GLM-5.3-Flash

320B total MoE, only 18B active per token - the first natively multimodal model in the GLM-5 series, released 2026-08-26 by Z.ai. The first open-source frontier model to pair sparse attention with linear attention in one hybrid architecture, cutting attention compute and KV cache by 3.01x and 4.44x vs GLM-5.3 while keeping precise long-context ability. Also adopts Manifold-Constrained Hyper-Connections (mHC). 1M context, MIT license on HuggingFace at zai-org/GLM-5.3-Flash, trained on a 30T-token multimodal corpus.

This is the ox-alpha reveal. Z.ai confirmed that the anonymous “ox-alpha” stealth model previewed free on OpenRouter/OpenCode from Aug 20 was an early version of GLM-5.3-Flash - and that the preview’s entire traffic load was served on Chinese AI chips. Per Z.ai (Zixuan Li), the official release is stronger and significantly more stable than the ox-alpha preview. OpenCode reported 42T tokens served in 6 days, making it the most-used model after DeepSeek Flash’s 56-day run. The ox-alpha row in this catalog is superseded by this model.

Native multimodal coding. Visual capabilities are built into the coding loop - the model observes interfaces, rendered results, and interaction feedback, then tests and improves its work. Coordinates across code, browsers, and GUIs (BUA/CUA) for frontend dev, game creation, and Blender 3D scenes. Beyond coding it handles Office, financial research, and document workflows, producing finished PPTX/PDF/DOCX/XLSX.

Benchmarks (Z.ai self-reported). GLM-5.3-Flash vs GLM-5.2 and the frontier field - strongest open-source multimodal-coder result across the board, beating GLM-5.2 on every coding and agentic row and leading open-source vision (OfficeQA Pro, CharXiv w/ tools, Chartography w/ tools):

Independent evaluation (Artificial Analysis, launch day). AA’s Intelligence Index run (max reasoning effort) confirms the launch story: 57 on the index, 3 points behind GLM-5.3 (60) and tied with GPT-5.6 Terra and Muse Spark 1.2, at a Cost per Task of $0.09 vs $0.68 for GLM-5.3 (~7.5x cheaper) - a spot on the intelligence-per-dollar Pareto frontier. The vendor’s agentic numbers replicate: AA measures GDPval-AA v2 Elo at 1770 (vendor claimed 1773), tied with GLM-5.3 and Grok 4.6 within the margin of error and behind only Claude Opus 5, and Terminal-Bench v2.1 at 84.3% - the exact vendor figure, and above AA’s own GLM-5.3 measurement (83.9%). It trails GLM-5.3 by 3.1 points on tau3-Banking (47.2%). The tradeoffs: AA-Omniscience accuracy of 28% (GLM-5.3: 34%, GPT-5.6 Terra: 47%), partially offset by a better hallucination rate (28% vs GLM-5.3’s 30%), and lower token efficiency than same-score peers - 149M output tokens to run the index, ~90% of them reasoning tokens, vs 133M for Kimi K3 and 136M for Qwen3.8 at the same 57. The cheap per-token price absorbs the extra tokens. One API caveat: AA lists a 400k context window for the launch endpoint, while Z.ai markets 1M native context.

Real-world deployment (Aug 26, 2026). A production team (Alec Fong, amplified by NVIDIA AI) reports switching their daily driver to GLM-5.3-Flash unquantized on 2x DGX Station with tensor parallelism 2: 881 tok/s at C=64, 232 tok/s single-stream, with realistic capacity of 4-8 simultaneous users at long context, and vision input working in production. That is a real multi-user serving data point for a frontier-class open model on a two-box desk cluster.

Access: model code glm-5.3-flash; OpenAI- and Anthropic-compatible APIs; GLM Coding Plan (3x GLM-5.3 quota, off-peak 50% points). First-party per-token API: $0.15/1M input, $0.50/1M output, cached input $0.026/1M (~80% discount) - just over 10% of GLM 5.3’s price. Recommended sampling: temperature 1, top_p 0.95, reasoning_effort max, thinking always on. Ollama Cloud carries the model at launch (private, US and Europe hosted, no data retention) with :cloud model codes wired into the usual harnesses - ollama launch claude --model glm-5.3-flash:cloud (Claude Code), ollama launch opencode --model glm-5.3-flash:cloud (OpenCode), or the Hermes Agent equivalent - while full GLM-5.3 rolls onto the same hosted pool; a plain API endpoint is available too.

Local-run status: runnable locally now. Unsloth published Dynamic GGUFs on Aug 28, 2026 (unsloth/GLM-5.3-Flash-GGUF, also one-click in Unsloth Desktop): 1-bit UD-IQ1_S (93GB) runs in ~100GB of RAM/VRAM, 2-bit UD-Q2_K_XL (109GB) needs ~115GB, 3-bit UD-IQ3_XXS (120GB) fits 128GB-class unified-memory devices - the M5 Max 128GB MacBook Pro and the NVIDIA DGX Spark - and 4-bit UD-Q4_K_XL (200GB) retains 93% of top-1 accuracy vs BF16 (642GB). Quant accuracy per Unsloth’s measurements: 71% (1-bit), 78% (2-bit), 82% (3-bit), 93% (4-bit). The variants are seeded, so the rig finder now matches this model against 128GB+ machines. The launch benchmarks are independently confirmed by Artificial Analysis (above), and a production team reports running it unquantized on 2x DGX Station at 232 tok/s single-stream (below).

Measured on the M5 Ultra Mac Studio (MacStories, Sep 21, 2026). In the same review that measured Flash-Next, the mixed 4/8-bit MLX build ran on the 256GB M5 Ultra vs the M3 Ultra 512GB: 41 tok/s vs 26 at low reasoning effort (+58%), prompt reads ~1,100 vs ~430 tok/s, first visible token on a 15.5K prompt 12.8s vs 32.6s. The long-context limits are the honest catch on a 256GB machine: at 128K a warm-cache re-run already ran out of memory, and at 256K neither Mac could answer a repeat request (the M3 Ultra 512GB handled the cold 258K prompt in 664s at 391 tok/s read). The mixed build’s reasoning tokens are included in its rates, so compare against thinking-mode numbers, not raw decode.