zai-org/GLM-5.3-Flash-BF16

TL;DR: GLM-5.3-Flash is Zhipu's first natively multimodal GLM-5 model — a 320B-parameter MoE with only 18B active, released in BF16 under MIT. The interesting part is architectural: it's the GLM series' first hybrid of sparse and linear attention plus a "Manifold-Constrained Hyper-Connections" (mHC) scheme, and together with this week's Qwen3.8-Flash-Next it confirms that hybrid-attention sparse MoEs are now the consensus recipe for cheap long-context agentic serving. The card claims it beats GLM-5.2 at one-tenth the price and "approaches Claude Opus 4.8" on coding/agentic work — but ships zero numbers in the README itself.

What it is

An image-text-to-text model (English/Chinese) that is the first natively multimodal member of the GLM-5 family, meaning vision was in from pre-training rather than grafted on via a bolt-on encoder stage. Pre-trained on a 30T-token multimodal corpus. The "Flash" positioning is explicit: the goal is frontier-adjacent capability at a fraction of the serving cost, and the design choices all point that way — 5.6% activation ratio (18B/320B), and long-context attention cost cut by mixing linear attention layers in with sparse attention layers.

The architecture bet

Three claims stack up here:

1. Hybrid sparse + linear attention. Most layers presumably use some linear/state-space-style attention (O(n) in context length), with sparse full-attention layers interleaved to preserve precise long-range retrieval. This is the same bet as MiniMax, Qwen-Next, and Kimi's recent releases: pure linear attention degrades retrieval, pure attention is too expensive at 300K–1M contexts, and a small ratio of real attention layers recovers most of the quality. The benchmark footnotes suggest they take long context seriously — HLE-with-tools at 300K context, NL2Repo at 1M context, DeepSWE at 400K.

2. Manifold-Constrained Hyper-Connections (mHC). Hyper-connections (ByteDance, 2024) generalize residual streams by learning multiple weighted residual paths between layers; the "manifold-constrained" variant here is claimed to improve scaling efficiency. The README gives no detail — this is one to pull from the GLM-5 technical report (arXiv:2602.15763).

3. Extreme sparsity with native multimodality. 18B active on a 320B base means single-request compute comparable to a mid-size dense model.

What's actually there, and what isn't

There: BF16 safetensors under MIT (genuinely permissive), day-one support across SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth, and a controllable thinking budget via reasoning_effort (low/high/max, defaulting to max). The evaluation footnotes are unusually candid about setup — judge models, context management, even an anti-hacking measure for NL2Repo blocking unauthorized pip/curl in agentic runs, which is a nice acknowledgment of reward-hacking in coding evals.

Not there: a single benchmark number. The README references HLE-with-tools, Terminal-Bench 2.1, DeepSWE, Toolathlon, GDPval-AA, BabyVision — footnotes for a table that isn't in the card. All headline claims ("one-tenth the price of GLM-5.2," "approaching Claude Opus 4.8") are vendor-supplied and live in the blog/tech report. Treat them as marketing until independent evals (Artificial Analysis is cited for GDPval-AA) land. Also note that "Flash" is relative: BF16 weights for 320B are ~640GB, and even NVFP4 quantization needs ~180GB — this is a multi-GPU or beefy-workstation deployment, not a laptop model.

Should you care?

Yes, for two reasons. First, if the cost claims hold even approximately, this is the strongest openly licensed multimodal agentic model at this serving cost point. Second, as a signal: two major labs shipping hybrid sparse/linear-attention MoEs in the same week means the pure-transformer long-context era is ending at the frontier of open releases. The mHC detail is worth reading the tech report for — hyper-connection variants haven't yet appeared in a production frontier model at this scale.

Trying it

Easiest path is the hosted API on Z.ai. For local serving with vLLM:

pip install -U vllm
vllm serve zai-org/GLM-5.3-Flash-BF16 \
  --tensor-parallel-size 8 \
  --reasoning-parser glm5

For chat use, pass clear_thinking=true in the chat template (it defaults to false), and set reasoning_effort to low or high if you don't want max-budget thinking. For benchmark reproduction, keep the default max. The SGLang cookbook and Unsloth guide (for quantized/low-VRAM runs) are linked from the model card: https://huggingface.co/zai-org/GLM-5.3-Flash-BF16