GLM-5.3-Flash: 320B Params, 18B Active, 10× Cheaper Than GLM-5.3 — and It Was "Ox Alpha" All Along (2026)

What GLM-5.3-Flash is

Z.ai (Zhipu AI) released GLM-5.3-Flash on August 26, 2026, and the launch has three stories folded into one. First, the model itself: a 320B-parameter Mixture-of-Experts transformer that activates just 18B parameters per token, trained on a brand-new base — not post-trained on GLM-5.2’s, the way GLM-5.3 was. It is the first natively multimodal model in the GLM-5 series: text, image, and video input over a 1M-token context, with up to 128K tokens of output.

Second, the price story. GLM-5.3-Flash lists at $0.15 per million input tokens and $0.50 per million output — about a tenth of GLM-5.3’s $1.40/$4.40 — and a launch promotion halves even that through September 9, 2026. Third, the mystery: this is the model that had already been serving anonymous traffic as "Ox Alpha" on OpenRouter since August 20, becoming the week’s most popular model without a name attached. The reveal turned a curiosity into a product decision.

The honest way to read the launch: Z.ai said at GLM-5.3’s release that open weights were coming "in about two weeks" for the flagship. Instead, the open-weights event that actually arrived was a different, cheaper, multimodal model — a genuine MIT release, while GLM-5.3’s own weights still haven’t shipped.

Was it really "Ox Alpha"? Yes

On August 20, OpenRouter began serving a free anonymous model called Ox Alpha. Nobody knew who built it. It drew attention for its 1M-token context and its coding-agent traffic, and tokenizer forensics pointed at Z.ai. On August 26, Z.ai confirmed it: GLM-5.3-Flash is Ox Alpha, tested anonymously through a router to gather real-world feedback before the official launch. In Z.ai’s telling, the experiment worked — it was the most popular model of the week, with all of that traffic flowing through on real workloads.

That story matters beyond marketing: it means this model already absorbed a large amount of genuine agentic-coding traffic before the official launch — the kind of signal that makes a launch-day benchmark table slightly more believable than usual.

Architecture: the hybrid-attention bet

The technical headline of GLM-5.3-Flash is architectural, not just parametric. Z.ai calls it the first open-source frontier model to combine sparse attention with linear attention in a hybrid design, alongside a new scaling trick it calls Manifold-Constrained Hyper-Connections (mHC). The motivation is economic: attention compute and KV-cache memory are what make long-context serving expensive, and Flash attacks both.

Compared with GLM-5.3, Z.ai says the design cuts attention compute by 3.01× and KV-cache size by 4.44× while keeping long-context accuracy intact. Across a fair comparison set (GLM-5.3, DeepSeek-V4-Flash, Kimi K3), Flash posts the lowest attention compute per head per layer of any of them; its KV cache is still slightly larger than DeepSeek-V4-Flash’s and Kimi K3’s, so there’s headroom left.

The base model is new — trained on a 30T-token multimodal corpus — and Z.ai published its base-model evals against its own lineage and DeepSeek-V4-Flash-Base:

The standouts: GLM-5.3-Flash-Base posts the highest LiveCodeBench-Base score of the set (37.6) — ahead of GLM-5-Base’s 34.4 — and lands within noise of it on MMLU, BBH, and HellaSwag. SimpleQA trails GLM-5-Base (33.5 vs 36.0), but the overall shape is what Z.ai claims: more intelligence per unit of compute than the previous generation.

Benchmarks: beats GLM-5.2, loses honestly to the frontier

Every figure in this section comes from Z.ai’s own launch table (or the Hugging Face model card, which matches it), with one exception — GDPval-AA v2, which Z.ai footnotes as run by Artificial Analysis. There is no independent reproduction of the full table yet, so treat the whole thing as a maker-set ceiling.

The clean result: vs GLM-5.2

This is the specific claim Z.ai makes, and the chart supports it: a win on all eight published benchmarks at less than half the size and roughly a tenth of the price. The biggest deltas are on the long-horizon agentic rows — AutomationBench nearly doubles, DeepSWE jumps by 17 points, Toolathlon by 18.5.

The honest result: vs the closed frontier

The launch-day summaries emphasized the wins, so the losses deserve equal airtime. On Terminal-Bench 2.1, Flash sits last-but-one (84.3) behind GPT-5.6 Terra (87.4), Gemini 3.7 Flash (85.8), and Claude Opus 4.8 (85.0). On Agents’ Last Exam it’s last of the models plotted (26.3). HLE with tools trails Opus 4.8 (55.3 vs 57.9). Gemini 3.7 Flash beats it on AutomationBench (52.3 vs 48.8), its best row against Anthropic and OpenAI. The genuine wins: GDPval-AA v2 (1,773 Elo — highest on the chart) and AutomationBench over Opus 4.8 and Terra.

What independent measurement says

The one third-party check available at launch is the Artificial Analysis Intelligence Index v4.1.1, where GLM-5.3-Flash scores 57 at about $0.045 per completed task at the discounted rate — pushing the Pareto frontier of the index. That puts it exactly level with Claude Opus 4.8 (57) and one point behind Qwen3.8 Max, at a tiny fraction of either’s price. On Z.ai’s own in-house eval — Z.ai Code Bench v1.0, run in Claude Code 2.1.207 — Flash at max effort nearly matches Opus 4.8: 29.0 vs 29.5.

Vision: the first multimodal GLM-5

GLM-5.3-Flash is Z.ai’s first attempt at native multimodality in the GLM-5 line, and the picture is genuinely mixed. On documents, charts, and technical reasoning it looks strong — OfficeQA Pro 62.4 (vs DeepSeek-V4-Vision-Exp’s 57.9 and Opus 4.8’s 48.9), CharXiv Reasoning with tools 89.4 (level with Opus 4.8’s 89.9), Chartography with tools 78.0 (clear of the field).

The weakness is exactly where Z.ai’s flash-tier competitor lives: Gemini 3.7 Flash beats it on BabyVision (70.9 vs 53.4), MVbench (82.2 vs 77.8), and MMVU (82.3 vs 80.5) — the multi-image and video-heavy evals. GPT-5.6 Terra also wins BabyVision. The pattern reads like a first generation: excellent on single-image document and chart work, behind on video and multi-image understanding. If the job is screenshots and PDFs, this is a value-tier option; if video reasoning matters, keep looking.

Pricing: the whole point

GLM Coding Plan: 3× the quota

Z.ai rolled GLM-5.3-Flash out to all GLM Coding Plan subscribers at launch with 3× the usable quota of GLM-5.3. The Coding Plan’s points-based system carries over, and off-peak calls — including all day on weekends — consume 50% of standard points. The plan now reads as: GLM-5.3 for the hardest agentic jobs, GLM-5.3-Flash for everything else, three times more of it.

Open weights: MIT on day one

This is where GLM-5.3-Flash parts company with its namesake flagship. The weights shipped on Hugging Face (zai-org/GLM-5.3-Flash) under an MIT licence on day one — the open-weights release everyone expected from GLM-5.3, delivered instead by the smaller sibling.

Self-hosting is real but not cheap: the FP8 checkpoint is roughly 306 GiB, and Z.ai’s guidance is Hopper-or-newer GPUs on an 8-GPU node minimum. At launch, SGLang, vLLM, and TokenSpeed are supported, with more inference frameworks promised. That’s a serious deployment requirement — a single H100/H200-class node — but for anyone already running an open-weights fleet (the DeepSeek crowd, for example), it slots into the same tooling: DeepSeek V4 Models, Harness, and API Discount Windows covers the open-weights playbook these models share.

The GLM-5.3 elephant in the room

GLM-5.3’s own open weights remain unreleased. At its August 14 launch, Z.ai said the ~750B flagship’s weights would follow "in about two weeks" — after safety evaluation of its emergent vulnerability-exploitation capability — and the Flash launch arrived on day 12 of that window without them. As of August 26, the zai-org Hugging Face organisation carries GLM-5.3-Flash’s weights, not GLM-5.3’s.

The asymmetry is worth stating plainly: the model Z.ai was willing to open-license immediately is the 320B multimodal value model; the one it is holding back is the text-only cyber-capable flagship. That ordering is consistent with the stated reason — the flagship’s vulnerability-exploitation gains were what triggered the delay — and Flash’s launch chart includes none of the cyber benchmarks GLM-5.3’s did. Whether that’s reassuring or not depends on how much you trust the safety framing. The full context is in the flagship’s guide: GLM-5.3: The Same Base Model, Six-Times-the-Agent, and the Weights Z.ai Held Back.

Who should use GLM-5.3-Flash

• Agentic-coding workloads on a budget. If you’re paying frontier prices for automation and coding-agent tokens measured in millions, this is the most interesting release of the month — near-Opus-4.8 coding performance at 1/50th of the output price.
• Document-heavy multimodal work. Screenshots, PDFs, charts, and screens: the strong rows (OfficeQA Pro, CharXiv, Chartography) make it a value-tier vision option — as long as the job isn’t video reasoning.
• Self-hosters. MIT weights, SGLang/vLLM/TokenSpeed support, and the cheapest frontier-adjacent per-token economics of the current open-weights crop. Bring an 8-GPU Hopper node.
• GLM Coding Plan subscribers. 3× quota for the same plan price, with off-peak discounts on top — the upgrade is already live.
• Not for: jobs that need the flagship’s cyber capability, anyone who needs GLM-5.3’s specific reasoning profile (no head-to-head is published), or video-heavy vision pipelines.

Frequently asked questions

Sources & further reading