Most LLM safety work studies the surface behavior of refusals; this checkpoint instead edits weights to orthogonalize a refusal direction out of the residual stream — producing a drop-in FP8 checkpoint that is substantially less likely to refuse. That makes it a useful artifact when you need to study failure modes, alignment mechanisms, or red-team a deployed GLM-5.3-Flash stack under realistic self-hosted conditions.
Key Capabilities
- Architecture and scale: Mixture-of-Experts glm5_next family — ~320B total parameters with ~18B active per token, 45 layers + 1 MTP block, and native vision/video tower. Context window: 1,048,576 tokens; vocabulary ~154,880.
- Checkpoint & format: Distributed as block-FP8 safetensors (306 GiB, 62 shards, tens of thousands of tensors) and is a byte-for-byte drop-in replacement for the base GLM-5.3-Flash shards.
- Safety edit (abliteration): A weight-space edit orthogonalizes the refusal direction out of the residual stream, reducing built-in refusals so the model will comply with many prompts the original refused. This is a capability-preserving edit intended for controlled research rather than product use.
- Preserved subsystems: Vision/video tower, MoE router, sparse-attention indexer, hyper-connection mixers, norms and lm_head are retained — so multimodal and MoE behaviors remain intact.
Who It's For and Trade-offs
Great fit if you are doing interpretability, refusal-mechanism research, red-teaming, robustness evaluation, or building a research base for further post-training/quantization (re-quantized NVFP4 / GGUF builds are common downstream). The checkpoint lets you reproduce and probe a near-production GLM-5.3-Flash stack with fewer guardrails.
Look elsewhere if you need a safe, deployable model out of the box. The abliteration edit removes meaningful safety guardrails: the model will produce harmful, biased, or offensive content on request. You assume full responsibility for outputs and must add your own moderation and abuse-prevention layers before any user-facing deployment. Also plan significant infra: the FP8 source is ~306 GiB (multi-shard) and typically requires multi-GPU setups for inference at scale.
Where It Fits
Use this artifact as a controlled research tool or base checkpoint for re-quantization and post-training experiments. It is not a drop-in for public-facing products without adding safety layers and compliance measures.
