Merge pull request #391 from NVIDIA-AI-IOT/feat/minmax

feat(benchmarks): add MiniMax M2.7 Model

Author: adsahu-nv

.agents/skills/jetson-model-catalog-inference/SKILL.md

@@ -280,6 +280,27 @@ Each value in the array must be another `models[].name` from `benchmarks.json`.
 
 ---
 
+## Discoverability artifacts (`/llms.txt`, `/llms-full.txt`)
+
+Two text routes surface the model catalog to external clients (search agents, LLMs, plain `curl` consumers). **Both are auto-generated from the content collection at build time — adding a model page is enough; no manual index edit required.**
+
+| Route | Source file | What it contains |
+|---|---|---|
+| `/llms.txt` | `src/pages/llms.txt.ts` | Short index. Curated tutorial header (static template literal) + auto-generated **Supported Models** section grouped by vendor. Models grouped by `family` per `MODEL_SECTIONS`, sorted by frontmatter `order`. |
+| `/llms-full.txt` | `src/pages/llms-full.txt.ts` | Full long-form index. Iterates `getCollection('tutorials')` and `getCollection('models')`, emits frontmatter + full Markdown body for every page. |
+
+**When you add a new model:**
+
+1. Set `family` to one of the existing families listed in `MODEL_SECTIONS` (in `src/pages/llms.txt.ts`). The model lands in the right section automatically.
+2. Set `short_description` clearly — it is the line text rendered in `/llms.txt`.
+3. Set `order` if you want a specific position within the family (lower = higher in the list).
+
+**When you introduce a brand-new vendor / family** (no existing entry matches): edit `MODEL_SECTIONS` in `src/pages/llms.txt.ts` to add a new section and list its `family` values. Without this edit the new model still appears, but in a fallback `### Other Models` section at the bottom.
+
+There is no static `public/llms.txt` — the dynamic route lives at `src/pages/llms.txt.ts`. Don't reintroduce a static file; Astro would let it shadow the route silently.
+
+---
+
 ## Validation (build will fail if wrong)
 
 `src/content/config.ts` `superRefine` on the models collection (skipped when `hide_run_button: true`):
@@ -305,6 +326,7 @@ Otherwise Zod reports a custom error on `supported_inference_engines`:
 - [ ] All matrix ids in `modules_supported` / `matrix_modules_disabled` / `one_shot_inference.modules_supported` are from `JETSON_MATRIX_MODULES`.
 - [ ] If `benchmark_key` is set, it exactly matches a `models[].name` in `src/data/benchmarks.json`. All product ids in `concurrency*` / `oom` / `dnr` arrays are valid benchmark product ids (not matrix ids).
 - [ ] OOM vs DNR vs null chosen per the [marker semantics](#oom-vs-dnr-vs-null-per-quant-markers) above.
+- [ ] `family` is set to a value listed in `MODEL_SECTIONS` (`src/pages/llms.txt.ts`). If the family is brand-new (new vendor), add a section to `MODEL_SECTIONS` so `/llms.txt` groups it correctly.
 - [ ] Run **`npm run build`** (see skill **`jetson-ai-lab-verify-build`**).
 
 ---
@@ -323,3 +345,5 @@ Otherwise Zod reports a custom error on `supported_inference_engines`:
 | One-shot snippet logic | `src/lib/evalRunModal.ts` |
 | Benchmark chart UI (OOM / DNR rendering) | `src/components/ModelBenchmarkSection.astro` |
 | Benchmark data file | `src/data/benchmarks.json` |
+| `/llms.txt` short index (auto-generated, vendor-grouped) | `src/pages/llms.txt.ts` (edit `MODEL_SECTIONS` for new vendors) |
+| `/llms-full.txt` long-form index (auto-generated) | `src/pages/llms-full.txt.ts` |

.agents/skills/skills.yaml

@@ -5,7 +5,7 @@
 skills:
   - name: jetson-model-catalog-inference
     path: .agents/skills/jetson-model-catalog-inference/SKILL.md
-    description: Configure Jetson AI Lab model pages — frontmatter for the Model Details sidebar (parameters, modalities, context length, license, precision pills), the Run on Jetson serve/call/one-shot UI, the three "block" levers (hide_run_button, matrix_modules_disabled, per-engine modules_supported), and the join to src/data/benchmarks.json (benchmark_key, benchmark_series, OOM vs DNR vs null). Use when adding or editing model pages, inference engines, serve/run commands, one-shot blocks, the details sidebar, benchmark wiring, or deciding whether to mark a Jetson module as OOM, blocked, or simply untested.
+    description: Configure Jetson AI Lab model pages — frontmatter for the Model Details sidebar (parameters, modalities, context length, license, precision pills), the Run on Jetson serve/call/one-shot UI, the three "block" levers (hide_run_button, matrix_modules_disabled, per-engine modules_supported), the join to src/data/benchmarks.json (benchmark_key, benchmark_series, OOM vs DNR vs null), and the auto-generated /llms.txt and /llms-full.txt discoverability indexes (src/pages/llms.txt.ts MODEL_SECTIONS — new vendors require a section entry). Use when adding or editing model pages, inference engines, serve/run commands, one-shot blocks, the details sidebar, benchmark wiring, deciding whether to mark a Jetson module as OOM, blocked, or simply untested, or introducing a brand-new model family.
 
   - name: jetson-ai-lab-verify-build
     path: .agents/skills/jetson-ai-lab-verify-build/SKILL.md

public/llms.txt

@@ -0,0 +1,75 @@
+---
+title: "MiniMax M2.7"
+model_id: "minimax-m2-7"
+short_description: "MiniMax's 230B agentic MoE flagship for software engineering and self-evolving agent harnesses with llama.cpp at 4-bit"
+family: "MiniMax M2.7"
+icon: "🌀"
+is_new: true
+order: 1
+type: "Text"
+vision_capable: false
+memory_requirements: "128GB unified memory (Thor T5000)"
+precision: "UD-IQ4_XS GGUF"
+parameters: "229B total / 10B activated"
+modalities: ["Text"]
+context_length: "196K"
+license: "MiniMax Model License"
+model_size: "100.96 GiB"
+hf_checkpoint: "unsloth/MiniMax-M2.7-GGUF"
+huggingface_url: "https://huggingface.co/unsloth/MiniMax-M2.7-GGUF"
+build_nvidia_url: "https://build.nvidia.com/minimaxai/minimax-m2.7"
+minimum_jetson: "Thor"
+# Gray out every non-T5000 tab globally — this 100.96 GiB model only fits on Thor T5000 (128GB unified memory).
+matrix_modules_disabled:
+  - thor_t4000
+  - orin_agx_64
+  - orin_nx_16
+  - orin_nano_8
+# Per-engine allowlist: only `thor_t5000` carries a serve command.
+serving:
+  entries:
+    - engine: "llama.cpp"
+      type: "Container"
+      modules_supported:
+        - thor_t5000
+      serve_command_thor: |-
+        sudo docker run -it --rm --pull always \
+          --runtime=nvidia --network host \
+          ghcr.io/nvidia-ai-iot/llama_cpp:latest-jetson-thor \
+          llama-server \
+            --hf-repo unsloth/MiniMax-M2.7-GGUF \
+            --hf-file UD-IQ4_XS/MiniMax-M2.7-UD-IQ4_XS-00001-of-00004.gguf \
+            --ctx-size 4096 \
+            --port 8080 \
+            --alias my_model \
+            --n-gpu-layers 999
+benchmark_key: "MiniMax M2.7"
+---
+
+[MiniMax M2.7](https://www.minimax.io/news/minimax-m27-en) is MiniMax's flagship agentic Mixture-of-Experts model, designed to build complex agent harnesses and complete highly elaborate productivity tasks. M2.7 is the first MiniMax model that deeply participates in its own evolution — during development the model autonomously updated its own memory, built dozens of complex skills for RL experiments, and improved its own learning process based on experiment results.
+
+This page describes serving the [Unsloth dynamic 4-bit GGUF](https://huggingface.co/unsloth/MiniMax-M2.7-GGUF) (`UD-IQ4_XS`, 100.96 GiB) on **Jetson AGX Thor T5000** with `llama.cpp`.
+
+## Inputs and Outputs
+
+**Input:** Text
+
+**Output:** Text (with optional reasoning traces between `<think>...</think>`)
+
+## Highlights
+
+- **229B total / 10B active** sparse MoE (`minimax-m2` arch), **196K context**.
+- **Strong real-world software engineering** and agentic tool use.
+- **Self-evolving training loop**: M2.7 helped optimize its own programming scaffold during RL.
+
+## Intended Use Cases
+
+- **Coding agents**: bug triage, refactors, code review, security analysis, and SRE-style root-cause investigations
+- **Long-running productivity agents**: document and spreadsheet automation with multi-turn tool use
+- **Agent harness research**: as a strong open-source backbone for tool-using and self-improving agent loops
+- **On-device RAG / repo Q&A** at the edge, when very large parameter counts matter more than minimum latency
+
+## Additional Resources
+
+- [Unsloth MiniMax-M2.7-GGUF on Hugging Face](https://huggingface.co/unsloth/MiniMax-M2.7-GGUF) — quantized weights (this page uses `UD-IQ4_XS`)
+- [MiniMaxAI/MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) — original BF16 weights and model card

src/content/models/minimax-m2-7.md

@@ -1177,6 +1177,41 @@
         }
       }
     },
+    {
+      "family": "MiniMax M2.7",
+      "name": "MiniMax M2.7",
+      "releaseDate": "2026-04-30",
+      "capabilities": {
+        "language": true,
+        "vision": false,
+        "audio": false,
+        "vla": false
+      },
+      "isl": 2048,
+      "osl": 128,
+      "engines": {
+        "llama.cpp": {
+          "UD-IQ4_XS GGUF": {
+            "concurrency1": {
+              "t5000": 17.05,
+              "t4000": null,
+              "agx_orin_64gb": null,
+              "orin_nx_16gb": null,
+              "orin_nano_8gb": null,
+              "dgx_spark": null
+            },
+            "concurrency8": {
+              "t5000": 34.17,
+              "t4000": null,
+              "agx_orin_64gb": null,
+              "orin_nx_16gb": null,
+              "orin_nano_8gb": null,
+              "dgx_spark": null
+            }
+          }
+        }
+      }
+    },
     {
       "family": "Pi",
       "name": "Pi 0.5",