SummaryGLM-4.7-Flash is Z.ai's 30B-A3B mixture-of-experts text model under the MIT license, with a configured context of 202,752 tokens and tool-calling support in its serving examples. Refusal-reduced variants include huihui-ai's abliterated weights, a Heretic build that Venice serves, and a HauhauCS GGUF release with an undisclosed method.
What the base model is
GLM-4.7-Flash is Z.ai’s lightweight model, which its card calls “a 30B-A3B MoE model”: about 31 billion parameters in total, of which about 3 billion are active per token, chosen from 64 experts with 4 used at a time. It is text-only. Its card shows serving commands for vLLM and SGLang with a GLM tool-call parser and a reasoning parser, and recommends a “Preserved Thinking” mode for multi-turn agentic work.
Its attention uses multi-head latent attention, which stores a compressed KV cache. That makes long contexts cheaper in memory than a standard model of the same size, but it also means the usual per-head formula does not apply, so the VRAM estimator shows weights and overhead for this model and leaves the cache out rather than guess.
The configuration allows 202,752 tokens of context. The card does not state a limit; hosted offerings of the standard model in our price data list 200,000 or 202,752 tokens.
The variants at a glance
- huihui-ai abliterated. Made with the remove-refusals-with-transformers approach, per the card, which does not say which layers were changed. GGUF files by mradermacher, a 4-bit MLX build for Apple silicon, and an Ollama build are linked from the card.
- Olafangensan heretic. Made with Heretic 1.1.0. The creator reports 15/100 refusals against 50/100 for the original on Heretic’s prompt set, with a KL divergence of 0.0039, which is very low: the outputs barely moved from the original. Note the starting point: the original refused only half of those prompts.
- HauhauCS Balanced (GGUF). The card says “no changes to datasets or capabilities” and calls the release lossless, without describing the method or a test. It recommends Balanced over its Aggressive sibling for agentic coding and warns that the model does not work well in Ollama because of chat-template issues.
Which one for which need
- Hosted, with long context and a published method: Venice serves Olafangensan’s Heretic build at $0.07 input and $0.40 output per 1M tokens, with a 200,000-token context and a cached-input rate. Among the refusal-reduced rows in our price data, no model above 20B parameters has a lower rate.
- Local on a single GPU: a 4-bit GGUF of huihui-ai’s release or the HauhauCS file (its card lists 17 GB at Q4_K_M). Follow the runtime notes on each card.
- Through one API with other models: MolochAI lists
glm-4.7-flash-uncensored, without naming the weights.
Limits and open questions
The Heretic build’s very low KL divergence is encouraging, but its refusal count was measured against an original that already refused only 50 of 100 prompts, so the change is smaller than for the Gemma and gpt-oss builds in this catalogue. The HauhauCS repo has no base_model field in its metadata, although its card names GLM-4.7-Flash as the base; we record the base from the card.
MIT is the most permissive license among the bases here: keep the copyright and license notice if you redistribute. See how the licenses compare.
Ordered as catalogued. Method, layers and claims are as each creator states them; Atlas has not tested any of these variants.
Estimate for GLM-4.7-Flash with 32,768 tokens of context, one sequence, FP16 KV cache and 10% overhead. Variants of the same architecture need about the same.