File size: 2,888 Bytes
b6d2234
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
---
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`.