Add MiniCPM5-2B cookbook (#38295)

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-09-07 21:30:17 +08:00
parent 62a4a6ea0e
commit e4008de757
6 changed files with 511 additions and 2 deletions
@@ -2,7 +2,6 @@
 title: MiniCPM-V 4.6
 metatags:
    description: "Deploy OpenBMB MiniCPM-V 4.6 (Qwen3.5-style hybrid GDN backbone + NaViT vision encoder) on NVIDIA GPUs with SGLang — multimodal text + image + video, slicing for high-resolution images."
-tag: NEW
 ---

 ## 1. Model Introduction
@@ -0,0 +1,220 @@
+---
+title: MiniCPM5-2B
+description: "Deploy OpenBMB MiniCPM5-2B with SGLang — a 2.5B dense Llama-architecture on-device model with 131K context, thinking mode, XML tool calling and DSpark speculative decoding on H200, RTX PRO 6000, RTX 5090 and DGX Spark."
+tag: NEW
+---
+
+## Deployment
+
+<a id="install" />
+
+<Accordion title="Install SGLang">
+
+For all methods and hardware platforms, see the [official SGLang installation guide](../../../docs/get-started/install). The two paths below match the **Python / Docker** toggle in the command panel.
+
+<Tabs>
+
+<Tab title="Python (pip / uv)">
+
+```bash Command
+pip install --upgrade pip
+pip install uv
+uv pip install --prerelease=allow "sglang>=0.5.12"
+```
+
+Then run the **Python** output of the command panel below in that environment.
+
+</Tab>
+
+<Tab title="Docker">
+
+```bash Command
+docker pull lmsysorg/sglang:dev
+```
+
+For how to launch the image, see [Install → Method 3: Using Docker](../../../docs/get-started/install#method-3-using-docker). Substitute the inner `sglang serve ...` with what the command generator below produces.
+
+</Tab>
+
+</Tabs>
+
+</Accordion>
+
+Pick your card to generate the launch command. MiniCPM5-2B is a 2.5B dense model and runs single-GPU at TP=1 on every supported card, so the page ships one operating point per card plus an optional **DSpark** speculative-decoding row.
+
+import { Deployment } from "/src/snippets/_deployment.jsx";
+import { config }     from "/src/snippets/configs/openbmb/minicpm5-2b.jsx";
+import { benchmarks } from "/src/snippets/configs/openbmb/minicpm5-2b-benchmarks.jsx";
+
+<Deployment config={config} benchmarks={benchmarks} />
+
+<Note>
+  Speed numbers exist for the RTX 5090 and DGX Spark cells — a single card at
+  random 1024/1024, recorded as Mean. The H200 and RTX PRO 6000 cells are
+  pending measurement, and no accuracy numbers have been taken on any platform
+  yet. The DSpark overlay carries no speed numbers on any card.
+</Note>
+
+## Playground
+
+The Playground is where you experiment with **SGLang features beyond the recipes above**. The Deploy panel emits this model's documented launch recipes; the Playground lets you turn on additional knobs on top of whichever cell the Deploy panel is currently showing.
+
+import { Playground } from "/src/snippets/_playground.jsx";
+
+<Playground config={config} />
+
+## 1. Model Introduction
+
+**MiniCPM5-2B** is the second model in OpenBMB's MiniCPM5 series, following MiniCPM5-1B. It is a dense 2B-class Transformer built for on-device assistants, local deployment, coding agents and tool-use workflows — scenarios where a compact model with a small deployment footprint is preferred. OpenBMB reports 2B-class open-source SOTA within its comparison set (average 53.9), with its clearest advantages in code reasoning, math reasoning, long-context understanding, tool use and agentic tasks.
+
+The checkpoint uses the standard `LlamaForCausalLM` architecture — no custom kernels and no model-code fork — so SGLang loads it through the stock Llama path. It is 2,516,756,480 parameters (1,981,982,720 non-embedding) over 42 layers with GQA (16 query heads, 2 KV heads), and a native context length of 131,072 tokens. Post-training runs SFT → RL → On-Policy Distillation, which merges 16 RL expert models into the single released checkpoint. Weights are released under the [Apache-2.0](https://github.com/OpenBMB/MiniCPM/blob/main/LICENSE) license.
+
+<table style={{width: "100%", borderCollapse: "collapse", tableLayout: "fixed"}}>
+  <colgroup>
+    <col style={{width: "34%"}} />
+    <col style={{width: "22%"}} />
+    <col style={{width: "44%"}} />
+  </colgroup>
+  <thead>
+    <tr style={{borderBottom: "2px solid #d55816"}}>
+      <th style={{textAlign: "left", padding: "10px 12px", fontWeight: 700}}>Checkpoint</th>
+      <th style={{textAlign: "left", padding: "10px 12px", fontWeight: 700}}>Precision</th>
+      <th style={{textAlign: "left", padding: "10px 12px", fontWeight: 700}}>Role on this page</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><a href="https://huggingface.co/openbmb/MiniCPM5-2B">openbmb/MiniCPM5-2B</a></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>BF16</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>The served model in every cell above (final release, post-trained with RL + OPD).</td>
+    </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.05)"}}><a href="https://huggingface.co/openbmb/MiniCPM5-2B-DSpark">openbmb/MiniCPM5-2B-DSpark</a></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>BF16 draft</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Draft model loaded by the <strong>DSPARK</strong> row of the Deploy panel; not served on its own.</td>
+    </tr>
+  </tbody>
+</table>
+
+OpenBMB also publishes SFT-only, mid-training and base checkpoints, plus GGUF, MLX and GPTQ-Int4 exports for llama.cpp / Ollama / LM Studio / Apple Silicon. Those target other runtimes and are not part of the SGLang matrix above.
+
+**Recommended generation:** `temperature=1.0`, `top_p=0.95` (informational — SGLang reads the checkpoint's `generation_config.json`; do not hardcode these in client code).
+
+**Resources:** [HuggingFace](https://huggingface.co/openbmb/MiniCPM5-2B) · [ModelScope](https://www.modelscope.cn/models/OpenBMB/MiniCPM5-2B) · [GitHub](https://github.com/OpenBMB/MiniCPM) · [Tech report](https://arxiv.org/pdf/2506.07900).
+
+## 2. Advanced Usage
+
+### 2.1 Thinking Mode
+
+Thinking is controlled by the chat template's `enable_thinking` flag, passed per request through `chat_template_kwargs`. The `qwen3` reasoning parser — already in every generated command, and toggleable from the **Parsers** card in the [Playground above](#playground) — splits the `<think>` segment into `message.reasoning_content` and leaves the final answer in `message.content`. Drop the flag and the thinking text stays inline in `content`, closing tag and all.
+
+<Accordion title="Thinking On / Off (Python)">
+
+```python Example
+from openai import OpenAI
+
+client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
+
+# Thinking on (chat-template default).
+resp = client.chat.completions.create(
+    model="openbmb/MiniCPM5-2B",
+    messages=[{"role": "user", "content": "What is 15% of 240?"}],
+    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
+)
+msg = resp.choices[0].message
+print("Reasoning:", getattr(msg, "reasoning_content", None))
+print("Answer:", msg.content)
+
+# Thinking off — a direct answer, no deliberation segment.
+resp = client.chat.completions.create(
+    model="openbmb/MiniCPM5-2B",
+    messages=[{"role": "user", "content": "What is 15% of 240?"}],
+    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
+)
+print("Without thinking:", resp.choices[0].message.content)
+```
+
+</Accordion>
+
+<Accordion title="Example Output">
+
+```text Output
+Reasoning: We are asked: "What is 15% of 240?" This is a simple percentage calculation. To find 15% of 240, we multiply 240 by 15% (which is 0.15). So: 240 × 0.15 = 36. Alternatively, we can think of it as (15/100) × 240 = (15 × 240)/100 = 3600/100 = 36. So the answer is 36.
+
+We need to respond in a helpful way. The user might be testing or seeking quick answer. Let's provide a clear response with explanation if needed, but since it's straightforward, we can just give the answer directly or briefly explain.
+
+Answer:
+
+15% of 240 is **36**.
+
+To calculate:
+\( 240 \times 0.15 = 36 \)
+Reasoning: None
+Answer: To find 15% of 240, multiply 240 by 15% (which is 0.15):
+
+\[
+240 \times 0.15 = 36
+\]
+
+So, 15% of 240 is **36**.
+```
+
+</Accordion>
+
+### 2.2 Tool Calling
+
+MiniCPM5-2B emits XML-style tool calls (`<function name="..."><param name="...">...</param></function>`), and SGLang's built-in `minicpm5` detector converts them to OpenAI-compatible `tool_calls`, parallel calls included. The parser is already in every command the Deploy panel generates; the **Tool Call Parser** chip in the Playground's **Parsers** card is an opt-out, not an opt-in.
+
+<Accordion title="Tool Calling (Python)">
+
+```python Example
+from openai import OpenAI
+
+client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
+
+tools = [{
+    "type": "function",
+    "function": {
+        "name": "get_weather",
+        "description": "Get the current weather for a city.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "city": {"type": "string", "description": "City name"},
+                "date": {"type": "string", "description": "YYYY-MM-DD"},
+            },
+            "required": ["city"],
+        },
+    },
+}]
+
+resp = client.chat.completions.create(
+    model="openbmb/MiniCPM5-2B",
+    messages=[{"role": "user", "content": "What is the weather in Beijing today?"}],
+    tools=tools,
+)
+
+msg = resp.choices[0].message
+# With the qwen3 reasoning parser on, the thinking segment lands in
+# `reasoning_content` and `content` may be empty on a tool-call turn.
+print("Reasoning:", getattr(msg, "reasoning_content", None))
+print("Content:", msg.content)
+for call in (msg.tool_calls or []):
+    print("Tool:", call.function.name, call.function.arguments)
+```
+
+</Accordion>
+
+<Accordion title="Example Output">
+
+```text Output
+Reasoning:
+The user is asking about the weather in Beijing today. I have access to a get_weather function that can help with this. Let me call it to get the current weather for Beijing.
+
+Content:
+
+
+Tool: get_weather {"city": "Beijing"}
+```
+
+</Accordion>
@@ -136,7 +136,7 @@ metatags:
  <Card
    title="OpenBMB"
    mode="card"
-    href="/cookbook/autoregressive/OpenBMB/MiniCPM-V-4_6"
+    href="/cookbook/autoregressive/OpenBMB/MiniCPM5-2B"
    img="/cards/logos/openbmb.png"
  />
  <Card
@@ -1404,6 +1404,7 @@
                  {
                    "group": "OpenBMB",
                    "pages": [
+                      "cookbook/autoregressive/OpenBMB/MiniCPM5-2B",
                      "cookbook/autoregressive/OpenBMB/MiniCPM-V-4_6"
                    ]
                  },
@@ -0,0 +1,55 @@
+// One entry per cell `match` tuple (same keys as the config cells).
+//
+// Only the RTX 5090 row carries numbers. They were measured on a single
+// RTX 5090 32GB (`zijiexia-5090-1gpu`, `lmsysorg/sglang:dev` @ 30705c004c) with
+// `python3 -m sglang.bench_serving`, random 1024/1024, against the exact cell
+// command on this page — including `--mem-fraction-static 0.75
+// --cuda-graph-max-bs 128`, which is what keeps concurrency 64 and 128
+// graph-backed (see the model page's Configuration Tips).
+//
+// `tokens_per_sec_per_gpu` is total (input+output) per GPU = the measured
+// output throughput x 2 at 1024/1024 on one GPU.
+//
+// Two operating points only: concurrency 1 (single-user latency) and 128 (the
+// saturated end of the sweep). The intermediate 8 / 32 / 64 rows were measured
+// but are not published here.
+//
+// The source runs did not restate a percentile, so the values are recorded as
+// Mean (bench_serving's headline lines) rather than claimed as P50 — same
+// convention as the LFM2.5 page. TTFT was not captured at concurrency 128, so
+// that cell carries `null` rather than a back-filled guess.
+//
+// No accuracy rows: the only GSM8K numbers taken so far came from a custom
+// chat-template harness, not the `sgl-eval run gsm8k` command the Reproduce
+// modal would show, and the two are not comparable.
+
+export const benchmarks = [
+  {
+    match: { hw: "rtx5090", variant: "default", quant: "bf16", nodes: "single" },
+    sglang_version: "dev @ 30705c004c",
+    latencyPercentile: "Mean",
+    speed: [
+      { workload: { dataset: "random", isl: 1024, osl: 1024, max_concurrency: 1 },
+        ttft_ms: 34, tpot_ms: 4.0, tokens_per_sec_per_gpu: 496 },
+      { workload: { dataset: "random", isl: 1024, osl: 1024, max_concurrency: 128 },
+        ttft_ms: null, tpot_ms: 11.8, tokens_per_sec_per_gpu: 19280 },
+    ],
+  },
+  {
+    // Single DGX Spark (GB10), same image and commit as the RTX 5090 row.
+    // TTFT is Mean; at concurrency 64 the Median is 246 ms, because the first
+    // wave of requests all queue behind one prefill.
+    match: { hw: "dgx-spark", variant: "default", quant: "bf16", nodes: "single" },
+    sglang_version: "dev @ 30705c004c",
+    latencyPercentile: "Mean",
+    speed: [
+      { workload: { dataset: "random", isl: 1024, osl: 1024, max_concurrency: 1 },
+        ttft_ms: 85, tpot_ms: 27.7, tokens_per_sec_per_gpu: 72 },
+      { workload: { dataset: "random", isl: 1024, osl: 1024, max_concurrency: 64 },
+        ttft_ms: 1373, tpot_ms: 42.9, tokens_per_sec_per_gpu: 2892 },
+    ],
+  },
+  // Pending — no numbers taken on these platforms yet.
+  { match: { hw: "h200",    variant: "default", quant: "bf16", nodes: "single" } },
+  { match: { hw: "rtx6000", variant: "default", quant: "bf16", nodes: "single" } },
+];
@@ -0,0 +1,234 @@
+// Single `export const config` literal — no spreads/calls/IIFE (Mintlify re-evals at hydration).
+// Cells are denormalized: no `--nnodes`/`--node-rank`/`--dist-init-addr`/`--host`/`--port` literals — engine injects them.
+//
+// MiniCPM5-2B: 2.5B dense `LlamaForCausalLM` (42 layers, GQA 16Q/2KV, 131072
+// context). Standard architecture — no custom kernels and no model-code fork —
+// so every supported card runs it single-GPU at TP=1 with the stock backend.
+// That leaves one recipe per card, hence a single variant / quantization /
+// strategy / node option and no parallelism flags in any cell.
+//
+// Recipes are the OpenBMB model card's SGLang commands, rewritten from
+// `python -m sglang.launch_server` to `sglang serve`, plus the parser pair the
+// model needs to be usable through the OpenAI API. Both are baked into every
+// cell, so the Parsers card in the Playground reads as an opt-OUT:
+//   --tool-call-parser minicpm5   the model emits XML-style
+//     `<function name="f"><param name="p">v</param></function>`; without the
+//     detector `tool_calls` comes back None and the XML lands in `content`.
+//   --reasoning-parser qwen3      the chat template is Qwen-style
+//     (`<|im_start|>` + `<think>`) and there is no `minicpm5` reasoning
+//     detector, so `qwen3` is the one that applies; without it `</think>`
+//     leaks into `content`.
+//
+// DSpark is the separately published draft checkpoint
+// (openbmb/MiniCPM5-2B-DSpark). It is orthogonal to the card grid, so it is an
+// overlay row rather than a match dim. No DSpark speed numbers are published:
+// the speedup tracks acceptance length, which moves with the prompt
+// distribution, and a random-token dataset inflates it above real traffic.
+
+export const config = {
+  modelName: "MiniCPM5-2B",
+
+  supportedHardware: ["h200", "rtx6000", "rtx5090", "dgx-spark"],
+
+  // RTX PRO 6000 and RTX 5090 (SM120 / Blackwell workstation + desktop) are not
+  // datacenter parts, so the shared HARDWARE_CATALOG in _deployment.jsx does not
+  // carry them. Ids/labels match the DeepSeek-V4 and Qwen3.8-27B configs.
+  hardware: [
+    { id: "rtx6000", label: "RTX PRO 6000", vram: "96GB", vendor: "blackwell" },
+    { id: "rtx5090", label: "RTX 5090", vram: "32GB", vendor: "blackwell" },
+  ],
+
+  matchDims: [
+    { id: "variant", title: "Model Variant", options: [
+      { id: "default", label: "Default" },
+    ] },
+    { id: "quant", title: "Quantization", options: [
+      { id: "bf16", label: "BF16" },
+    ] },
+    { id: "nodes", title: "Nodes", options: [
+      { id: "single", label: "Single Node" },
+    ] },
+  ],
+
+  overlayDims: [
+    {
+      id: "spec",
+      title: "Speculative Decoding",
+      default: "none",
+      options: [
+        { id: "none", label: "None" },
+        {
+          id: "dspark", label: "DSPARK",
+          // Verbatim from the model card's DSpark command, including
+          // `--trust-remote-code`: the base checkpoint is plain Llama and does
+          // not need it, the draft checkpoint's config does. gamma = 7, so the
+          // verify window is 8 tokens.
+          flags: [
+            "--trust-remote-code",
+            "--speculative-algorithm DSPARK",
+            "--speculative-draft-model-path openbmb/MiniCPM5-2B-DSpark",
+            "--speculative-dspark-block-size 7",
+          ],
+        },
+      ],
+    },
+  ],
+
+  modelNames: {
+    "default|bf16": "openbmb/MiniCPM5-2B",
+  },
+
+  placeholders: {
+    HOST_IP:   { target: "command", label: "Bind host",         default: "0.0.0.0"  },
+    PORT:      { target: "command", label: "Bind port",         default: "30000"    },
+    HF_TOKEN:  { target: "command", label: "HF token (Docker)", default: "<your-hf-token>" },
+    CURL_HOST: { target: "curl",    label: "Server host",       default: "localhost" },
+    CURL_PORT: { target: "curl",    label: "Server port",       default: "30000"     },
+  },
+
+  curl: `curl http://{{CURL_HOST}}:{{CURL_PORT}}/v1/chat/completions \\
+-H 'Content-Type: application/json' \\
+-d '{ "model": "{{MODEL_NAME}}", "messages": [{"role":"user","content":"Who are you? Please briefly introduce yourself."}] }'`,
+
+  // Reproduce command for the Benchmark card's "⚡ Reproduce" modal. No
+  // `accuracy` entry: the page carries no accuracy numbers yet.
+  benchmarkCommands: {
+    speed:
+`python3 -m sglang.bench_serving \\
+  --backend sglang \\
+  --host {{CURL_HOST}} --port {{CURL_PORT}} \\
+  --model {{MODEL_NAME}} \\
+  --dataset-name {{DATASET}} \\
+  --random-input-len {{ISL}} --random-output-len {{OSL}} \\
+  --num-prompts {{NUM_PROMPTS}} --max-concurrency {{MAX_CONCURRENCY}} \\
+  --flush-cache`,
+    numPromptsByConc: { 1: 10, 128: 512 },
+  },
+
+  // MiniCPM5 support (the `minicpm5` tool-call parser and the DSPARK draft
+  // worker) ships in the SGLang dev image.
+  dockerImages: {
+    h200:    "lmsysorg/sglang:dev",
+    rtx6000: "lmsysorg/sglang:dev",
+    rtx5090: "lmsysorg/sglang:dev",
+    "dgx-spark": "lmsysorg/sglang:dev",
+  },
+
+  // Pre-selects the issue template's `model` field on "Submit verified cell".
+  github: {
+    cookbookModel: "openbmb/MiniCPM5-2B",
+  },
+
+  playgroundFeatures: {
+    // The model fits one GPU on every supported card, so TP=1 is the verified
+    // shape; TP=2 is exposed for experimentation only.
+    attention: {
+      knobs: [
+        { id: "tp", label: "TP", values: [null, 1, 2] },
+      ],
+    },
+
+    // ----- Card: "Parsers" -----
+    // Opt-OUT: both flags are already in every cell, so the handler derives
+    // each chip as on and strips the flag when one is toggled off. The
+    // reasoning slug is `qwen3`, not `minicpm5` — see the header note.
+    parsers: {
+      items: [
+        { id: "reasoning", label: "Reasoning Parser", flag: "--reasoning-parser qwen3" },
+        { id: "toolCall",  label: "Tool Call Parser", flag: "--tool-call-parser minicpm5" },
+      ],
+    },
+
+    // ----- Card: "Speculative Decoding" -----
+    // Same DSpark flags as the Deploy panel's overlay row, so the two paths
+    // compose an identical command.
+    speculative: {
+      options: [
+        { id: "current", label: "Inherited from base" },
+        { id: "off",     label: "Off (greedy)" },
+        { id: "dspark",  label: "DSpark",
+          flags: ["--trust-remote-code",
+                  "--speculative-algorithm DSPARK",
+                  "--speculative-draft-model-path openbmb/MiniCPM5-2B-DSpark",
+                  "--speculative-dspark-block-size 7"] },
+      ],
+    },
+  },
+
+  // One recipe per card — the model card's SGLang launch line, plus the
+  // `minicpm5` tool-call parser it recommends for agent workloads and the
+  // `qwen3` reasoning parser its `<think>` template needs.
+  cells: [
+    {
+      match: { hw: "h200", variant: "default", quant: "bf16", nodes: "single" },
+      verified: true,
+      env: [],
+      flags: [
+        "--model-path {{MODEL_NAME}}",
+        "--reasoning-parser qwen3",
+        "--tool-call-parser minicpm5",
+        "--host {{HOST_IP}}",
+        "--port {{PORT}}",
+      ],
+    },
+    {
+      // Verification round still open on this card. `verificationStatus` alone,
+      // with NO `verified: true` baseline: the boolean is what the Playground
+      // reads for its own badge, so leaving it on would make the Playground
+      // claim "Verified" while the Deploy panel says the round is in progress.
+      match: { hw: "rtx6000", variant: "default", quant: "bf16", nodes: "single" },
+      // Flat string, not a predicate: this cell is in-progress with or without
+      // the DSPARK overlay, so there is nothing for the selection to switch on.
+      verificationStatus: "in-progress",
+      env: [],
+      flags: [
+        "--model-path {{MODEL_NAME}}",
+        "--reasoning-parser qwen3",
+        "--tool-call-parser minicpm5",
+        "--host {{HOST_IP}}",
+        "--port {{PORT}}",
+      ],
+    },
+    {
+      // The 32GB card is the one where the default KV pool starves decode
+      // CUDA-graph capture: with defaults the pool takes 473,718 tokens / 19 GB
+      // and leaves 4.6 GB, so capture stops around bs=48 and every larger batch
+      // runs eager (3810 tok/s at concurrency 64). The pair below gives back
+      // 4% of the pool -- still hugely oversized for a 2.5B model -- and keeps
+      // batches up to 128 graph-backed (7454 tok/s at the same concurrency).
+      // The two flags go together: raising the cap without freeing the memory
+      // just lets SGLang clamp capture back down.
+      match: { hw: "rtx5090", variant: "default", quant: "bf16", nodes: "single" },
+      verified: true,
+      env: [],
+      flags: [
+        "--model-path {{MODEL_NAME}}",
+        "--reasoning-parser qwen3",
+        "--tool-call-parser minicpm5",
+        "--mem-fraction-static 0.75",
+        "--cuda-graph-max-bs 128",
+        "--host {{HOST_IP}}",
+        "--port {{PORT}}",
+      ],
+    },
+    {
+      // GB10 has no discrete VRAM, so `mem_get_info()` reports all 128GB of
+      // unified system memory and the default fraction claims ~89GB for KV --
+      // leaving ~5GB for the OS, which kills the node during warmup with no
+      // traceback and no OOMKilled event. 0.30 is required, not tuning; it
+      // still leaves a 658k-token pool.
+      match: { hw: "dgx-spark", variant: "default", quant: "bf16", nodes: "single" },
+      verified: true,
+      env: [],
+      flags: [
+        "--model-path {{MODEL_NAME}}",
+        "--reasoning-parser qwen3",
+        "--tool-call-parser minicpm5",
+        "--mem-fraction-static 0.30",
+        "--cuda-graph-max-bs 128",
+        "--host {{HOST_IP}}",
+        "--port {{PORT}}",
+      ],
+    },
+  ],
+};