GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Mr-ezdno
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Mr-ezdno
Model Tree
Blackfrost-AI/RED-SNOW-5.3-FLASH-BF16
Input Modalities
Output Modalities
Supported Functionality
Mixed-rate plan, average 2.46 bpw on the routed experts:
| Layers | Rate |
|---|---|
| 3–8 and 33–44 | K3 (18 layers — early and late MoE blocks, where quantization error compounds most) |
| 9–32 and the MTP layer 45 | K2 (25 layers) |
| attention + shared experts | K4 |
| dense MLP 0–2 | K3, lm_head 5, vision 6 |
Everything else follows the proven recipe of the 2bpw pack: codebook mcg (required for
sushi's fast GLM kernels — decode lanes, prefill grid, DFlash verify; a mul1 quant silently
falls back to slow paths), window 16, own exllamav3 1.6.2 conversion of the RED-SNOW BF16
weights with 250×2048 calibration, trunk rebuilt from BF16 over HTTP Range, pack skeleton from the
beamster/GLM-5.3-Flash-Sushi-2bpw template (MIT, Z.AI), the DFlash2 drafter shipped in that
template (Inco.ai, CC BY-NC-ND 4.0, see below),
chat_template.jinja and generation_config.json from RED-SNOW.
98 GB on disk. Verified: every repacked layer matches the converter's final output; round-trip repack against the author's template is byte-identical; pack-vs-BF16 audit clean.
KLD against the BF16 master, measured on the EXL3 conversions the packs are repacked from (exllamav3 model_diff, wikitext-2 test, 20×2048; the packs' experts are byte-identical, their affine trunk is rebuilt from BF16):
| this pack (2.46 bpw) | 2bpw pack (2.05 bpw) | BF16 | |
|---|---|---|---|
| KLD | 0.197 | 0.270 | — |
| top-1 agreement with BF16 | 85.6% | 82.2% | — |
| perplexity | 3.91 | 4.21 | 3.32 |
Not comparable with sushi kld readings.
sh
sushi serve --model RED-SNOW-5.3-FLASH-Sushi-2.5bpw --kv-quant 16 --ctx-size 300000# requires sushi >= 1.2.2; 128 GB tier (resident ~95 GiB)Reasoning model: give max_tokens headroom (3000+); if content returns empty with
finish_reason: stop, the answer is in reasoning_content. Tool calling works; vision and
the DFlash2 drafter are included.
GLM-5.3-Flash-DFlash2/ under its own CC BY-NC-ND 4.0 license (non-commercial, no derivatives; see its README).GLM-5.3-Flash-DFlash2/ (CC BY-NC-ND 4.0, above).sushi serve --model RED-SNOW-5.3-FLASH-Sushi-2.5bpw --kv-quant 16 --ctx-size 300000# requires sushi >= 1.2.2; 128 GB tier (resident ~95 GiB)