SKILLEMALL.ai
GitHub ↗ RU Sign in

CB nemotron-lightning35

Reference desk for NVIDIA Nemotron 3.5 Lightning (30B-A3B), a hybrid Mamba-Transformer MoE with MTP built from Nano 3 Base by distillation from Super 3 Base, single-stage text SFT, RLVR, MTP boosting, and W4A16 NVFP4 quantization. Use when the user asks facts about Lightning 3.5 rather than building a pipeline.

The skillemall take

A reference desk for NVIDIA's Nemotron 3.5 Lightning—a 30B hybrid model combining Mamba, Transformer, and MoE. The skill promises facts about its architecture, distillation process, and W4A16 quantization. Twenty-one files, no scripts; 1593 body tokens. Checks found 10 medium and low-level issues, nothing critical. Safety score 90, quality 80, process 68. No sandbox or model runs. Use it as a lookup tool, not a pipeline builder. If you need specs on Lightning 3.5, it delivers. If you're deploying or integrating it, keep looking. Worth installing if your questions are about the model itself.

NVIDIA-NeMo/Nemotron Agent Skills author: NVIDIA-NeMo 21 files body ≈ 1 593 tokens Open the sourcegithub.com↗ analyzed 5 d ago

Reference desk for NVIDIA Nemotron 3.5 Lightning (30B-A3B), a hybrid Mamba-Transformer MoE with MTP built from Nano 3 Base by distillation from Super 3 Base…

As a process B 68/100 · Nearly there — weak spots: result and completion, inputs and preconditions, running it twice

ReferenceInfrastructuretype and topics are labelled automatically from the skill text
Runs in: Claude / Claude Code ChatGPT / Codex Gemini CLI Cursor GitHub Copilot / VS Code Qwen Code Kimi Code CLI Z.ai (GLM) DeepSeek (Deep Code) Mistral Vibe LangChain (deepagents) OpenClaw Goose OpenCode
JSON
Technical rating
C
86/100
safety, quality, tests
Safety 60%
90
Quality 40%
80
Run on models
none yet
Process rating
B
68/100
Nearly there
Inputs and preconditions w 11
0
Progress reporting w 2
0
Running it twice w 4
30
the three weakest of ten parameters · all ten

How to improve

    For the model run — optional
    • Your own cases (evals/evals.json, 4–6 real requests with expected answers): the full check would then run those instead of a model-drafted suite.
    • A spec.yaml with trigger phrases and assertions — a behaviour contract for CI; `skilltest init` writes a template.

    Guard findings · 10

    ✓ No critical or high findings

    Medium and low: 10
    • low Secrets in code secret-high-entropy-token paper/data.md:23
      High-entropy token-like string (may be an id, hash or a credential) (quoted — discussed, not commanded)
      - "Repo open-data blend (data_blend_raw.json): 23 subsets, all weight 1.0, from Nemo…-v1, Nemo…-v2, Nemo…2.1, Nemo…-v1; tokenized with the Lightning
      quoted
    • low Secrets in code secret-high-entropy-token paper/data.md:48
      High-entropy token-like string (may be an id, hash or a credential)
      Context: the report built a synthetic Indic corpus (Nemo…v1: about 2.83T generated tokens across 22 Indian languages and 34 language-script configurations, translated mostly from
    • low Secrets in code secret-high-entropy-token paper/mtp.md:22
      High-entropy token-like string (may be an id, hash or a credential) (quoted — discussed, not commanded)
      - "§3.7 boosting recipe (stated generally, then detailed for Super): hidden-state mixing plus a knowledge-distillation loss with the frozen backbone as teacher, the LK loss (Samarin et al., 2026) in p
      quoted
    • low Secrets in code secret-high-entropy-token paper/mtp.md:67
      High-entropy token-like string (may be an id, hash or a credential)
      Nemo…-v2 + Nemotron-RL-Super-Training-Blends seeds) and the "Training
    • low Secrets in code secret-high-entropy-token paper/quantization.md:144
      High-entropy token-like string (may be an id, hash or a credential)
      | QAD data | public blend of 13 tokenized Nemotron datasets (top weights: Pret…-v1 General 20, SFT-Math-v3 17, SFT-…-v2 python 15) |
    • low Secrets in code secret-high-entropy-token paper/sft.md:22
      High-entropy token-like string (may be an id, hash or a credential) (quoted — discussed, not commanded)
      - "Repo recipe open data: 12 blend entries from Nemo…-v1 (chat_if, structured_outputs), Nemo…-v1 (lean), Nemo…-v1, Nemo…-v1 (interactive_
      quoted
    • low Secrets in code secret-high-entropy-token paper/sft.md:79
      High-entropy token-like string (may be an id, hash or a credential)
      Open data blend in the repo config (12 entries): Nemo…-v1 (chat_if,
    • low Secrets in code secret-high-entropy-token paper/sft.md:81
      High-entropy token-like string (may be an id, hash or a credential)
      (interactive_agent, tool_calling), and Nemo…-v1 (C++, Python, InfiniByte parts
    • low Secrets in code secret-high-entropy-token recipes/stage0_pretrain.md:23
      High-entropy token-like string (may be an id, hash or a credential)
      | `config/data_prep/data_blend_raw.json` | open pretraining blend: Nemo…-v2, Nemo…2.1, Nemo…-v1, Nemo…-v1 (plus `_small` / `_large` variants) |
    • low Secrets in code secret-high-entropy-token recipes/stage1_sft.md:23
      High-entropy token-like string (may be an id, hash or a credential)
      | `config/data_prep/data_blend_raw.json` | open SFT blend: Nemo…-v1, Nemo…-v1, Nemo…-v1, Nemo…-v1, Nemo…-v1 |

    Files scanned: 21. Evidence is masked. Grey chips explain why severity was lowered.

    Against the Agent Skills spec

    ✓ No remarks against the Agent Skills spec

    Process rating: all ten parameters 68/100

    • 0Inputs and preconditions. Does not say what the process needs to start
    • 0Progress reporting. Says nothing while it works
    • 30Running it twice. 4 mutating operations with no state check
    • 40Result and completion. Does not say what the result is
    • 55Failures and branches. 1 branches
    • 70When it triggers. States when to use, but not when not to
    • 100Tools and files. No external tools needed
    • 100Steps. 34 steps
    • 100Consistency. Name and required fields are in place
    • 100Execution cost. Instruction body is 1593 tokens

    Everything here is measured from the skill text rather than judged by a model, so the numbers are checkable. A parameter weighs more when it is a more common reason for the process to stall.

    Quality signals

    • +5Description has no quoted example phrases that should trigger the skill
    • +4Description does not say when NOT to use the skill (false activations)
    • +3Output format is not stated: the model decides each time
    • +4No input/output examples
    • +1No license
    • +2Single-language instructions
    • +3Description length 312: enough signal without eating the budget
    • +4Structure: 9 headings
    • +3Step-by-step instructions: 34 items

    Quality base 70; lint remarks subtract, signals add up to 100. Result: 80.