32, u32)> = (0..META_TAGS_PER_SPAN)
+                    .map(|m| {
+                        (
+                            intern(format!("tag.key.{trace_idx}.{span_idx}.{m}"), &mut dict),
+                            intern(format!("tag-value-{trace_idx}-{span_idx}-{m}"), &mut dict),
+                        )
+                    })
+                    .collect();
+                let metric_keys: Vec<u32> = (0..METRICS_PER_SPAN)
+                    .map(|m| intern(format!("metric.{trace_idx}.{span_idx}.{m}"), &mut dict))
+                    .collect();
+
+                (service, name, resource, ty, meta, metric_keys)
+            } else {
+                // High sharing: reuse the shared dictionary entries.
+                (
+                    shared_service,
+                    shared_name,
+                    shared_resource,
+                    shared_type,
+                    shared_meta.clone(),
+                    shared_metric_keys.clone(),
+                )
+            };
+
+            let meta_map: HashMap<u32, u32> = meta.into_iter().collect();
+            let metrics_map: HashMap<u32, f64> = metric_keys
+                .into_iter()
+                .enumerate()
+                .map(|(i, k)| (k, i as f64 + 1.0))
+                .collect();
+
+            spans.push((
+                service,
+                name,
+                resource,
+                100_000_000_000u64 + trace_idx as u64,
+                span_id,
+                parent_id,
+                1_700_000_000_000_000_000i64,
+                123_456i64,
+                0i32,
+                meta_map,
+                metrics_map,
+                ty,
+            ));
+        }
+        traces.push(spans);
+    }
+
+    let payload: V05Payload = (dict, traces);
+    rmp_serde::to_vec(&payload).expect("Failed to serialize V05 test payload.")
+}
+
+/// Runs the V05 decode benchmark matrix against the given criterion harness, which lets us reuse
+/// the same scenarios for both the wall-time and the bytes-allocated measurements.
+///
+/// Note: `Throughput::Bytes` is reported per scenario so each series can be read as a decode rate,
+/// but high-sharing and low-sharing payloads have very different encoded sizes for the same span
+/// count, so their throughput numbers are *not* directly comparable across the sharing dimension.
+fn bench_v05_matrix<M: criterion::measurement::Measurement>(
+    c: &mut Criterion<M>,
+    group_name: &str,
+) {
+    // A representative chunked payload: 20 traces (the upper-bound trace count for a tracer flush)
+    // with a varying number of spans each. Span counts stay realistic for a single flush while
+    // exercising the dictionary-dedup path across many spans.
+    const NUM_TRACES: usize = 20;
+    let span_counts = [10, 100, 500];
+
+    let mut group = c.benchmark_group(group_name);
+
+    for &spans_per_trace in &span_counts {
+        for (sharing_label, unique_per_span) in [("high_sharing", false), ("low_sharing", true)] {
+            let data = build_v05_payload(NUM_TRACES, spans_per_trace, unique_per_span);
+            let data_as_bytes = libdd_tinybytes::Bytes::copy_from_slice(&data);
+
+            group.throughput(Throughput::Bytes(data.len() as u64));
+            group.bench_with_input(
+                BenchmarkId::new(sharing_label, spans_per_trace * NUM_TRACES),
+                &data_as_bytes,
+                |b, data_as_bytes| {
+                    b.iter_batched(
+                        || data_as_bytes.clone(),
+                        |data_as_bytes| {
+                            let result = black_box(msgpack_decoder::v05::from_bytes(data_as_bytes));
+                            assert!(result.is_ok());
+                            // Return the result to avoid measuring the deallocation time.
+                            result
+                        },
+                        criterion::BatchSize::LargeInput,
+                    );
+                },
+            );
+        }
+    }
+
+    group.finish();
+}
+
+fn deserialize_msgpack_v05(c: &mut Criterion) {
+    bench_v05_matrix(c, "msgpack_decoder::v05");
+}
+
+/// Allocation-measured counterpart. The dictionary-dedup path clones a `Bytes` slice for every
+/// string slot, so the amount allocated per decode is the metric most directly affected by the
+/// sharing ratio.
+fn deserialize_msgpack_v05_allocs(c: &mut Criterion<AllocatedBytesMeasurement<System>>) {
+    bench_v05_matrix(c, "msgpack_decoder::v05 (allocs)");
+}
+
+criterion_group!(deserialize_v05_benches, deserialize_msgpack_v05);
+criterion_group!(
+    name = deserialize_v05_alloc_benches;
+    config = memory_allocated_measurement(&super::GLOBAL);
+    targets = deserialize_msgpack_v05_allocs
+);
diff --git a/libdd-trace-utils/benches/main.rs b/libdd-trace-utils/benches/main.rs
index 0d86f25ee5..20a32599e6 100644
--- a/libdd-trace-utils/benches/main.rs
+++ b/libdd-trace-utils/benches/main.rs
@@ -10,10 +10,13 @@ use libdd_common::bench_utils::ReportingAllocator;
 pub static GLOBAL: ReportingAllocator<System> = ReportingAllocator::new(System);
 
 mod deserialization;
+mod deserialization_v05;
 mod serialization;
 
 criterion_main!(
     serialization::serialize_benches,
     deserialization::deserialize_benches,
-    deserialization::deserialize_alloc_benches
+    deserialization::deserialize_alloc_benches,
+    deserialization_v05::deserialize_v05_benches,
+    deserialization_v05::deserialize_v05_alloc_benches
 );