diff --git a/deploy/docker-compose.validator.yml b/deploy/docker-compose.validator.yml
index 389e4c9..fb844ad 100644
--- a/deploy/docker-compose.validator.yml
+++ b/deploy/docker-compose.validator.yml
@@ -17,6 +17,10 @@ services:
       MT_RELAY_URL: "${MT_RELAY_URL:-https://api.microtensor.cloud}"
       MT_GPUS: "${MT_GPUS:-0,1,2}"
       MT_SANDBOX_RUNTIME: "${MT_SANDBOX_RUNTIME:-runsc}"
+      MT_JUDGE_SHARE_GPU: "${MT_JUDGE_SHARE_GPU:-}"
+      MT_JUDGE_MODEL: "${MT_JUDGE_MODEL:-}"
+      MT_JUDGE_REVISION: "${MT_JUDGE_REVISION:-}"
+      MT_SLM_MEMORY_SHARE: "${MT_SLM_MEMORY_SHARE:-0}"
       HF_HOME: /var/lib/microtensor/cache/hf
     volumes:
       - /var/run/docker.sock:/var/run/docker.sock
diff --git a/deploy/validator-entrypoint.sh b/deploy/validator-entrypoint.sh
index 3715b95..25a6107 100644
--- a/deploy/validator-entrypoint.sh
+++ b/deploy/validator-entrypoint.sh
@@ -6,8 +6,8 @@ if ! docker info >/dev/null 2>&1; then
     exit 1
 fi
 
-if ! docker info --format '{{json .Runtimes}}' | grep -q runsc; then
-    echo "gVisor (runsc) is not registered with Docker; miner code will not run" >&2
+if [ "${MT_SANDBOX_RUNTIME:-runsc}" = "runsc" ] && ! docker info --format '{{json .Runtimes}}' | grep -q runsc; then
+    echo "gVisor (runsc) is not registered with Docker; set MT_SANDBOX_RUNTIME=runc to run miner code in plain containers" >&2
     exit 1
 fi
 
diff --git a/docs/validator_setup.md b/docs/validator_setup.md
index 1b2adea..7868063 100644
--- a/docs/validator_setup.md
+++ b/docs/validator_setup.md
@@ -116,6 +116,20 @@ docker compose -f docker-compose.validator.yml run --rm validator epoch run --ar
 
 Evaluates twenty items per test, posts no report and sets no weights. Use it after installation and after every image update.
 
+## Testnet on one GPU
+
+The published mechanism gives the context compression probe agent its own card. On testnet a validator may run everything on one GPU and swap the probe agent for a quantised build of the same model so both fit:
+
+```
+MT_GPUS=0
+MT_SANDBOX_RUNTIME=runc
+MT_JUDGE_SHARE_GPU=1
+MT_JUDGE_MODEL=Qwen/Qwen2.5-Coder-7B-Instruct-AWQ
+MT_SLM_MEMORY_SHARE=0.38
+```
+
+The probe agent then takes half the card and the model under test the share you set. `MT_JUDGE_MODEL` is refused on mainnet; `MT_SANDBOX_RUNTIME=runc` runs miner code in a plain container, which is acceptable only on testnet.
+
 ## Updating
 
 ```bash
diff --git a/microtensor_subnet/cli/epoch.py b/microtensor_subnet/cli/epoch.py
index bb9e214..153064f 100644
--- a/microtensor_subnet/cli/epoch.py
+++ b/microtensor_subnet/cli/epoch.py
@@ -95,6 +95,7 @@ def register(subparsers: argparse._SubParsersAction[argparse.ArgumentParser]) ->
     )
     run.add_argument("--concurrency", type=int, default=3)
     run.add_argument("--sandbox-runtime", default=os.environ.get("MT_SANDBOX_RUNTIME", "runsc"))
+    _add_gpu_sharing(run)
     run.add_argument("--item-limit", type=int, default=0, help="development only")
     run.add_argument("--dry-run", action="store_true", help="evaluate and report, set no weights")
     add_chain_arguments(run)
@@ -110,6 +111,7 @@ def register(subparsers: argparse._SubParsersAction[argparse.ArgumentParser]) ->
     )
     loop.add_argument("--concurrency", type=int, default=3)
     loop.add_argument("--sandbox-runtime", default=os.environ.get("MT_SANDBOX_RUNTIME", "runsc"))
+    _add_gpu_sharing(loop)
     loop.add_argument("--poll-seconds", type=int, default=600)
     loop.add_argument("--once", action="store_true", help="one poll, then exit")
     loop.add_argument("--item-limit", type=int, default=0, help="development only")
@@ -488,6 +490,27 @@ def _probe(args: argparse.Namespace) -> int:
             close_reference(docker, reference)
 
 
+def _add_gpu_sharing(parser: argparse.ArgumentParser) -> None:
+    parser.add_argument(
+        "--judge-share-gpu",
+        action="store_true",
+        default=os.environ.get("MT_JUDGE_SHARE_GPU", "") == "1",
+        help="run the probe agent on the same GPU as the model under test",
+    )
+    parser.add_argument(
+        "--judge-model",
+        default=os.environ.get("MT_JUDGE_MODEL", ""),
+        help="testnet only: a different probe agent repository",
+    )
+    parser.add_argument("--judge-revision", default=os.environ.get("MT_JUDGE_REVISION", ""))
+    parser.add_argument(
+        "--slm-memory-share",
+        type=float,
+        default=float(os.environ.get("MT_SLM_MEMORY_SHARE", "0") or 0),
+        help="vLLM memory share for the model under test; 0 picks the default",
+    )
+
+
 def _settings(args: argparse.Namespace, config: Any) -> EpochConfig:
     return EpochConfig(
         chain=config,
@@ -500,6 +523,10 @@ def _settings(args: argparse.Namespace, config: Any) -> EpochConfig:
         sandbox_runtime=args.sandbox_runtime,
         dry_run=args.dry_run,
         item_limit=args.item_limit,
+        judge_share_gpu=bool(getattr(args, "judge_share_gpu", False)),
+        judge_model=str(getattr(args, "judge_model", "") or ""),
+        judge_revision=str(getattr(args, "judge_revision", "") or ""),
+        slm_memory_share=float(getattr(args, "slm_memory_share", 0.0) or 0.0),
     )
 
 
diff --git a/microtensor_subnet/serving/host/config.py b/microtensor_subnet/serving/host/config.py
index d7d7057..d989d2f 100644
--- a/microtensor_subnet/serving/host/config.py
+++ b/microtensor_subnet/serving/host/config.py
@@ -32,6 +32,7 @@ class HostConfig:
     sandbox_image: str = SANDBOX_IMAGE
     gateway_image: str = GATEWAY_IMAGE
     sandbox_runtime: str = "runsc"
+    slm_memory_share: float = 0.0
     sandbox_cpus: float = SANDBOX_CPUS
     sandbox_memory: str = SANDBOX_MEMORY
     sandbox_pids: int = SANDBOX_PIDS
diff --git a/microtensor_subnet/validator/config.py b/microtensor_subnet/validator/config.py
index f0ac38b..ee5c80e 100644
--- a/microtensor_subnet/validator/config.py
+++ b/microtensor_subnet/validator/config.py
@@ -42,6 +42,10 @@ class EpochConfig:
     slm_ready_seconds: int = SLM_READY_SECONDS
     host_services_port: int = 9701
     allow_unsandboxed: bool = False
+    judge_share_gpu: bool = False
+    judge_model: str = ""
+    judge_revision: str = ""
+    slm_memory_share: float = 0.0
     dry_run: bool = False
     item_limit: int = 0
     extra: dict[str, str] = field(default_factory=dict)
diff --git a/microtensor_subnet/validator/epoch.py b/microtensor_subnet/validator/epoch.py
index fc5624a..c72463a 100644
--- a/microtensor_subnet/validator/epoch.py
+++ b/microtensor_subnet/validator/epoch.py
@@ -8,7 +8,7 @@
 from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Any
+from typing import Any, Final
 
 from microtensor_subnet.artifact.params import ParameterError, footprint_bytes
 from microtensor_subnet.chain.epochs import Epoch, epoch_seed
@@ -43,9 +43,10 @@
 from microtensor_subnet.validator.sandbox import tools as tool_faults
 from microtensor_subnet.validator.sandbox.docker import Docker
 from microtensor_subnet.validator.sandbox.services import Services
-from microtensor_subnet.validator.sandbox.slm import SlmClient
+from microtensor_subnet.validator.sandbox.slm import MEMORY_SHARE, SlmClient
 
 log = logging.getLogger("microtensor_subnet.validator.epoch")
+JUDGE_SHARED_SHARE: Final[float] = 0.5
 DEFAULT_CONTEXT = 8192
 DEFAULT_MEMORY_BYTES = 24 * 1024**3
 DEFAULT_MAX_PARAMETERS = 8_000_000_000
@@ -587,9 +588,12 @@ def run_epoch(
     gpus = tuple(config.gpus or ("0",))
     agent: probe_judge.ProbeAgent | None = None
     if context.grader == compress.GRADER:
-        if len(gpus) < 2:
-            raise EpochError("context compression needs a second GPU for the probe agent")
-        agent, gpus = _judge(config, context, engine, gpus[-1]), gpus[:-1]
+        if config.judge_share_gpu:
+            agent = _judge(config, context, engine, gpus[-1], memory_share=JUDGE_SHARED_SHARE)
+        else:
+            if len(gpus) < 2:
+                raise EpochError("context compression needs a second GPU for the probe agent")
+            agent, gpus = _judge(config, context, engine, gpus[-1]), gpus[:-1]
     services.start()
     host = SandboxHost(config, engine, services, host_ip())
     runner = Runner(config, context, host, relay, GpuPool(gpus), desk)
@@ -721,7 +725,12 @@ def _owner(chain: Any) -> str:
 
 
 def _judge(
-    config: EpochConfig, context: EpochContext, engine: Docker, gpu: str
+    config: EpochConfig,
+    context: EpochContext,
+    engine: Docker,
+    gpu: str,
+    *,
+    memory_share: float = MEMORY_SHARE,
 ) -> probe_judge.ProbeAgent:
     grading = context.grading
     spec = grading.get("probe_agent") or {}
@@ -729,6 +738,11 @@ def _judge(
     if not model:
         raise EpochError("the arena names no probe agent")
     revision = str(spec.get("revision") or "main")
+    if config.judge_model:
+        if config.chain.network != "test":
+            raise EpochError("a probe agent override is only allowed on testnet")
+        log.warning("testnet probe agent override: %s instead of %s", config.judge_model, model)
+        model, revision = config.judge_model, config.judge_revision or "main"
     model_dir = collect.download(model, revision, config.cache_dir)
     tokenizer = str(grading.get("tokenizer") or model)
     tokenizer_dir = (
@@ -740,6 +754,7 @@ def _judge(
         model_dir,
         gpu,
         max_tokens=int(spec.get("max_tokens") or 64),
+        memory_share=memory_share,
     )
     agent.start(config.slm_ready_seconds)
     context.judge = compress.Judge(
diff --git a/microtensor_subnet/validator/sandbox/__init__.py b/microtensor_subnet/validator/sandbox/__init__.py
index d175a16..8268e93 100644
--- a/microtensor_subnet/validator/sandbox/__init__.py
+++ b/microtensor_subnet/validator/sandbox/__init__.py
@@ -59,6 +59,9 @@ def gateway_image(self) -> str: ...
     @property
     def sandbox_runtime(self) -> str: ...
 
+    @property
+    def slm_memory_share(self) -> float: ...
+
     @property
     def sandbox_cpus(self) -> float: ...
 
@@ -206,7 +209,9 @@ def _slm_container(
                 cpu_container(name, self.config.llama_image, model_dir, max_len, limits, network)
             )
             return name
-        share = SHARED_MEMORY_SHARE if limits.router_on_gpu else MEMORY_SHARE
+        share = self.config.slm_memory_share or (
+            SHARED_MEMORY_SHARE if limits.router_on_gpu else MEMORY_SHARE
+        )
         self.docker.start(
             ContainerSpec(
                 name=name,
diff --git a/microtensor_subnet/validator/sandbox/judge.py b/microtensor_subnet/validator/sandbox/judge.py
index 4780fb0..978b006 100644
--- a/microtensor_subnet/validator/sandbox/judge.py
+++ b/microtensor_subnet/validator/sandbox/judge.py
@@ -11,7 +11,12 @@
 import httpx
 
 from microtensor_subnet.validator.sandbox.docker import ContainerSpec, Docker
-from microtensor_subnet.validator.sandbox.slm import MODEL_NAME, SlmClient, vllm_command
+from microtensor_subnet.validator.sandbox.slm import (
+    MEMORY_SHARE,
+    MODEL_NAME,
+    SlmClient,
+    vllm_command,
+)
 
 PROMPT: Final[str] = (
     "You read a summary of a coding agent's session and answer one question about that session "
@@ -36,6 +41,7 @@ class ProbeAgent:
     model_dir: Path
     gpu: str
     max_tokens: int = 64
+    memory_share: float = MEMORY_SHARE
     name: str = ""
     network: str = ""
     url: str = ""
@@ -50,7 +56,9 @@ def start(self, ready_seconds: int) -> None:
                 name=self.name,
                 image=self.image,
                 network=self.network,
-                command=vllm_command("/model", max_model_len=MAX_MODEL_LEN),
+                command=vllm_command(
+                    "/model", max_model_len=MAX_MODEL_LEN, memory_share=self.memory_share
+                ),
                 env={"HF_HUB_OFFLINE": "1", "VLLM_NO_USAGE_STATS": "1"},
                 volumes=((str(self.model_dir), "/model", "ro"),),
                 gpus=(self.gpu,),