GLM-5.3-Flash

Vendor card: 320B total, 18B active, hybrid sparse plus linear attention, natively multimodal, first multimodal model in the GLM-5 series. Safetensors header on the same page says 321,323,031,390 params (BF16 6.93B, F8_E4M3 314.40B, F32 295,518). GEOMETRY filled 2026-10-17 from config.json, which the card does not restate. model_type glm5_next: hidden 4096, 45 layers = 34 linear_attention (KDA, linear_attn_config num_heads 64, head_dim 128, conv kernel 4) + 11 deepseek_sparse_attention at indices 3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43. MLA: kv_lora_rank 512, qk_rope_head_dim 0, qk_nope_head_dim 256, v_head_dim 256, q_lora_rank 1536, 64 heads and 64 kv heads so gqa_ratio 1, mla_use_nope true. head_dim is recorded as 0 because the glm5_next config class forces head_dim = qk_rope_head_dim, and this checkpoint sets that to 0; the content width lives in qk_nope_head_dim. KV (bf16, sparse layers only): transformers modeling_glm5_next.py caches the compressed latent, k_pass of width kv_lora_rank plus k_rot of width qk_rope_head_dim, so 11 layers x (512 + 0) x 2 bytes = 11,264 B/token = 11 KiB. The 34 KDA layers hold a fixed recurrent state (64 heads x 128 head_dim) and a conv state of kernel 4, neither of which grows with context, so they are not in that number. The DSA indexer (index_n_heads 32, index_head_dim 128, index_topk 2048, index_kpool 4, all 45 indexer_types full) keeps its own per-token state of width index_head_dim + hidden_size + 1 = 4,225 and is also excluded; it selects positions, it is not the KV cache. MoE: 288 routed + 1 shared, 8 routed per token, first 3 layers dense at intermediate 12288, remaining 42 sparse at moe_intermediate_size 2048, silu with swiglu_limit 10. Context 1,048,576 with no rope_scaling key. ENGINES: the card names SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth. The weights ship FP8 e4m3 dynamic. NVIDIA's GLM-5.3-Flash-FP4-Local (NVFP4, QAD) says the quant uses about 166 GiB of vRAM and is validated on 2x RTX PRO 6000 Blackwell and 2x DGX Spark 128 GB, so a single 128 GB box does not hold it and nothing stays memory-mapped on disk. The Q2 GGUF in the benchmark records is a local DwarfStar conversion, not a vendor quant. flops_per_token_active = 2 x 18e9 = 3.6e10 uses the vendor's 18B-active figure. MIT.