{"id":78326,"date":"2026-08-29T00:50:31","date_gmt":"2026-08-29T04:50:31","guid":{"rendered":"https:\/\/overcentral.com\/en\/?p=78326"},"modified":"2026-08-29T00:50:31","modified_gmt":"2026-08-29T04:50:31","slug":"z-ai-alibaba-identical-ai-models-78326","status":"publish","type":"post","link":"https:\/\/overcentral.com\/en\/z-ai-alibaba-identical-ai-models-78326\/","title":{"rendered":"Z.ai and Alibaba Release Nearly Identical AI Models"},"content":{"rendered":"

Two frontier open-weight AI models landed within 24 hours of each other this week, and their architectural blueprints read like photocopies. Z.ai released GLM-5.3-Flash<\/a>, a 320-billion-parameter multimodal mixture-of-experts (MoE) model with 18 billion active parameters. Alibaba\u2019s Qwen team followed with Qwen3.8-Flash-Next, a 125-billion-parameter model (6 billion active) that previews the Qwen4 architecture. The two teams worked independently, yet their configurations converge on nearly every major design decision: a 3:1 hybrid of linear and full attention, a compressed indexer capped at 2048 tokens, four gated residual streams instead of one, and the Muon optimizer with fused matrix splitting. This is not a coincidence\u2014it is a signal that the industry is rapidly consolidating around a shared recipe for efficient, long-context AI.<\/p>\n

GLM-5.3-Flash and Qwen3.8-Flash-Next: A Side-by-Side Look<\/h2>\n

GLM-5.3<\/a>a-Flash is the first natively multimodal model in Z.ai\u2019s GLM-5 series, released under the MIT license on Hugging Face. Tested anonymously as Ox Alpha<\/em> on OpenRouter, it became the most popular model of the week. Trained on a 30-trillion-token multimodal corpus, it serves a 1-million-token context window. Z.ai claims it outperforms GLM-5.2 across benchmarks at one-tenth the price, while approaching Claude Opus 4.8<\/a> on coding and agentic tasks. List pricing is $0.15 per million input tokens and $0.50 per million output tokens.<\/p>\n

Qwen3.8-Flash-Next plays the role Qwen3-Next played for Qwen3.5: an early public preview of the next architecture family. The model card lists a 125-billion main model plus an additional 51-billion n-gram embedding table, with 6 billion parameters activated per token. Native context is 262,144 tokens, extensible to 1 million via YaRN. The Qwen team reports that training required only about one-ninth the compute of Qwen3.7-Plus. The accompanying technical report is titled \u201cOn the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability.\u201d<\/em><\/p>\n

Convergence Point 1: Three of Every Four Attention Layers Are Linear<\/h2>\n

Both models use a 3:1 ratio of linear to full-attention layers. GLM-5.3-Flash stacks 45 layers: 34 linear-attention layers and 11 sparse-attention layers. Qwen3.8-Flash-Next stacks 48 layers in a repeating block of 3 Gated DeltaNet (GDN) layers plus 1 Qwen Sparse Attention (QSA) layer. The linear layers are the cheap layers. Instead of a KV cache that grows with the text, they compress all history into a fixed-size recurrent state. Compute per token stays constant regardless of context length.<\/p>\n

GLM uses Kimi Delta Attention (KDA), the linear-attention design introduced by Moonshot AI\u2019s Kimi Linear, which applies a fine-grained per-channel decay gate. Qwen uses its own Gated DeltaNet, which gates at the per-head level. Different granularity, same delta-rule family, same job: replace the growing cache with a compact, updateable summary. The remaining quarter of layers do precise long-range retrieval. GLM uses NoPE multi-head latent attention (MLA) in the DeepSeek style. Qwen uses grouped-query attention inside QSA. This is where the KV cache actually lives, and where the second shared trick comes in.<\/p>\n

Convergence Point 2: Compress 4x, Score, Keep 2048 Tokens<\/h2>\n

Neither model lets its full-attention layers attend over the entire context. Both attach a small learned indexer that scores chunks of history and keeps only the winners. GLM\u2019s sparse layers use a 32-head lightning indexer with top-2048 selection, descended from DeepSeek\u2019s DSA. To cut indexer cost at 1-million-token contexts, Z.ai introduces IndexPool, which compresses four indexer key vectors into one through weighted pooling before scoring. Qwen\u2019s QSA operates at micro-block granularity: the compressed lightweight indexer scores 4-token blocks and keeps the top 512 blocks\u2014exactly 2048 tokens.<\/p>\n

Both models compress the context 4x before scoring and cap the attention budget at 2048 tokens. Qwen credits QSA with up to 7.6x prefill and 4.9x decoding speedups over full attention at 1 million tokens. On GLM\u2019s side, the Flash architecture cuts attention compute by roughly 3x and KV cache size by 4.4x compared with the full GLM-5.3 model, while nearly halving active parameters (18B vs 32B) and layer count (45 vs 92).<\/p>\n

Convergence Point 3: Four Residual Streams Instead of One<\/h2>\n

Both models abandon the single residual stream that has defined transformers since 2017. Both widen it into four parallel branches, with gates controlling what each block reads back and writes out. GLM adopts Manifold-Constrained Hyper-Connections (mHC), a DeepSeek-originated design, configured with 4 branches in the shipped weights. Qwen wrote its own variant, Gated Residual, which modulates flow through 4 widened streams via an element-wise data-dependent read gate and a per-branch scalar write gate.<\/p>\n

According to the Qwen team, Gated Residual removes the extra branch-mixing step used by Hyper-Connections, reducing memory-access overhead, and the gate suppresses activation outliers well enough to allow FP8 residual storage. Notably, the Qwen team ablated both approaches and found them roughly equal in quality. Two labs, two implementations, one identical conclusion: four gated streams beat one.<\/p>\n

Convergence Point 4: Muon, With Fused Matrices Split Per Component<\/h2>\n

Both models train with the Muon optimizer. And both apply the same subtle refinement: fused projection matrices are split into their independent transformations before Muon orthogonalizes them. Qwen documents splitting fused QKV, SwiGLU, and GDN projections, assigning Muon to genuine 2-D linear maps and AdamW to embeddings, routers, and low-rank parameters. Qwen also refit its scaling laws for the new architecture and dropped batch-size warmup entirely, after measuring that warmup cost 18.8% more optimizer steps without improving results.<\/p>\n

Where They Disagree: Positional Encoding<\/h2>\n

The one clean split is rotary position embeddings in the full-attention layers. GLM-5.3-Flash drops them: the config sets qk_rope_head_dim = 0<\/code>codecodecode, making its sparse MLA layers fully NoPE. Position information flows implicitly through the recurrent linear layers. Qwen tried the same thing and kept RoPE. According to the Qwen3.8-Next technical report, NoPE produced no measurable difference during pre-training. The failure surfaced later: after post-training, the NoPE variant often failed to stop generating. That is a useful cautionary result for the field. Pre-training loss curves can hide behavioral defects that only appear after RLHF-stage tuning.<\/p>\n

The Broader Convergence, and the One Dissenter<\/h2>\n

This recipe is not limited to two labs. DeepSeek pioneered the sparse-indexer-plus-2048-budget pattern with DSA in DeepSeek-V3.2-Exp, and mHC is a DeepSeek design now shipping in GLM. Moonshot\u2019s Kimi contributed KDA, the exact linear-attention layer GLM adopted. Chinese open models are visibly cross-pollinating architecture components and converging on shared settings.<\/p>\n

The notable dissenter is MiniMax. During M2 development, the team extensively tested linear and sliding-window attention at scale and found severe deficits in multi-hop reasoning, especially beyond 32K context after SFT. M2 shipped with full softmax attention in every layer. For M3, MiniMax adopted MiniMax Sparse Attention (MSA), which sparsifies softmax attention via block selection but includes no linear-attention layers at all<\/strong>. So the field has not fully settled. Z.ai, Qwen, DeepSeek, and Kimi are betting that a 3:1 linear hybrid preserves reasoning. MiniMax\u2019s ablations say it does not, at least for their stack.<\/p>\n

The emergence of nearly identical architectures from two independent teams signals that the underlying optimization landscape is narrowing. The combination of linear attention for streaming cost, a tiny indexer for retrieval precision, and multiple residual branches for representation capacity appears to be a sweet spot for long-context models that need to be both cheap and capable. The one open question\u2014positional encoding\u2014may prove minor, as Qwen\u2019s post-training failure suggests that NoPE is workable but requires extra stabilization. Meanwhile, MiniMax\u2019s dissent reminds the field that linear attention still carries risks for complex reasoning chains. For practitioners evaluating open-weight models, the convergence means less guesswork: the recipe is being validated by multiple labs simultaneously, and the remaining debates are narrow enough to be tested in-house.<\/p>\n

    \n
  • GLM-5.3-Flash (34:11) and Qwen3.8-Flash-Next (36:12) independently landed on the same 3:1 linear-to-full attention ratio.<\/li>\n
  • Both compress context 4x and cap sparse attention at a 2048-token budget, a pattern DeepSeek\u2019s DSA started.<\/li>\n
  • Both replace the single residual stream with 4 gated branches; Qwen ablated its Gated Residual against mHC and found them equal.<\/li>\n
  • They split on positional encoding: GLM drops RoPE (NoPE), while Qwen kept it after NoPE models failed to stop generating post-training.<\/li>\n
  • MiniMax is the dissenter: its scaled ablations found linear attention hurts multi-hop reasoning, so M3 uses sparse softmax attention only.<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"

    Two frontier open-weight AI models landed within 24 hours of each other this week, and their architectural blueprints read like photocopies. Z.ai released GLM-5.3-Flash, a 320-billion-parameter multimodal mixture-of-experts (MoE) model with 18 billion active parameters. Alibaba\u2019s Qwen team followed with Qwen3.8-Flash-Next, a 125-billion-parameter model (6 billion active) that previews the Qwen4 architecture. The two teams […]<\/p>\n","protected":false},"author":7,"featured_media":78328,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/pub-4d4fc17555de4152be07eaf2a416a31e.r2.dev\/en\/ocie_1787979048378.jpg","fifu_image_alt":"Z.ai and Alibaba Release Nearly Identical AI Models","footnotes":""},"categories":[31],"tags":[],"class_list":["post-78326","post","type-post","status-publish","format-standard","has-post-thumbnail","category-technology"],"fifu_image_url":"https:\/\/pub-4d4fc17555de4152be07eaf2a416a31e.r2.dev\/en\/ocie_1787979048378.jpg","fifu_image_alt":"Z.ai and Alibaba Release Nearly Identical AI Models","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/78326","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=78326"}],"version-history":[{"count":0,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/78326\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/78328"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=78326"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=78326"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=78326"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}