--- license: mit base_model: - zai-org/GLM-5.3-Flash pipeline_tag: text-generation tags: - slipstream - dgx-spark - nvfp4 - exl3 - glm --- # GLM-5.3-Flash-Slipstream GLM-5.3-Flash for the Slipstream inference engine: NVIDIA's NVFP4 checkpoint converted for Slipstream, with the routed and shared experts re-quantized by us to EXL3 4.0 bpw from the BF16 weights (Hessian-calibrated), plus the vision tower. These are **Slipstream weight files** (`weights.vbin` + `index.json` per directory), the engine's own pre-packed format. They run only with Slipstream; they are not safetensors and do not load in other runtimes. Built with Slipstream's converters at commit `20c2dd9` (2026-10-31). Base model: [zai-org/GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash). Engine: [Slipstream](https://github.com/Interactify/slipstream) (engine code release coming soon). Hardware: two NVIDIA DGX Spark (GB10, 128 GB each) connected over their 200 GbE link. ## Contents | Directory | What | Node | |---|---|---| | `glm53c_tp2_r0` | model weights, rank 0 (NVFP4 + MXFP8 prefill copies, MTP head) | first | | `glm53c_tp2_r1` | model weights, rank 1 | second | | `x3o_tp2_r0` | EXL3 4.0 bpw experts, rank 0 | first | | `x3o_tp2_r1` | EXL3 4.0 bpw experts, rank 1 | second | | `glm53_vision` | vision tower (BF16) | first | ## Download On the first Spark (rank 0, runs the HTTP server): ```bash hf download Interactify/GLM-5.3-Flash-Slipstream --local-dir ~/vmodels \ --include "glm53c_tp2_r0/*" "x3o_tp2_r0/*" "glm53_vision/*" ``` On the second Spark (rank 1): ```bash hf download Interactify/GLM-5.3-Flash-Slipstream --local-dir ~/vmodels \ --include "glm53c_tp2_r1/*" "x3o_tp2_r1/*" ``` The tokenizer, chat template and generation config: ```bash hf download nvidia/GLM-5.3-Flash-NVFP4 --exclude "*.safetensors" # on the first Spark ``` The directory names match the defaults in Slipstream's `server/models.json`, so no configuration is needed. ## Run ```bash python server/slipstream.py serve glm-5.3-flash --profile balanced ``` The server speaks the OpenAI and Anthropic APIs on port 8080. See the Slipstream README for profiles, the router and engine settings. ## Measured on Slipstream - Decode: 46-65 tok/s with MTP drafts (chat, code, JSON). - Prefill: 1120 (0.9k) / 1680 (3.5k) / 1750 (13.8k) tokens/s. - Experts: KL 0.053 against BF16 (the NVFP4 experts alone: 0.066). - Standard profile: perplexity +0.7 % against the checkpoint, KL 0.046. - tool-eval-bench: 91/100, median agent turn 3.4 s (92 scenarios). ## Notes - Context up to 1,048,576 tokens; Slipstream sizes the window to the memory of the placement. - An optional DFlash2 drafter is supported by Slipstream but not included here (non-commercial licence; download it yourself). ## License MIT, as the base model (Z.ai) and NVIDIA's NVFP4 checkpoint it is built from. See `LICENSE` and `NOTICE`.