Cruz

🚨 𝗜 𝗝𝗨𝗦𝗧 𝗔𝗖𝗛𝗜𝗘𝗩𝗘𝗗 𝗦𝗢𝗠𝗘𝗧𝗛𝗜𝗡𝗚 𝗜𝗡𝗖𝗥𝗘𝗗𝗜𝗕𝗟𝗘 𝗢𝗡 𝗢𝗡𝗘 𝗗𝗚𝗫 𝗦𝗣𝗔𝗥𝗞! 🧪🔥
With the support of @AWERXAI, I'm excited to release something VERY special for the local AI community!

🧠 @deepseek_ai 𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸 𝗩𝟰.𝟭 𝗙𝗹𝗮𝘀𝗵 𝟭𝟬𝟭𝟳
📦 𝗦𝗔𝗚𝗘-𝗘𝗫𝗟𝟯 𝗮𝘁 𝗷𝘂𝘀𝘁 𝟭.𝟲𝟮 𝗯𝗽𝘄 𝗲𝘅𝗽𝗲𝗿𝘁𝘀
💻 𝟵𝟵.𝟲𝟭% 𝗖𝗢𝗗𝗘 𝗮𝗻𝘀𝘄𝗲𝗿-𝘁𝗼𝗸𝗲𝗻 𝗮𝗴𝗿𝗲𝗲𝗺𝗲𝗻𝘁!
⚡ 𝟯𝟱.𝟱 𝘁𝗼𝗸/𝘀 𝗱𝗲𝗰𝗼𝗱𝗲 𝗼𝗻 𝘃𝗟𝗟𝗠!
🚀 𝟭,𝟭𝟴𝟵 𝘁𝗼𝗸/𝘀 𝗠𝗔𝗫 𝗣𝗥𝗘𝗙𝗜𝗟𝗟!
🧠 𝟮𝟮𝟵𝗞 𝗖𝗧𝗫 (𝘀𝘁𝗶𝗹𝗹 𝘄𝗼𝗿𝗸𝗶𝗻𝗴 𝘁𝗼 𝗽𝘂𝘀𝗵 𝗶𝘁 𝗳𝘂𝗿𝘁𝗵𝗲𝗿!)
𝗔𝗟𝗟 𝗢𝗡 𝗢𝗡𝗘. 𝗗𝗚𝗫. 𝗦𝗣𝗔𝗥𝗞. 🤯

And YES, the quantized weights are AVAILABLE NOW! 💚
━━━━━━━━━━━━━━━━━━━━
🧪 𝗧𝗛𝗘 𝗟𝗔𝗦𝗧 𝟳𝟮 𝗛𝗢𝗨𝗥𝗦 𝗛𝗔𝗩𝗘 𝗕𝗘𝗘𝗡 𝗜𝗡𝗦𝗔𝗡𝗘!

Just a few days ago, I published my article about @nvidiaai's FP4-Local format: The Good, The Bad & The Ugly.

One thing really stuck with me...

NVIDIA demonstrated what was possible when aggressive quantization and intelligent optimization come together.

But there's a problem.

𝗥𝗘𝗣𝗥𝗢𝗗𝗨𝗖𝗜𝗡𝗚 𝗠𝗔𝗦𝗦𝗜𝗩𝗘 𝗤𝗨𝗔𝗡𝗧𝗜𝗭𝗔𝗧𝗜𝗢𝗡 𝗧𝗥𝗔𝗜𝗡𝗜𝗡𝗚 𝗣𝗜𝗣𝗘𝗟𝗜𝗡𝗘𝗦 𝗜𝗦 𝗢𝗨𝗧 𝗢𝗙 𝗥𝗘𝗔𝗖𝗛 𝗙𝗢𝗥 𝗠𝗢𝗦𝗧 𝗟𝗢𝗖𝗔𝗟 𝗔𝗜 𝗘𝗡𝗧𝗛𝗨𝗦𝗜𝗔𝗦𝗧𝗦.

That inspired me.

I've been developing my own SAGE-EXL3 process, and I wanted to see just how far I could push the latest DeepSeek Flash release while keeping the results accessible to the community.

So I spent the last 72 hours pushing the limits.

And the results genuinely surprised me! 🔥

━━━━━━━━━━━━━━━━━━━━

🏆 𝗧𝗛𝗘 𝗥𝗘𝗦𝗨𝗟𝗧𝗦 𝗔𝗥𝗘 𝗖𝗥𝗔𝗭𝗬!

I measured my 1.62 bpw expert quant against DeepSeek's native FP4 release checkpoint using held-out sequences and full-vocabulary token predictions.

And when I saw the CODE results, I could barely believe it!

🧪 𝗙𝗜𝗗𝗘𝗟𝗜𝗧𝗬 𝗩𝗦 𝗗𝗘𝗘𝗣𝗦𝗘𝗘𝗞'𝗦 𝗡𝗔𝗧𝗜𝗩𝗘 𝗙𝗣𝟴/𝗙𝗣𝟰 𝗥𝗘𝗟𝗘𝗔𝗦𝗘

💻 𝗖𝗢𝗗𝗘 — 𝗛𝘂𝗺𝗮𝗻𝗘𝘃𝗮𝗹

✅ 99.08% Top-1 Agreement
✅ 𝟵𝟵.𝟲𝟭% 𝗔𝗡𝗦𝗪𝗘𝗥-𝗢𝗡𝗟𝗬 𝗔𝗚𝗥𝗘𝗘𝗠𝗘𝗡𝗧!
✅ KL Divergence: 0.0186

🧠 𝗥𝗘𝗔𝗦𝗢𝗡𝗜𝗡𝗚 — 𝗠𝗔𝗧𝗛-𝟱𝟬𝟬

✅ 96.04% Top-1 Agreement
✅ 97.98% Answer-Only Agreement
✅ KL Divergence: 0.0349

🔢 𝗠𝗔𝗧𝗛 — 𝗚𝗦𝗠𝟴𝗞

✅ 95.56% Top-1 Agreement
✅ 94.19% Answer-Only Agreement
✅ KL Divergence: 0.0477
📊 𝗢𝗩𝗘𝗥𝗔𝗟𝗟 𝗛𝗘𝗟𝗗-𝗢𝗨𝗧 𝗙𝗜𝗗𝗘𝗟𝗜𝗧𝗬
🏆 94.52% Top-1 Agreement
📉 Mean KL: 0.0982
🎯 Median KL: 0.00019

𝗡𝗘𝗔𝗥-𝗟𝗢𝗦𝗦𝗟𝗘𝗦𝗦 𝗖𝗢𝗗𝗘 𝗧𝗢𝗞𝗘𝗡 𝗙𝗜𝗗𝗘𝗟𝗜𝗧𝗬 𝗔𝗧 𝗝𝗨𝗦𝗧 𝟭.𝟲𝟮 𝗕𝗣𝗪 𝗘𝗫𝗣𝗘𝗥𝗧𝗦! 🤯

These are direct teacher-forced fidelity measurements against native FP8, not task-level pass rates.
But seeing this level of agreement at such aggressive compression is EXACTLY what I was working toward!
━━━━━━━━━━━━━━━━━━━━

⚡ 𝗔𝗡𝗗 𝗬𝗘𝗦, 𝗜𝗧 𝗔𝗖𝗧𝗨𝗔𝗟𝗟𝗬 𝗥𝗨𝗡𝗦 𝗢𝗡 𝗢𝗡𝗘 𝗦𝗣𝗔𝗥𝗞!

Not two Sparks.
Not four Sparks.
Not an expensive rented GPU cluster.

𝗢𝗡𝗘 𝟭𝟮𝟴𝗚𝗕 𝗗𝗚𝗫 𝗦𝗣𝗔𝗥𝗞.

My latest vLLM measurements:
⚡ Decode: 35.5 tok/s
🚀 Maximum measured prefill: 1,189 tok/s
For a model this large, compressed this aggressively, and maintaining this level of fidelity...

𝗧𝗛𝗔𝗧'𝗦 𝗔 𝗕𝗜𝗚 𝗗𝗘𝗔𝗟 𝗙𝗢𝗥 𝗟𝗢𝗖𝗔𝗟 𝗔𝗜! 💚

And remember, 1.62 bpw refers to the routed experts, not every parameter in the full model.

━━━━━━━━━━━━━━━━━━━━

💚 𝗪𝗛𝗬 𝗧𝗛𝗜𝗦 𝗠𝗔𝗧𝗧𝗘𝗥𝗦.

I have a simple goal.

𝗠𝗔𝗞𝗘 𝗣𝗢𝗪𝗘𝗥𝗙𝗨𝗟 𝗔𝗜 𝗔𝗖𝗖𝗘𝗦𝗦𝗜𝗕𝗟𝗘 𝗧𝗢 𝗣𝗘𝗢𝗣𝗟𝗘 𝗪𝗛𝗢 𝗗𝗢𝗡'𝗧 𝗛𝗔𝗩𝗘 𝗔𝗖𝗖𝗘𝗦𝗦 𝗧𝗢 𝗗𝗔𝗧𝗔𝗖𝗘𝗡𝗧𝗘𝗥𝗦.

I don't want the future of local AI to be determined by who can rent the most GPUs or afford the largest training runs.

I want developers, independent researchers and enthusiasts to have the freedom to experiment, quantize, optimize and run powerful models on hardware they actually OWN.

NVIDIA's FP4-Local research showed me just how much potential remains in aggressive quantization.

My goal with SAGE-EXL3 is to keep pushing that frontier while making advanced quantization more practical for the wider community.

Today, I'm sharing the quantized checkpoint and the one-Spark serving recipe.

𝗧𝗛𝗜𝗦 𝗜𝗦 𝗝𝗨𝗦𝗧 𝗧𝗛𝗘 𝗕𝗘𝗚𝗜𝗡𝗡𝗜𝗡𝗚.

━━━━━━━━━━━━━━━━━━━━

🧪 𝗡𝗢𝗪 𝗙𝗢𝗥 𝗧𝗛𝗘 𝗥𝗘𝗔𝗟 𝗧𝗘𝗦𝗧!

The token-level fidelity results are incredible, but I'm not stopping there.

I'm currently running a broader evaluation campaign to measure how closely this quant preserves the native FP8 model's actual task performance.

💻 HumanEval+ / MBPP+
🧠 MMLU-Pro STEM / MATH-500
🔬 SciCode / LiveCodeBench
🔧 BFCL / Multi-step tool chains
🌍 MGSM

𝗧𝗛𝗘 𝗚𝗢𝗔𝗟 𝗜𝗦 𝗧𝗢 𝗣𝗥𝗢𝗩𝗘 𝗧𝗛𝗔𝗧 𝗔𝗚𝗚𝗥𝗘𝗦𝗦𝗜𝗩𝗘 𝗖𝗢𝗠𝗣𝗥𝗘𝗦𝗦𝗜𝗢𝗡 𝗗𝗢𝗘𝗦𝗡'𝗧 𝗛𝗔𝗩𝗘 𝗧𝗢 𝗠𝗘𝗔𝗡 𝗔𝗚𝗚𝗥𝗘𝗦𝗦𝗜𝗩𝗘 𝗤𝗨𝗔𝗟𝗜𝗧𝗬 𝗟𝗢𝗦𝗦!

Those task-level results are coming next! 👀

━━━━━━━━━━━━━━━━━━━━

👀 𝗔𝗡𝗗 𝗢𝗡𝗘 𝗟𝗜𝗧𝗧𝗟𝗘 𝗧𝗘𝗔𝗦𝗘𝗥...

@mr_r0b0t and I have also been working on an GB10 specialized inference engine called 𝗥𝗘𝗫𝗟𝟯.

Early testing has been VERY exciting, including 60 tok/s with bursts reaching 74tok/s!

We're putting the finishing touches on it and making sure everything is solid before releasing it.

More on that VERY soon. 😏

━━━━━━━━━━━━━━━━━━━━

🙏 𝗔 𝗛𝗨𝗚𝗘 𝗧𝗛𝗔𝗡𝗞 𝗬𝗢𝗨!

To @AWERXAI for supporting my work and helping make this contribution to the community possible.

To @nvidiaai for the FP4-Local research that inspired me to push this experiment even further.

To @justinn_builds for his initial vllm TP4 recipe that got me started in the right direction for loading this model up.

To @mr_r0b0t for collaborating with me on what's coming next!

And to @deepseek_ai for continuing to release powerful open-weight models that make community experiments like this possible.

𝗧𝗛𝗘 𝗕𝗘𝗦𝗧 𝗣𝗔𝗥𝗧 𝗢𝗙 𝗢𝗣𝗘𝗡 𝗔𝗜 𝗜𝗦 𝗪𝗛𝗔𝗧 𝗧𝗛𝗘 𝗖𝗢𝗠𝗠𝗨𝗡𝗜𝗧𝗬 𝗖𝗔𝗡 𝗕𝗨𝗜𝗟𝗗 𝗢𝗡 𝗧𝗢𝗣 𝗢𝗙 𝗜𝗧. 💚

🔗 𝗗𝗢𝗪𝗡𝗟𝗢𝗔𝗗 𝗠𝗬 𝗦𝗔𝗚𝗘-𝗘𝗫𝗟𝟯
huggingface.co/vcruz305/DSV4.…

🛠️ 𝗢𝗡𝗘-𝗦𝗣𝗔𝗥𝗞 𝗥𝗘𝗖𝗜𝗣𝗘
github.com/vcruz305/DSV4.…

𝗢𝗡𝗘 𝗦𝗣𝗔𝗥𝗞.
𝟭.𝟲𝟮 𝗕𝗣𝗪 𝗘𝗫𝗣𝗘𝗥𝗧𝗦.
𝟵𝟵.𝟲𝟭% 𝗖𝗢𝗗𝗘 𝗔𝗡𝗦𝗪𝗘𝗥 𝗔𝗚𝗥𝗘𝗘𝗠𝗘𝗡𝗧.

𝗟𝗘𝗧'𝗦 𝗞𝗘𝗘𝗣 𝗣𝗨𝗦𝗛𝗜𝗡𝗚 𝗟𝗢𝗖𝗔𝗟 𝗔𝗜 𝗙𝗢𝗥𝗪𝗔𝗥𝗗! 🚀🔥

🚨 NEW RELEASE ALERT! 𝗥𝗜𝗣𝗣𝗟𝗘-𝟲𝟱𝗕 𝗦𝗔𝗚𝗘-𝗥𝗘𝗫𝗟𝟯 - 𝗠𝗬 𝗙𝗜𝗥𝗦𝗧 𝟵𝟵% 𝗧𝗢𝗣-𝟭 𝗤𝗨𝗔𝗡𝗧! 🔥
🧠 𝟲𝟱𝗕 𝗗𝗘𝗡𝗦𝗘 • 𝟱.𝟰𝟭 𝗕𝗣𝗪
📦 𝟰𝟱.𝟮 𝗚𝗕 weights
⚡ Up to 𝟲𝟰.𝟴 𝗧𝗢𝗞/𝗦 - RTX PRO 6000 Blackwell
🎯 𝟵𝟵.𝟯𝟰% top-1 agreement vs BF16
🐈 𝟴𝟴.𝟴𝟵% sixcat-eval - just 𝟭.𝟵𝟭 𝗣𝗢𝗜𝗡𝗧𝗦 below my Qwen Flash Next FP4-Local result of 90.8%! 🤯

And I also built a SECOND version: just 𝟮𝟮.𝟯 𝗚𝗕 with 𝟵𝟲.𝟰𝟵% top-1 agreement!

Two different memory budgets. The same 65B dense Ripple model from @RicursiveAI.

━━━━━━━━━━━━━━━━━━

🔬 𝗧𝗛𝗘 𝟵𝟵% 𝗠𝗜𝗟𝗘𝗦𝗧𝗢𝗡𝗘

My 5.41 bpw release measured against Ricursive's original BF16 checkpoint:

🎯 99.34% top-1 agreement
📉 0.0005 mean KL divergence
✅ 99.998% top-5 agreement
📊 55,535 evaluated assistant-turn tokens

That's next-token prediction agreement, NOT a 99% task pass rate or literal losslessness.

After all the models I've quantized, crossing 99% overall top-1 agreement is a personal milestone!

━━━━━━━━━━━━━━━━━━

📦 𝗔𝗡𝗗 𝗬𝗘𝗦 — 𝗧𝗛𝗘𝗥𝗘’𝗦 𝗔 ~𝟮𝟮𝗚𝗕 𝗩𝗘𝗥𝗦𝗜𝗢𝗡!

Ripple-65B-Preview-SAGE-REXL3-2.57bpw

📦 22.3 GB weight pack
🎯 96.49% top-1 agreement vs BF16
📉 0.0155 mean KL divergence
💾 ~20.8 GiB resident weights
🐈 83.82% sixcat-eval (119/120 scored; provisional)

Roughly HALF the larger pack's size, while retaining over 96% next-token agreement on the same 55,535-token evaluation.

That gives the community another choice without pretending the smaller version has identical task performance. Runtime, draft and KV-cache memory still need room beyond the weights.

━━━━━━━━━━━━━━━━━━

🐈 𝗛𝗢𝗪 𝗗𝗢𝗘𝗦 𝗜𝗧 𝗗𝗢 𝗢𝗡 sixcat-eval?

My 5.41 bpw result:

🏆 88.89% overall — 118/120 scored
🧮 100% Math
💻 95% Coding
🛠️ 90% Tool Usage
📋 85% Instruction Following
🧠 83.3% Knowledge
🔬 80% Truthfulness

Just under TWO points below my earlier Qwen3.8 Flash Next FP4-Local score!

Different evaluation versions, sampling and serving configurations—not a controlled head-to-head. Both Ripple scores are provisional; unscored items are excluded and some responses truncated.

━━━━━━━━━━━━━━━━━━

⚡ 𝗗𝗘𝗡𝗦𝗘 𝟲𝟱𝗕. 𝗢𝗡𝗘 𝗚𝗣𝗨. 𝗥𝗘𝗔𝗟 𝗦𝗣𝗘𝗘𝗗.

5.41 bpw target + 4.55 bpw EXL3 DFlash2 draft, on ONE RTX PRO 6000 Blackwell 96GB, using SGLang:

🚀 Up to 64.8 tok/s greedy decode
🔥 Up to 53.1 tok/s sampled decode
📥 ~1,796 tok/s cold prefill
🧠 117,376-token full-attention KV pool

Decode peaks: 403-token prompt. Prefill: 3,152-token prompt. These speeds belong to the 5.41 bpw setup, not the 22GB version.

The quantized DFlash2 draft uses 2.19GB VRAM versus 7.60GB for BF16. The task evaluations used the BF16 draft; speed tests used the quantized draft.

I also built out the SGLang EXL3 integration, with fused kernels and Ripple-specific reasoning/tool-call support. Setup scripts, measurements and troubleshooting are in the recipe.

Credit to @RicursiveAI Ricursive Intelligence team (@annadgoldie @Azaliamirh @swyx ) for Ripple, and the @sgl_project contributors. 💚

🤗 𝟱.𝟰𝟭 𝗕𝗣𝗪 — 𝟵𝟵.𝟯𝟰% 𝗔𝗚𝗥𝗘𝗘𝗠𝗘𝗡𝗧
huggingface.co/vcruz305/Rippl…

🤗 𝟮.𝟱𝟳 𝗕𝗣𝗪 — 𝟮𝟮.𝟯 𝗚𝗕 / 𝟵𝟲.𝟰𝟵%
huggingface.co/vcruz305/Rippl…

🛠️ SGLang recipe + benchmarks
github.com/vcruz305/Rippl…

⚡ Quantized DFlash2 draft
huggingface.co/vcruz305/Rippl…

🐈 sixcat-eval — benchmark harness
github.com/vcruz305/sixca…

🚨 At last a USA based company who invested $32M in compute to train this model from scratch! They beat Qwen3.8-27B and 35
B-A3B and all the leading USA frontier models like Beam 501B and Meta Muse Spark Open 816B on multiple benchmarks! 🚀🔥🤯

Ripple-65B Preview (65B dense) is punching well above its size on agentic coding:
• TerminalBench 2.1: 80.5%
• SWE-Bench Pro v2.0-Hard: 79.1%
• DeepSWE v1.1: 44.3%
For context vs smaller models:
Qwen3.8-27B → 73.0% / 61.7% / 42.2%
Qwen3.6-35B-A3B → ~51.5% (TB 2.0) / 49.5%

As per their read me “a 65B parameter dense Transformer model pretrained on 12 trillion tokens with support for up to 1M context length. Ripple-65B Preview delivers strong results on agentic coding, scoring 80.52 on TerminalBench 2.1 and surpassing significantly-larger US open-weight models such as Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4).” 🤯🤯

Going to put this model through its paces and really validate these claims! Stay tuned!!

Kimi K3.1 Pro weights are finally released! 🚀 Let’s see what magic I can pull together for my 2x - 4x spark folk 😎

🚀 What an achievement of 40 tok/s decode and Prefill 1,020 tok/s for my latest Deepseek-4.1-Flash-1017 EXL3 1.55 bpw quant by the magician himself Mr @AgentSparko on just ONE DGX SPARK! ⚡️

I’ve updated my recipe repo with his work! I also recommend you take a look at this recipe that includes his super shiny LIMCARS dashboard! Amazing work! Thank you my brother!

github.com/vcruz305/DSV4.…