GLM-5.3-DSpark by Alaya NeW: Open-Weight Model

Runs On

What it takes to serve GLM-5.3-DSpark (2B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

Precision	Weights	Memory needed	Cheapest setup	Per hour	Also fits
16-bit	4.1 GB	4.9 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00
8-bit	2.0 GB	2.4 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00
4-bit	1.0 GB	1.2 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.

GLM-5.3-DSpark on every accelerator the SAVRN Index prices, at every precision

Model Card

By Alaya NeW, published under mit, revision 7b88f72ec567.

This repository releases a DSpark draft model for GLM-5.3. It is a speculator, not a standalone LLM: pair it with a GLM-5.3 target (BF16 or FP8) in SGLang or vLLM. Requires a recent SGLang or vLLM build with native DSpark support. This draft is for GLM-5.3 only. --speculative-num-draft-tokens is block size + 1. The draft is bf16 dense. Keep the draft KV cache dtype on auto so it does not inherit the target's FP8 MLA KV cache. - Context-length numbers are the mean of 3 runs (50 prompts, SPEED-Bench low-entropy throughput) - Qualitative numbers are one run on SPEED-Bench categories (80 prompts each) Acceptance length is the per-request mean of completion tokens divided by verification steps.…

Read Alaya NeW's full model card

This repository releases a DSpark draft model for GLM-5.3. It is a speculator, not a standalone LLM: pair it with a GLM-5.3 target (BF16 or FP8) in SGLang or vLLM.

Draft model: AlayaNeW/GLM-5.3-DSpark
Target model: zai-org/GLM-5.3 / zai-org/GLM-5.3-FP8
Algorithm: DSpark (parallel block draft + Markov head + confidence head)

Architecture

Field	Value
Architecture	`Qwen3DSparkModel` (`qwen3_dspark`)
Draft layers	5 (4× sliding attention + 1× full attention)
Hidden size / head dim	6144 / 128
Attention	GQA, 64 Q heads, 16 KV heads
Sliding window	2048
Block size $\gamma$	8
Target layers (KV inject)	`[1, 20, 38, 56, 75]` (GLM-5.3 has 78 layers)
Markov head	vanilla, rank 256
Confidence head	enabled (with Markov features)
Vocab	154880 (same as GLM-5.3)
Precision	bfloat16

Quick Start

Requires a recent SGLang or vLLM build with native DSpark support. This draft is for GLM-5.3 only.

Serve with SGLang

export SGLANG_ENABLE_SPEC_V2=1

sglang serve \
  --model-path zai-org/GLM-5.3-FP8 \
  --trust-remote-code \
  --tp-size 8 \
  --reasoning-parser glm45 \
  --tool-call-parser glm47 \
  --context-length 202752 \
  --mem-fraction-static 0.83 \
  --cuda-graph-max-bs 256 \
  --max-running-requests 128 \
  --disable-shared-experts-fusion \
  --host 0.0.0.0 \
  --port 8100 \
  --speculative-algorithm DSPARK \
  --speculative-draft-model-path AlayaNeW/GLM-5.3-DSpark \
  --speculative-num-draft-tokens 9 \
  --speculative-dspark-block-size 8

--speculative-num-draft-tokens is block size + 1.

Serve with vLLM

export VLLM_USE_V2_MODEL_RUNNER=1

vllm serve zai-org/GLM-5.3-FP8 \
  --trust-remote-code \
  --kv-cache-dtype fp8 \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --tool-call-parser glm47 \
  --enable-auto-tool-choice \
  --reasoning-parser glm45 \
  --enable-prefix-caching \
  --max-model-len 202752 \
  --max-num-seqs 256 \
  --gpu-memory-utilization 0.90 \
  --max-cudagraph-capture-size 256 \
  --host 0.0.0.0 \
  --port 8100 \
  --served-model-name GLM-5.3 \
  --speculative-config '{
    "method": "dspark",
    "model": "AlayaNeW/GLM-5.3-DSpark",
    "num_speculative_tokens": 8,
    "draft_sample_method": "probabilistic",
    "kv_cache_dtype": "auto"
  }'

The draft is bf16 dense. Keep the draft KV cache dtype on auto so it does not inherit the target's FP8 MLA KV cache.

Call the server

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8100/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="GLM-5.3",
    messages=[{"role": "user", "content": "Write a Python function that checks if a number is prime."}],
    max_tokens=1024,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(resp.choices[0].message.content)

Evaluation

Runtime: SGLang on 8× NVIDIA H200 (TP8)
Target: zai-org/GLM-5.3-FP8
Speculation block size: 8 (--speculative-dspark-block-size 8, --speculative-num-draft-tokens 9)
Sampling: temperature 0.7, concurrency 10, max generation 4096
Context-length numbers are the mean of 3 runs (50 prompts, SPEED-Bench low-entropy throughput)
Qualitative numbers are one run on SPEED-Bench categories (80 prompts each)

Acceptance length is the per-request mean of completion tokens divided by verification steps. Higher is better. Per-position rates are percent.

Speed-Bench: Low Entropy

Context Len	AL	Pos 0	Pos 1	Pos 2	Pos 3	Pos 4	Pos 5	Pos 6	Pos 7
1k	4.55	82.65	66.53	53.35	43.16	35.14	29.07	24.29	20.48
2k	4.76	84.08	68.73	55.92	45.99	38.31	32.29	27.45	23.47
8k	4.93	84.99	70.61	58.61	48.69	40.82	34.52	29.31	25.14
16k	4.98	85.48	71.22	59.17	49.60	41.76	35.42	30.19	25.75
32k	4.89	85.23	70.77	58.32	48.30	40.20	33.72	28.64	24.18

Speed-Bench: Qualitative

Category	AL	Pos 0	Pos 1	Pos 2	Pos 3	Pos 4	Pos 5	Pos 6	Pos 7
coding	4.60	84.01	67.95	54.73	44.02	35.76	29.25	24.20	20.22
stem	3.64	78.47	58.40	42.27	30.42	21.54	15.17	10.59	7.45
humanities	3.46	76.52	55.61	39.13	27.54	19.21	13.35	9.03	6.18
math	3.66	78.04	58.33	42.49	30.49	21.81	15.54	11.04	8.07
multilingual	4.77	83.99	68.72	56.02	46.37	38.54	32.49	27.46	23.18
qa	3.53	77.12	56.00	39.71	28.15	20.14	14.46	10.37	7.30
rag	4.71	84.06	68.93	56.67	46.01	38.02	31.06	25.52	21.01
roleplay	2.90	70.13	44.52	27.86	17.95	11.77	8.12	5.80	4.27
reasoning	3.90	78.72	59.27	44.31	33.53	25.73	20.00	15.95	12.73
summarization	4.95	86.97	73.33	60.79	50.35	41.10	33.40	27.50	22.14
writing	3.02	70.20	46.49	30.46	20.29	13.81	9.65	6.89	4.99

Citation

@misc{glm53dspark2026,
  title={GLM-5.3-DSpark},
  year={2026},
  howpublished={\url{https://huggingface.co/AlayaNeW/GLM-5.3-DSpark}}
}

Configuration

Architecture: Qwen3DSparkModel
Context length (tokens): 1,048,576
Layers: 5
Hidden size: 6,144
Feed-forward size: 12,288
Attention heads: 64
Key/value heads: 16
Head dimension: 128
Vocabulary size: 154,880
Sliding window (tokens): 2,048
RoPE base: 8,000,000
Model type: qwen3

Identity and Version

Repository: AlayaNeW/GLM-5.3-DSpark
Publisher: Alaya NeW
Task: Text generation
Modality: Text
Library: transformers
Parameters: 2B parameters
Languages: zh, en
Revision: 7b88f72ec56765e72475ca7a491b65943ab20969
First published: 2026-09-24
Last updated: 2026-09-24

Files and Weights

4 files, 4.1 GB in total. The weights are 1 file totalling 4.1 GB in safetensors.

Weights1 file · 4.1 GB

Configuration1 file · 1.7 KB

Documentation1 file · 5.8 KB

Repository1 file · 1.5 KB

Every file

File	Type	Size	SHA-256
model.safetensors	Weights	4.1 GB	43997ba601d5
config.json	Configuration	1.7 KB	—
README.md	Documentation	5.8 KB	—
.gitattributes	Repository	1.5 KB	—

License and Download

License: mit
Access: Open weights, no gate
Download size: 4.1 GB

Download from Alaya NeW

Released by Alaya NeW through its official repository on Hugging Face. Read the license.

Built From

Derived from zai-org/GLM-5.3
Described by arXiv:2607.05147

Memory Requirements

Precision	Weights in memory
As published	4.1 GB
16-bit	4.1 GB
8-bit	2.0 GB
4-bit	1.0 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About GLM-5.3-DSpark

How much GPU memory does GLM-5.3-DSpark need?

About 4.9 GB at 16-bit and 1.2 GB at 4-bit: the weights (2B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run GLM-5.3-DSpark on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use GLM-5.3-DSpark commercially?

Yes. GLM-5.3-DSpark is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

What is GLM-5.3-DSpark's context length?

1,048,576 tokens, from the maximum position embeddings in its published configuration.

Similar Models

Model · Text generation

Qwen3-1.7B

Qwen

Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support, with the following key features: - Uniquely support of seamless switching between thinking mode (for complex logical reasoning, math, and coding) and non-thinking mode (for efficient, general-purpose dialogue) within single model, ensuring optimal performance across various scenarios. - Significantly enhancement in its reasoning capabilities, surpassing previous QwQ (in thinking mode) and…

Open weights apache-2.0 2B parameters 40,960 tokens transformers

View model

Model · Text generation

TopologicalQwen

Convergent Intelligence

Topology-Aware Knowledge Distillation from Qwen3-30B-A3B → 1.7B TopologicalQwen is a 1.7B parameter model distilled from Qwen3-30B-A3B using Topological Knowledge Distillation (TKD) — a methodology that treats the teacher's output distribution over a concatenated token stream as a bounded variation (BV) function and decomposes knowledge transfer into three channels via the Mesh Fundamental Identity: 1. Smooth distillation (AC component) — Standard KL divergence over regions where the teacher's distribution varies continuously. This is what every other KD method does and stops at. 2. Jump corrections (D^j f) — Explicit correction terms at conceptual boundaries where the teacher's…

Open weights apache-2.0 2B parameters 40,960 tokens transformers

View model

Model · Text generation

Qwen3-1.7B-Thinking-Distil

Convergent Intelligence

Extended Reasoning Distillation from Qwen3-30B-A3B-Thinking → 1.7B The most downloaded model in the Convergent Intelligence portfolio. Qwen3-1.7B-Thinking-Distil captures extended deliberation patterns from the Qwen3-30B-A3B Thinking teacher — the variant that generates long-form reasoning chains before committing to an answer — and compresses them into a 1.7B student via supervised fine-tuning on the longwriter-6k dataset. The Thinking teacher produces the richest signal of the three teacher variants in the DistilQwen family (Instruct, Thinking, Coder). Where Instruct distillation captures clean instruction-following and Coder captures hierarchical decomposition, Thinking distillation…

Open weights apache-2.0 2B parameters 40,960 tokens transformers

View model

Model · Text generation

Qwen3-1.7B-Distilled-30B-A3B

Convergent Intelligence

A 1.7B-parameter causal language model distilled from Qwen3-30B-A3B on 6,122 STEM chain-of-thought samples using discrepancy-informed knowledge distillation. The training objective emphasizes proof structure, detects reasoning pivot tokens through token-level divergence dynamics, smooths high-entropy student singularities before distillation, and monitors structural drift through discrepancy energy. Standard knowledge distillation treats all tokens uniformly. Even proof-weighted approaches typically apply a static multiplier over the entire derivation span. That helps, but it still misses the internal structure of reasoning: some regions are smooth procedural continuation, while others are…

Open weights apache-2.0 2B parameters 40,960 tokens transformers

View model

Model · Text generation

Qwen3-1.7B-Coder-Distilled-SFT

Convergent Intelligence

A 1.7B model built in two stages: knowledge distillation from a 30B Coder teacher to establish a structured reasoning backbone, then supervised fine-tuning on ~54,600 logical inference problems. The Coder teacher's decomposition patterns meet formal propositional logic. The hypothesis: a model that learned STEM derivation from a Coder teacher (Stage 1) already has latent structure for sequential logic, state tracking, and compositional reasoning. Logical inference SFT (Stage 2) activates that structure explicitly — the model doesn't learn logic from scratch, it surfaces what the Coder teacher already gave it. Qwen3-1.7B distilled from Qwen3-Coder-30B-A3B-Instruct — the coding-specialized…

Open weights apache-2.0 2B parameters 40,960 tokens transformers

View model

Model · Text generation

DistilQwen3-1.7B-uncensored

Convergent Intelligence

This is the model card of a transformers model that has been pushed on the Hub. Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019). Part of the DistilQwen3 Series by Convergent Intelligence LLC: Research Division This model is part of a distillation chain built on Discrepancy Calculus — a measure-theoretic framework where the teacher's output distribution is decomposed via the Mesh Fundamental Identity into smooth (AC), jump, and Cantor components. The discrepancy operator…

Open weights 2B parameters 40,960 tokens transformers

View model