dstackai · peterschmidt85 · Jul 31, 2026
diff --git a/mkdocs/docs/concepts/presets.md b/mkdocs/docs/concepts/presets.md
@@ -33,6 +33,11 @@ base: Qwen/Qwen2.5-7B-Instruct
 # The number of benchmarked trials
 max_trials: 3
 
+# The benchmark workload (defaults to 8, 1024, 1024)
+concurrency: 16
+input_tokens: 2048
+output_tokens: 512
+
 env:
   - HF_TOKEN
 ```
@@ -66,9 +71,11 @@ $ dstack preset create -f preset.dstack.yml --resume a1b2c3d4
 
 </div>
 
+When resuming, `model`, `prompt`, `min_context_length`, `max_trials`, `concurrency`, `baseline`, and `fleets` are read from the original session, not from the configuration file. Editing them and resuming has no effect. To change any of them, create a new preset.
+
 To stop a creation and its runs, use `dstack preset stop`.
 
-!!! info "Claude configuration"
+??? info "Claude configuration"
     By default, preset creation uses the existing `claude` login. To use an Anthropic API key instead, set:
 
     ```shell
@@ -84,6 +91,11 @@ To stop a creation and its runs, use `dstack preset stop`.
 
     Supported effort levels are `low`, `medium`, `high`, `xhigh`, and `max`.
 
+!!! info "Recommended setup"
+    - Profile settings such as `spot_policy`, `max_price`, and `backends` are ignored during preset creation. Configure them on the fleet instead, and pass a specific fleet so the agent uses the offers you intend.
+    - The agent chooses the hardware that gets the best performance. If you have a preference on hardware (e.g. price or specific GPUs), use a fleet that is pre-configured correspondingly.
+    - If `min_context_length` is unset, the agent may choose a short context to get a better benchmark result. Set it explicitly.
+
 ## Configuration options
 
 ### Model
@@ -106,15 +118,21 @@ To stop a creation and its runs, use `dstack preset stop`.
 
 ### Trials
 
-`max_trials` is required and sets how many benchmarked trials the agent runs before promoting the best one. Set `concurrency` to control the benchmark concurrency.
+`max_trials` is required and sets how many benchmarked trials the agent runs before promoting the best one.
+
+`concurrency`, `input_tokens`, and `output_tokens` define the benchmark workload, and default to 8, 1024, and 1024. The same values are used for every trial and for the final service benchmark, so results are comparable.
+
+### Baseline
+
+Set `baseline: true` to make the first trial a baseline: the agent serves the model the way the chosen serving framework recommends, without tuning it for performance. Later trials are optimization attempts.
 
 ### Context length
 
-Set `context_length` to require a minimum supported context length.
+Set `min_context_length` to require a minimum supported context length.
 
 ### Fleets
 
-Set `fleets` to restrict creation and reuse to specific [fleets](fleets.md). Placement properties such as `backends`, `max_price`, and `spot_policy` constrain both creation and reuse too.
+Set `fleets` to restrict creation and reuse to specific [fleets](fleets.md).
 
 ### Prompt
 
@@ -147,7 +165,7 @@ $ dstack preset apply -f preset.dstack.yml --id 532f3f4b
  Spot policy    on-demand
  Max price      off
  Model          Qwen/Qwen3.5-27B (base)
- Preset         532f3f4b (ctx=8K con=8 387 tok/s TTFT 582ms)
+ Preset         532f3f4b (ctx=8K io=512/512 con=8 tps=387 ttft=582)
 
  #  BACKEND           RESOURCES                                         INSTANCE TYPE                  PRICE
  1  runpod (EU-RO-1)  cpu=12 mem=54GB disk=100GB gpu=RTXPRO4500:32GB:1  NVIDIA RTX PRO 4500 Blackwell  $0.74
@@ -167,20 +185,30 @@ Use `dstack preset` to list presets:
 
 ```shell
 $ dstack preset list
- BASE               ID        GPU                  BENCHMARK                    STATUS              SUBMITTED     NAME
- Qwen/Qwen2.5-0.5B
-                    bc592b38                                                    clauding (0/3)      23 sec ago    qwen05
- Qwen/Qwen3-32B
-                    f91d6b60  RTX5090:32GB:1       con=8 576 tok/s TTFT 368ms   verified (10/10)    2 days ago    qwen3-32b
- Qwen/Qwen3.5-27B
-                    3c4d5e6f                                                    verifying (3/3)     2 min ago     qwen35-27b-2
-                    532f3f4b  RTXPRO4500:32GB:1..  con=8 387 tok/s TTFT 582ms   verified (4/4)      yesterday     qwen35-27b
-                    d1c2e12b  RTX5090:32GB:1       con=8 266 tok/s TTFT 2.15s   verified (7/10)     yesterday
+ ID        BASE              GPU                WORKLOAD                    BEST                       STATUS           SUBMITTED
+ bc592b38  Qwen/Qwen3.6-27B  RTXPRO4500:32GB:1  ctx=32K io=1024/1024 con=8  tps=422 ttft=501  ▁▄█      trialing (3/10)  2 min ago
 ```
 
 </div>
 
-Presets are grouped by base model. In-progress creations appear too, with a live status like `clauding` or `verifying`. Pass `-w` to watch in realtime. Pass `-v` to include validation resources and all benchmark metrics, or `--json` for complete preset objects. Filter with `--base` or `--repo`.
+By default, `dstack preset` shows creations that are still running, or the most recent one if none are. Pass `-a` to show every preset, or `-n` to show the last N:
+
+<div class="termy">
+
+```shell
+$ dstack preset list -a
+ ID        BASE              GPU                WORKLOAD                    BEST                       STATUS           SUBMITTED
+ bc592b38  Qwen/Qwen3.6-27B  RTXPRO4500:32GB:1  ctx=32K io=1024/1024 con=8  tps=422 ttft=501  ▁▄█      trialing (3/10)  2 min ago
+ f91d6b60  Qwen/Qwen3-32B    RTX5090:32GB:1     ctx=32K io=1024/512 con=8   tps=576 ttft=368  ▁▁▅▅▄▅▇█ verified (8/8)   2 days ago
+ 532f3f4b  Qwen/Qwen3.5-27B  RTXPRO4500:32GB:1  ctx=8K io=512/512 con=8     tps=387 ttft=582  ▁▇██     verified (4/4)   yesterday
+ d1c2e12b  Qwen/Qwen3.5-27B  RTX5090:32GB:1     ctx=32K io=1024/128 con=8   tps=266 ttft=2154 █▆█▁███  verified (7/10)  yesterday
+```
+
+</div>
+
+The `BEST` column is the best trial so far, followed by one glyph per trial: height is throughput, `·` is a trial that produced no benchmark. The shape shows whether a run converged or wandered.
+
+Pass `-w` to watch in realtime. Pass `-v` to include validation resources, the model repo, and the preset name, or `--json` for complete preset objects. Filter with `--base` or `--repo`.
 
 ### Delete presets
 

diff --git a/src/dstack/_internal/cli/commands/preset.py b/src/dstack/_internal/cli/commands/preset.py
@@ -235,12 +235,26 @@ def _list(self, args: argparse.Namespace) -> None:
         verbose = args.verbose
         if not getattr(args, "watch", False):
             presets, sessions = self._list_presets_and_sessions(base=base, repo=repo)
-            print_presets(presets, sessions=sessions, verbose=verbose)
+            print_presets(
+                presets,
+                sessions=sessions,
+                verbose=verbose,
+                all_presets=args.all_presets,
+                limit=args.limit,
+            )
             return
         with Live(console=console, refresh_per_second=LIVE_TABLE_REFRESH_RATE_PER_SEC) as live:
             while True:
                 presets, sessions = self._list_presets_and_sessions(base=base, repo=repo)
-                live.update(get_presets_table(presets, sessions=sessions, verbose=verbose))
+                live.update(
+                    get_presets_table(
+                        presets,
+                        sessions=sessions,
+                        verbose=verbose,
+                        all_presets=args.all_presets,
+                        limit=args.limit,
+                    )
+                )
                 time.sleep(LIVE_TABLE_PROVISION_INTERVAL_SECS)
 
     def _list_presets_and_sessions(
@@ -408,6 +422,21 @@ def _add_list_args(parser: argparse.ArgumentParser) -> None:
         action="store_true",
         help="Output in JSON format",
     )
+    parser.add_argument(
+        "-a",
+        "--all",
+        action="store_true",
+        dest="all_presets",
+        help="Show all presets. By default, it only shows unfinished creations or the last one.",
+    )
+    parser.add_argument(
+        "-n",
+        "--last",
+        metavar="COUNT",
+        type=int,
+        dest="limit",
+        help="Show only the last N presets. Implies --all",
+    )
     model_filter = parser.add_mutually_exclusive_group()
     model_filter.add_argument(
         "--base",

diff --git a/src/dstack/_internal/cli/models/configurations.py b/src/dstack/_internal/cli/models/configurations.py
@@ -12,6 +12,9 @@
 from dstack._internal.utils.json_schema import add_extra_schema_types
 
 DEFAULT_CONCURRENCY = 8
+DEFAULT_INPUT_TOKENS = 1024
+DEFAULT_OUTPUT_TOKENS = 1024
+DEFAULT_BASELINE = False
 
 
 class PresetModelRepo(CoreModel):
@@ -134,7 +137,7 @@ class PresetConfiguration(
             )
         ),
     ] = None
-    context_length: Annotated[
+    min_context_length: Annotated[
         Optional[PositiveInt], Field(description="The minimum required context length")
     ] = None
     max_trials: Annotated[
@@ -155,6 +158,34 @@ class PresetConfiguration(
             )
         ),
     ] = None
+    input_tokens: Annotated[
+        Optional[PositiveInt],
+        Field(
+            description=(
+                "The number of input tokens per request used for benchmarks during"
+                f" preset creation. Defaults to `{DEFAULT_INPUT_TOKENS}`"
+            )
+        ),
+    ] = None
+    output_tokens: Annotated[
+        Optional[PositiveInt],
+        Field(
+            description=(
+                "The number of output tokens per request used for benchmarks during"
+                f" preset creation. Defaults to `{DEFAULT_OUTPUT_TOKENS}`"
+            )
+        ),
+    ] = None
+    baseline: Annotated[
+        Optional[bool],
+        Field(
+            description=(
+                "Whether the first trial must be a baseline that serves the model with the"
+                " serving framework's recommended defaults instead of an optimization attempt."
+                " Defaults to `false`"
+            )
+        ),
+    ] = None
     gateway: Annotated[
         Optional[Union[bool, EntityReference, str]],
         Field(
@@ -173,6 +204,18 @@ class PresetConfiguration(
     def effective_concurrency(self) -> int:
         return self.concurrency if self.concurrency is not None else DEFAULT_CONCURRENCY
 
+    @property
+    def effective_input_tokens(self) -> int:
+        return self.input_tokens if self.input_tokens is not None else DEFAULT_INPUT_TOKENS
+
+    @property
+    def effective_output_tokens(self) -> int:
+        return self.output_tokens if self.output_tokens is not None else DEFAULT_OUTPUT_TOKENS
+
+    @property
+    def effective_baseline(self) -> bool:
+        return self.baseline if self.baseline is not None else DEFAULT_BASELINE
+
     @root_validator(pre=True)
     def apply_model_shorthand(cls, values: Any) -> Any:
         if not isinstance(values, dict):
@@ -211,9 +254,12 @@ class PresetConstraints(CoreModel):
 
     run_name_prefix: str
     model: PresetModelSpec
-    context_length: Optional[PositiveInt] = None
+    min_context_length: Optional[PositiveInt] = None
     max_trials: PositiveInt
     concurrency: PositiveInt
+    input_tokens: PositiveInt
+    output_tokens: PositiveInt
+    baseline: bool = False
     fleets: list[str] = Field(min_items=1)
     env: list[str] = []
 

diff --git a/src/dstack/_internal/cli/services/presets/agent.py b/src/dstack/_internal/cli/services/presets/agent.py
@@ -423,6 +423,14 @@ async def _session_tailers(
             offset_key="trials",
             echo=agent_session.echo,
         ),
+        _RecordMirror(
+            source=workspace.verifications_path,
+            target=agent_session.verifications_path,
+            redacted_values=redacted_values,
+            offset_store=offset_store,
+            offset_key="verifications",
+            echo=agent_session.echo,
+        ),
     ]
     tailer_tasks = [
         asyncio.create_task(tailer.run()) for tailer in [progress_tailer, *record_mirrors]

diff --git a/src/dstack/_internal/cli/services/presets/apply.py b/src/dstack/_internal/cli/services/presets/apply.py
@@ -8,6 +8,7 @@
 from dstack._internal.cli.services.configurators.run import ServiceConfigurator
 from dstack._internal.cli.services.presets.output import (
     format_preset_benchmark,
+    format_preset_objective,
 )
 from dstack._internal.cli.services.presets.store import PresetStore
 from dstack._internal.core.errors import CLIError
@@ -58,11 +59,11 @@ def _validate_preset_matches(preset: Preset, *, configuration: PresetConfigurati
     service_model = preset.service.model
     if service_model is None or service_model.name.lower() != model_name.lower():
         raise CLIError(f"Preset {preset.id} does not serve {model_name}")
-    if configuration.context_length is not None:
-        if preset.context_length < configuration.context_length:
+    if configuration.min_context_length is not None:
+        if preset.context_length < configuration.min_context_length:
             raise CLIError(
                 f"Preset {preset.id} does not support context length"
-                f" {configuration.context_length}"
+                f" {configuration.min_context_length}"
             )
     if configuration.model.allows_variant_selection:
         if preset.base.lower() != model_name.lower():
@@ -94,5 +95,7 @@ def _format_requested_model(configuration: PresetConfiguration) -> str:
 
 
 def _format_selected_preset(preset: Preset) -> str:
-    details = format_preset_benchmark(preset, verbose=True)
-    return f"{escape(preset.id)} ([secondary]{details}[/])"
+    # The formatter dims its own keys; wrapping it again would flatten that.
+    # One line, so the objective and the result are joined rather than columned.
+    details = f"{format_preset_objective(preset)} {format_preset_benchmark(preset, verbose=True)}"
+    return f"{escape(preset.id)} ({details})"
diff --git a/src/dstack/_internal/cli/services/presets/create.py b/src/dstack/_internal/cli/services/presets/create.py
@@ -554,7 +554,10 @@ async def _create_preset(
             workspace=setup.workspace,
             token=token,
         )
-    prompt = get_preset_agent_system_prompt(user_prompt=setup.user_prompt)
+    prompt = get_preset_agent_system_prompt(
+        user_prompt=setup.user_prompt,
+        baseline=configuration.effective_baseline,
+    )
     if setup.write_constraints:
         if setup.user_prompt:
             agent_session.write_user_prompt(setup.user_prompt)
@@ -849,9 +852,12 @@ def _build_constraints(
         {
             "run_name_prefix": build_name,
             "model": json.loads(configuration.model.json(exclude_none=True)),
-            "context_length": configuration.context_length,
+            "min_context_length": configuration.min_context_length,
             "max_trials": configuration.max_trials,
             "concurrency": configuration.effective_concurrency,
+            "input_tokens": configuration.effective_input_tokens,
+            "output_tokens": configuration.effective_output_tokens,
+            "baseline": configuration.effective_baseline,
             "fleets": list(allowed_fleets),
             "env": list(configuration.env),
         }