Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    NVFP4 Large-MoE on RTX — Single-Box Field Playbook

    1

    No BF16 A/B multipliers we didn't run, no calibration stories we don't have.

    $9.99

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    NVFP4 Large-MoE on RTX — Single-Box Field Playbook

    NVFP4 Large-MoE on RTX — Single-Box Field Playbook

    Example session with this skill installed

    I have an RTX 5070 Ti 16 GB and a Ryzen 9 (has AVX-512) with 96 GB RAM. Can I run a ~120 GB MoE in NVFP4, and what decode speed should I expect?

    • Read your context and instructions
    • Compiled the nvfp4 large-moe on

    A grounded assessment: yes via experts-in-RAM + GPU FP4 path, AVX-512 CPU unlocks the fast kernel path (2× CPU-side gain per measured comparison), expected bands referenced against our 5060 Ti measurements (52–75 tok/s decode [measured] on 16 GB Blackwell), the vram-reserve floor to set first, concurrency=1 design point, and which of the 12 pitfalls apply to your exact combo.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Everything here comes from a real production single-node deployment (RTX 5060 Ti 16 GB + 96 GB DDR5, Blackwell sm_120) — not copied documentation. Every number is tagged by nature: [measured] / [official] / [config] / [estimated]. No BF16 A/B multipliers we didn't run, no calibration stories we don't have.

    What's inside

    • Architecture decision tree: NVFP4 on Blackwell (native FP4 Tensor Cores) vs Ada (capacity play + FP8 decode) vs Ampere (use INT4 honestly); why experts-in-RAM + GPU-FP4 is the only way a 100 GB-class MoE fits a 16 GB consumer card; why KV-cache precision is often a bigger lever than FP4 weights.
    • Measured performance bands: decode 52–75 tok/s healthy band with MTP accept 65–70%; prefill 1475–1556 tok/s, and a documented +16.3% cold-prompt prefill win from two specific env flags (needle 6/6); engine resident ~15.4 GiB of 16.3 GiB.
    • Tuned parameter sheet: context, KV dtype/residency, speculative-decoding depths, expert slots/cache, vram-reserve, pcie_frac, and why effective concurrency is 1 by design on 16 GB cards.
    • 12 field-tested failure modes (all [measured], all with fixes): the <512 MiB free-VRAM 7× decode collapse, the AVX-512 kernel trap that silently halves your CPU pool, warmup illusion, benchmark methodology that fakes 10× differences, RAM→swap death spiral with lying engine self-reporting, ETXTBSY binary swaps, /health false positives, updaters silently eating your tuned config, reasoning-model API quirks, the concurrency red line, and OS auto-update 'mystery restarts'.
    • Quality ladder grounding: NVFP4 92.1% top-token agreement vs Q8_0 on sm_120 plus the measured 3-rung quant comparison on the same box.

    Honest by design: we run a pre-quantized NVFP4-GPTQ pack and say so; no fabricated calibration war stories. Engine is the open-source Strata MoE stack (public GitHub) — no weights are redistributed with this product.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse All Claude Skills.

    Single-node consumer-hardware scope; measured bands are from one specific reference box (RTX 5060 Ti 16 GB + 96 GB DDR5) and other hardware will shift absolute numbers. We use a pre-quantized NVFP4-GPTQ pack — no first-party calibration-dataset guidance. No model weights are included or redistributed.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 16 days ago

    Permissions

    No special permissions declared or detected

    Listed16 days ago

    What's inside

    Frequently Asked Questions

    For the native FP4 Tensor Core fast path, yes — the decision tree in section 5 covers Ada (capacity play with FP8 decode) and Ampere (INT4 instead). The pitfalls and methodology apply to any single-box MoE deployment.

    No. Every number carries an evidence tag ([measured]/[official]/[config]/[estimated]) and the measurement conditions (idle, warm, coherent text). We deliberately refuse to quote BF16 speedup multipliers because we never ran that A/B on the same box.

    The playbook's home stack is the open-source Strata MoE inference engine (public GitHub repo with full docs). Architecture reasoning, parameter thinking and all 12 pitfalls transfer to any experts-in-RAM style stack.