🚀 FULL GLM-5.3. FOUR DGX SPARKS. A NEW PARETO-FRONTIER CONTENDER. This is a HUGE personal milestone for my SAGE work! 🔥 My GLM-5.3 SAGE EXL3 3.38 bpw brings the full model down to 319 GB while preserving ALL 19,200 routed experts. No pruning. No expert merging. Every bit matters. 🎯 INDEPENDENT-SET FIDELITY • KL divergence: 0.082388 nats • BF16 top-1 agreement: 91.1523% • 10,240 scored positions on text excluded from calibration ⚡ FOUR-SPARK PERFORMANCE Around 41 tok/s average (69 tok/s burst) across six real prompts with DFlash2, with drafted replies matching serial decoding token for token. 256K-context serving demonstrated. Sized toward a 1M-context memory target, with full-length inference validation still ahead. The goal is the highest-fidelity full GLM-5.3 possible on four Sparks WITHOUT sacrificing the memory needed for 256K–1M context. Not just making the weights fit, but leaving room to actually USE the model. 🙏 A MASSIVE thank you to @cfontes, who personally funded the compute required to make this model possible. This release would not have happened without his support! And we’re not done. He’ll be releasing a recipe soon that’s going to blow everyone’s minds! 🤯🔥 This is what community-backed local AI looks like. More capability, less memory, and people coming together to push what’s possible. 💪 Model 👇

RED-SNOW-5.3-FLASH EXL3 is LIVE for both ONE and TWO DGX Sparks. 🥵❄️ ONE Spark: 2.49 bpw, 90.32% top-1, up to 61.7 tok/s TWO Sparks: 4.91 bpw, 96.85% top-1, KL 0.01174 New collab with @Blackfrost_AI. RED-SNOW is an abliteration of GLM-5.3-Flash with a cyber/security fine-tune, built for authorized red-team operations. I took the BF16 release and built two SAGE MixedK EXL3 targets depending on how much hardware you have. 𝗢𝗡𝗘 𝗗𝗚𝗫 𝗦𝗣𝗔𝗥𝗞 2.49 bpw Top-1 agreement vs BF16: 90.32% Mean KLD: 0.1190 Single-Spark decode: ExLlamaV3: 38.7 tok/s TensorFold: 41.0 tok/s REXL3, co-collab stealth engine with @mr_r0b0t : 61.7 tok/s 👀🤫 That is the pack for people trying to squeeze the entire model onto one Spark. 𝗧𝗪𝗢 𝗗𝗚𝗫 𝗦𝗣𝗔𝗥𝗞𝗦 4.91 bpw Top-1 agreement: 96.85% Mean KLD: 0.01174 That is 9,917 / 10,240 positions choosing the same next token as the BF16 reference runtime. This is the quality-first build. SAGE does not just flatten every expert to the same precision. It allocates bits across the model based on sensitivity, which is exactly why I wanted separate 1-Spark and 2-Spark releases instead of forcing one compromise pack on everyone. Both are standalone EXL3 checkpoints. Vision components are packaged. MTP is preserved. Authorized research/security use only. Huge thanks to @Blackfrost_AI for trusting me with another collaboration. And yes, that 61.7 tok/s REXL3 result is something I’ll talk a lot more about soon. 🥶 Each repo has a proper deploy.md file for proper deployment to your Spark setup! Actively working on the 2 spark deploy.md so its validation stage but it will be updated shortly after validation! The one spark deploy.md is already validated! ONE Spark: TWO Sparks: RED-SNOW-5.3-FLASH COLLECTION:

🐳 DeepSeek-V4-Flash-Vision-Exp is 305B, and I packed it into a ~95 GiB mixed-precision EXL3 release that now SERVES on one DGX Spark. Receipts + Recipe below. 🚀 K2 across the routed-expert stack. K3 only where my layer sensitivity scan ranked the extra precision highest. MTP/DSpark drafter tensors preserved from source. 𝗠𝗜𝗫𝗘𝗗 𝗞, 𝗡𝗢𝗧 𝗙𝗟𝗔𝗧 𝟮-𝗕𝗜𝗧 The source checkpoint is already mixed format: MXFP4 routed experts FP8 attention MXFP4 MTP/DSpark drafter I dequantized each routed-expert tensor to BF16 before trellis encoding, then rebuilt the expert bank as: 43 MoE layers on a K2 base K3 on layers 3, 13, 21, 22, 28, 41 attention / shared experts / embeddings kept at high precision (BF16) 48 output shards, ~95 GiB total Those six layers came from a complete K2 vs K3 proxy-error scan, not guesses. The ranking was nearly flat across all 43 layers, so I am not pretending these are six magical outliers. The real question is whether the extra bits there produce a measurable fidelity gain over uniform K2 that comparison is still open. 𝗦𝗧𝗜𝗟𝗟 𝗢𝗣𝗘𝗡 prefill/decode numbers + max ctx PPL/KLD vs uniform K2 MTP/DSpark speculative serving CUDA-graph decode numbers sixcat battery (first full run is in; publishing after a chat-template fix that currently understates tool use) Model: Serve it (one Spark, open source):