UkisAI
Website  •  Learn more  •  BF16  •  Collection  •  Evaluation  •  Enterprise licensing

Swift 1.5 GLM-5.3-Flash

NVFP4. Quantized from Swift 1.5 GLM-5.3-Flash with NVIDIA Model Optimizer, following NVIDIA's published recipe: NVFP4 on the routed experts and dense MLP, with FP8 KV cache. Native NVFP4 execution requires compatible NVIDIA Blackwell hardware.

Swift 1.5 GLM-5.3-Flash is UkisAI's reasoning-efficient derivative of GLM-5.3-Flash. It uses 57.7% fewer thinking tokens, with a 2.4x speed up while keeping the accuracy loss <1% vs base on max reasoning effort.

The prompt is a sample from LiveCodeBench v6

Training approach

We made Swift GLM-5.3-Flash efficient by figuring out which tokens were linked to pathological overthinking and penalizing them without "attacking" the reasoning length directly, then regained the accuracy with further post-training, leading to "compressed" token usage while maintaining accuracy.

Swift 1.5 produces shorter reasoning traces. In our testing, we also observe fewer overthinking errors.

This release also features our previously mentioned post-training methods adapted specifically for coding and long-horizon agent work such as terminal use and software engineering.

Our training data is viewable here: https://huggingface.co/datasets/ukisai/Qwen3.8-27B-multi-turn-agent-sft albeit it is not used out of the box, but rather re-sampled, turned into proper RL environments etc.

Evaluation

The external results below compare GLM-5.3-Flash, the foundation base model, and Swift 1.5. Both models use the same saved evaluation protocols, and all scores are reported as final aggregate percentages.

The main table reports BF16 results. Quantized results are in Quantized models.

Benchmark Score Mean tokens Median tokens
BaseSwift 1.5 BaseSwift 1.5 ReductionReduction
General reasoning
GPQA-Diamond90.51%89.90%10,2486,052↓ 40.9%↓ 57.7%
MMLU-Pro87.38%86.25%2,7551,064↓ 61.4%↓ 44.4%
C-Eval93.42%93.68%1,214752↓ 38.0%↓ 19.9%
IFBench56.40%55.73%11,6608,299↓ 28.8%↓ 22.9%
Mathematics
AIME 202693.33%92.67%24,19318,145↓ 25.0%↓ 45.8%
HMMT (Nov 2025)95.33%91.33%25,92418,797↓ 27.5%↓ 31.7%
Coding
LiveCodeBench v686.69%87.94%21,50614,645↓ 31.9%↓ 36.1%
Agentic coding
Terminal-Bench 2.1*76.40%75.96%118,87879,473↓ 33.1%↓ 36.8%
DeepSWE v1.146.02%48.67%75,67159,513↓ 21.4%↓ 20.9%

* Note: Terminal-Bench 2.1 ran with a 5-hour limit (18,000 seconds) on every task, not the native per-task limits, for both base and Swift 1.5.

How to reproduce

Serving: BF16 · vLLM · GLM-4.7 reasoning parser · context 262,144 · reasoning effort max · MTP disabled.
Sampling: temperature 1.0 · top_p 0.95 · presence_penalty 0 · repetition_penalty 1.
Benchmarks: five seeds (0–4) for the question benchmarks. LiveCodeBench is full release v6 mean pass@1 over seeds.
Terminal-Bench 2.1: Harbor 0.20.0 / Terminus-2, pinned 89-task dataset, five attempts per task, temperature 1, top_p 1, 18,000-second task limit.
DeepSWE v1.1: 113 tasks × 3 attempts, Pier 0.3.1 with mini-swe-agent 2.4.6, GLM-4.7 tool-call parser.

BenchmarkOutput cap
GPQA-Diamond100,000
MMLU-Pro100,000
C-Eval16,384
IFBench81,920
AIME 2026250,000
HMMT Nov 2025250,000
LiveCodeBench v6100,000

Efficiency across reasoning efforts

GPQA-Diamond at each reasoning_effort setting, Swift 1.5 against the base at the same setting:

Reasoning effort Score Mean tokens Median tokens
BaseSwift 1.5 BaseSwift 1.5 ReductionReduction
Max90.51%89.90%10,2486,052↓ 40.9%↓ 57.7%
High88.79%86.97%7,6234,158↓ 45.5%↓ 31.1%
Low77.27%73.33%2,8151,377↓ 51.1%↓ 13.7%

At max, Swift 1.5 trails base by 0.61 percentage points while using 40.9% fewer mean and 57.7% fewer median thinking tokens. High and low save tokens but also lose 1.82 and 3.94 percentage points respectively.

Quantized models

Format Repository Runtime
NVFP4 Swift-1.5-GLM-5.3-Flash-NVFP4 NVIDIA Blackwell
AWQ INT4 (W4A16) Swift-1.5-GLM-5.3-Flash-W4A16-AWQ vLLM (compressed-tensors)
GGUF Swift-1.5-GLM-5.3-Flash-GGUF llama.cpp

Each quantized Swift 1.5 against a quantized base of the same format, on the same protocol as the BF16 evaluation:

Benchmark Score Mean tokens Median tokens
BaseSwift 1.5 BaseSwift 1.5 ReductionReduction
NVFP4
GPQA-Diamond90.10%90.20%13,0697,148↓ 45.3%↓ 59.5%
AIME 202694.00%91.33%30,85529,768↓ 3.5%↓ 48.3%
W4A16-AWQ
GPQA-Diamond90.81%89.19%10,6016,718↓ 36.6%↓ 55.3%
AIME 202692.67%94.00%20,13822,546↑ 12.0%↓ 46.8%

NVFP4 base: nvidia/GLM-5.3-Flash-NVFP4, the same recipe. W4A16-AWQ base: wtdcode/GLM-5.3-Flash-AWQ-W4A16, the closest public AWQ.

How to use

Native NVFP4 execution requires compatible NVIDIA Blackwell hardware and a runtime with GLM-5.3-Flash ModelOpt support. The quantized evaluation uses FP8 KV cache (--kv-cache-dtype fp8).

For installation and hardware-specific settings, see the vLLM GLM-5.3-Flash recipe and NVIDIA's GLM-5.3-Flash NVFP4 reference.

We evaluated this format on 8x B200 with a GLM-5.3-Flash vLLM build (vllm/vllm-openai:glm53-flash-x86_64-cu129), tensor parallelism 8, a 262,144-token context and MTP disabled. Load the checkpoint directly from this repository:

vllm serve ukisai/Swift-1.5-GLM-5.3-Flash-NVFP4 \
  --dtype bfloat16 \
  --tensor-parallel-size 8 \
  --kv-cache-dtype fp8 \
  --max-model-len 262144 \
  --reasoning-parser glm47 \
  --enable-auto-tool-choice \
  --tool-call-parser glm47 \
  --port 8000

Reasoning effort is set per request through the chat template, for example "chat_template_kwargs": {"reasoning_effort": "high"}. max is the default; high and low are also supported.

Quantization

Built with NVIDIA Model Optimizer (main @ 90ba9fb) using NVIDIA's published recipe for this architecture, GLM-5.3-Flash-BF16/ptq/nvfp4_experts_dense_mlp-kv_fp8_cast, the same recipe as nvidia/GLM-5.3-Flash-NVFP4:

  • NVFP4 (W4A4, group size 16): the routed experts (layers 3-44) and the dense MLP (layers 0-2).
  • FP8 KV cache (cast mode).
  • BF16: attention (KDA and sparse MLA), shared experts, router, vision tower, embeddings, lm_head and the MTP head.
  • Calibration: max, 1,024 samples of 512 tokens from cnn_dailymail and Nemotron-Post-Training-Dataset-v2 (512 each).

The one-layer MTP head (layer 45) is included unquantized in BF16, as in NVIDIA's release, so self-speculative decoding works:

--speculative-config '{"method":"mtp","num_speculative_tokens":5}'

License and access

Swift 1.5 GLM-5.3-Flash is a derivative of GLM-5.3-Flash (Copyright (c) 2026 Z.AI Co., Ltd, MIT License). UkisAI's contribution, including the adapted weights, is licensed under the Swift Open License v1.0. See NOTICE for the change notice and attribution details.

Personal, research, educational, evaluation, and commercial use of the Swift contribution are free for individuals and organizations with gross annual revenue, including affiliates, of up to US$1,000,000. Above that threshold, commercial use requires a separate Swift Enterprise License. Contact UkisAI for terms.

The base model's MIT License still applies to the base model. Nothing in the Swift Open License limits your rights in GLM-5.3-Flash itself under the MIT License.

Citation

@misc{swift-1.5-glm-5.3-flash,
  title  = {Swift 1.5 GLM-5.3-Flash},
  author = {UkisAI},
  year   = {2026},
  url    = {https://huggingface.co/ukisai/Swift1.5-GLM-5.3-Flash}
}

Acknowledgements

We acknowledge Amazon Web Services for providing compute credits and infrastructure support for Swift's development, training, and evaluation.

Downloads last month
299
Safetensors
Model size
161B params
Tensor type
F32
·
BF16
·
U8
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ukisai/Swift-1.5-GLM-5.3-Flash-NVFP4

Quantized
(5)
this model

Dataset used to train ukisai/Swift-1.5-GLM-5.3-Flash-NVFP4