Swift-1.5-GLM-5.3-Flash-NVFP4: Token-Efficient Reasoning on Blackwell GPUs

LinkUkisAITool

UkisAI has released an NVFP4 quantized version of GLM-5.3-Flash built with NVIDIA ModelOpt for Blackwell architectures. By penalising overthinking via post-training, the model cuts thinking tokens by 57.7% and speeds up inference with less than 1% accuracy loss. It caters to teams self-hosting agentic and coding workflows.

Key points

  1. Achieves a 57.7% reduction in median thinking tokens and a 2.4x speedup on maximum reasoning effort with <1% accuracy loss versus base.
  2. Quantised using NVIDIA Model Optimizer with NVFP4 (W4A4) on routed experts and dense MLP, alongside an FP8 KV cache.
  3. Post-training targets and penalises pathological overthinking tokens, tuned for terminal tasks and software engineering benchmarks.
  4. Evaluated on 8x NVIDIA B200 GPUs using vLLM and SGLang with a 262,144-token context window.
  5. Available under Swift Open License 1.0: free for entities earning up to US$1,000,000 annually, requiring an enterprise license beyond.

Why it matters

Reasoning models often burn significant compute through excessive deliberation on intermediate steps. Pairing token compression with native Blackwell NVFP4 quantisation substantially cuts inference costs and latency for long-horizon agentic workflows.

Analysis

Mitigating Overthinking in Reasoning Models

UkisAI's Swift 1.5 derivation of GLM-5.3-Flash (originally developed by Z.AI Co., Ltd) tackles inference latency by addressing token bloat. Rather than bluntly capping reasoning output lengths, the authors analysed and penalised tokens strongly correlated with repetitive or circular reasoning loops. Post-training and reinforcement learning environments then recovered task accuracy, producing significantly shorter reasoning traces without degrading benchmark outputs (for instance, 89.90% vs. 90.51% on GPQA-Diamond).

Quantisation Architecture and Runtime Details

The NVFP4 build adheres to NVIDIA's ModelOpt quantisation profile:

  • NVFP4 (W4A4, group size 16): Applied to dense MLPs (layers 0–2) and routed experts (layers 3–44).
  • FP8: Deployed for the KV cache to conserve memory.
  • BF16: Kept for attention mechanisms, routing components, vision modules, and the one-layer MTP head used for speculative decoding.

Executing NVFP4 natively requires NVIDIA Blackwell silicon (such as B200 GPUs) and compatible vLLM or SGLang container builds. For standard hardware, UkisAI provides AWQ INT4 and GGUF variants. Commercial deployments must also account for the revenue threshold stipulated by the Swift Open License v1.0.

Mentioned links

Found in

Summary generated by AI (Gemini 3.8 Flash, 9 November 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • X post:O

    Post-Trained GLM-5.3-Flash Variants for Coding and Security

    In an online discussion on GLM-5.3-Flash, user O highlights community post-trained derivatives hosted on Hugging Face. UkisAI's Swift 1.5 significantly cuts reasoning tokens for general daily use, while Cantina Security's apex-flash-1-abliterated targets cyber research. The author ranks both variants ahead of stock GLM-5.3-Flash and DeepSeek.

    13bookmarks468views

    AI & LLMs· Discussion

  • Link:Z.ai

    Z.ai Releases GLM-5.3-Flash as an Open-Weights Multimodal Model

    Z.ai has launched GLM-5.3-Flash under the MIT licence on Hugging Face, marking the first natively multimodal model in the GLM-5 family. Featuring 320B total parameters and 18B active parameters, it targets high efficiency across coding and agentic tasks.

    AI & LLMs· Tool

  • Repository:MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks

    Serving GLM-5.3 Flash EXL3 on Dual NVIDIA DGX Spark Kits

    Mia's AI Lab provides an optimised vLLM deployment overlay for GLM-5.3-Flash using 4 bpw EXL3 quantisation tailored for 2x NVIDIA GB10 systems.

    431starsPython

    AI & LLMs· Tool

  • Repository:NVIDIA/Model-Optimizer

    NVIDIA Model Optimizer: Unified Library for Model Compression and Inference Acceleration

    NVIDIA Model Optimizer (ModelOpt) provides state-of-the-art model optimization techniques such as quantization, pruning, distillation, and speculative decoding. It accepts PyTorch, Hugging Face, and ONNX models and outputs optimized checkpoints ready for downstream engines like TensorRT-LLM, vLLM, and SGLang.

    5kstarsPython

    AI & LLMs· Tool

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.