
Tencent Hy4 vs GLM-5.5: One of These Models Does Not Exist
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 182 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 749 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 1902 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 149 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 96 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 83 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 109 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 211 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
GLM-5.5 does not exist. Z.ai's release log has not moved past GLM-5.3-Flash since 26 August 2026, the Hugging Face organisation's newest upload is the GLM-5.3 family from a day earlier, and the documentation path for a 5.5 model returns 404 — as does its Artificial Analysis page. Tencent Hy4, meanwhile, has been generally available since 4 November 2026 and open-weighted since 20 November, when Tencent published a 770-billion-parameter checkpoint under its own licence with prebuilt serving images. So the comparison this page can honestly run is Tencent Hy4 against GLM-5.3, the flagship Z.ai actually ships, plus an inventory of what would have to appear before the name in the title became a model.
That is not a technicality. GLM-5.3 is a serious open-weight model — 753 billion parameters with 40 billion active, a 1M-token context, a custom Z.ai licence and a full weight release — and it competes with Hy4 on almost the same axes. Pricing it against Tencent's new model is a real exercise. Pricing GLM-5.5 against anything is not, because there is no rate to quote.
Five surfaces, checked today
A claim that something does not exist decays the moment it is published, so here is the method rather than only the conclusion. Each line was run on 20 November 2026 and each is repeatable in an afternoon.
• Z.ai's release notes — the newest entry is GLM-5.3-Flash, dated 26 August 2026, above GLM-5.3 at 18 August, GLM-5.2 at 16 June, GLM-5.1 at 7 April and GLM-5 at 12 February. Nothing numbered 5.4 or 5.5 sits above it.
• Z.ai's documentation — the model guide path for a 5.5 model returns 404. The documentation index enumerates GLM-5.3-Flash, GLM-5.3 and GLM-5.2, and stops.
• The zai-org Hugging Face organisation — its four newest text repositories, all created 25 August 2026, are GLM-5.3, GLM-5.3-BF16, GLM-5.3-Flash and GLM-5.3-Flash-BF16. A repository named for 5.5 does not resolve.
• Artificial Analysis — a page exists for GLM-5.3 and returns a live index score. The path for a 5.5 model is a 404.
• Rate card — GLM-5.3 and GLM-5.2 sit at $1.40 input and $4.40 output, GLM-5.3-Flash at $0.15 and $0.50, GLM-5.3-FlashX at $0.37 and $1.25. There is no row above GLM-5.3.
None of that is a statement about obscurity, and it is worth saying why. Z.ai leaves a checkable trail for everything it ships: a documentation page, a rate row, a weight repository on the same day. GLM-5.3 has all three. A model with none of them has not shipped, whatever a version number implies about the cadence.
What GLM-5.3 actually is
Since the real comparison is against Z.ai's shipped flagship, its specification matters more than a rumour's.
• Architecture — 753 billion total parameters, 40 billion active per token, mixture-of-experts. The Hugging Face safetensors index confirms 753.3 billion, so this is a measured figure rather than a vendor round number.
• Context — 1M tokens, same as Tencent Hy4, which makes long-document work a straight comparison rather than a capability difference.
• Price — $1.40 per million input tokens, $4.40 per million output, with an 81% cache discount and a blended rate of $0.90 at a 7:2:1 cache-hit/input/output mix.
• Independent score — 32 on the Artificial Analysis Intelligence Index at maximum effort, 23 at low effort. Twenty-six point nine minutes per index task, 74 tokens per second, 75.5 seconds to first answer token.
• Licence — a custom Z.ai licence rather than a standard permissive one, with weights and an FP8 build published.
The interesting number is the index score. Tencent Hy4 scores 35, GLM-5.3 scores 32. Three points is inside the noise of a single evaluation run, which means these two models are, on independent evidence, roughly equivalent in aggregate capability. What separates them is everything else on the card.
Two open-weight flagships, side by side
This is the comparison GLM-5.5 was probably supposed to be, and it is close enough to be useful.
• Parameters — Tencent Hy4 is 770 billion total with 49 billion active; GLM-5.3 is 753 billion with 40 billion active. Nearly identical totals, nine billion fewer active parameters on the Z.ai side, which should make GLM-5.3 the cheaper model to serve at equal utilisation.
• Price — $0.834 and $2.501 for Hy4 against $1.40 and $4.40 for GLM-5.3. Tencent is 40% cheaper on input and 43% cheaper on output, despite activating more parameters per token. That is a deliberate pricing decision, not an architectural consequence.
• Cache — $0.042 for Hy4 against $0.26 for GLM-5.3, which is Z.ai's published 81% discount off $1.40. Hy4's cache economics are six times better, which matters most in agent loops that re-read a fixed context.
• Modality — Hy4 takes text, images and video frames. GLM-5.3's card and tags describe a text model, with the vision capability living in the Flash sibling. For an image-heavy pipeline that is a hard split rather than a preference.
• Output ceiling — 64K for Hy4, 128K for GLM-5.3 per its published limits. Z.ai wins that row.
• Speed and latency — 93 tokens per second and 115.9 seconds to first token for Hy4; 74 tokens per second and 75.5 seconds for GLM-5.3. Z.ai is the faster model on first-token latency and the slower one on throughput, which is an odd combination and probably reflects different reasoning-budget defaults rather than different engineering.
• Cost per completed index task — $1.03 for Hy4 against $5.01 for GLM-5.3. That is the widest gap in this comparison and it is not a token-price artefact: GLM-5.3 burns 135 million output tokens across the index against Hy4's 160 million at a blended rate of $0.90 per million against $0.45.
The licence difference nobody prices
Both models are downloadable, and the terms are not equivalent.
Tencent Hy4 ships under the Tencent Hy License 1.0 — Apache 2.0 with a Section 10 appended. That section requires a separate written licence for a model-as-a-service business with worldwide gross revenue above US$50 million in any twelve-month period, or a product with more than 100 million monthly active users. Internal use is carved out entirely: if the model and its outputs stay inside your organisation, Section 10 does not apply.
GLM-5.3 ships under a custom Z.ai licence, also not Apache, also with commercial conditions. The practical difference for most readers is nil — both permit research, fine-tuning, quantisation and internal deployment — but for a platform planning to resell inference, both require a conversation, and the thresholds differ.
The comparison that matters more is against Hy4's own predecessor. Tencent Hy4 Preview shipped in August under Apache 2.0 with no strings. The November GA build, and the 20 November weight upload, moved to the custom licence. Anyone who assumed "open weights" meant "same terms as August" should read Section 10 before building a product on it.
Why the version number keeps circulating
A name with no product behind it still gets typed into search boxes, and there are two mechanisms at work.
The first is cadence. Z.ai has shipped a numbered GLM release roughly every two months for a year — February, April, June, August — and the gaps are regular enough that a 5.5 in October or November is the naive extrapolation. That is a reasonable inference and a bad schedule. A release cadence is not a release date, and the four-month gap since GLM-5.3 is now longer than any in the sequence, which means the pattern has already broken rather than confirmed itself.
The second is that GLM-5.3-Flash exists and is easy to confuse with a generation bump. It is a different model at a different size — 320 billion total parameters with 18 billion active, vision-capable, priced at $0.15 and $0.50 — not a version of the flagship. Pages that treat "5.3-Flash" as evidence of an imminent 5.5 are reading a product line as a countdown.
Neither mechanism produces a model card. When one does, the trigger is specific: a documentation page that resolves, a rate row above $1.40, and a weight repository created in the zai-org organisation. Any one of those alone is a weak signal; all three together is a launch.
What to run while waiting
If the actual requirement is a GLM-class model today, GLM-5.3 is it, and the choice against Hy4 comes down to three questions.
• Does the input include images or video? Hy4, because GLM-5.3's flagship does not read them and the Flash sibling that does is a smaller model.
• Is the workload cache-heavy — a fixed repository, a fixed contract set, an agent that keeps notes in context? Hy4, on a 6x cache-price advantage and a 40% lower base rate.
• Does the output need to exceed 64K tokens in one call, or does first-token latency dominate the user experience? GLM-5.3, which publishes 128K output and gets to its first token in two-thirds the time.
• Is independent capability the only criterion? Neither. Three index points is not a decision, and both sit well below the closed frontier — Claude Opus 5.5 at 59 and Gemini 4 Argon at 50 are in a different bracket.
Both on one key

Tencent Hy4 and GLM-5.3 are both on OrcaRouter under their own strings — tencent/hy4 at $0.834 and $2.501, z-ai/glm-5.3 at $1.26 and $3.96 on our catalogue, below the $1.40 and $4.40 Z.ai publishes on its own card. Both are passed through at provider list price with zero markup on what we are billed; nothing is added.
The useful experiment is not which model is better in the abstract, because three index points will not settle it. It is which one completes your task more cheaply once retries are counted. A weighted route across the two, with a validation check before accepting an answer, measures that in a day — and because both are open-weight, the result is not locked to either vendor's roadmap.
When a GLM-5.5 does appear, it will be a model-string change on the same key rather than a second contract, and the routing table above it will not need rewriting.
The honest summary


Two claims survive scrutiny and one does not.
Tencent Hy4 is real, callable since 4 November, open since 20 November, and on independent evaluation sits three points above Z.ai's current flagship while costing 40% less per input token and carrying a six-fold better cache rate. It is the more capable of the two on text and the only one of the two that reads images or video.
GLM-5.3 is real, shipped since August, cheaper to run at equal utilisation by virtue of nine billion fewer active parameters, and better on two specific rows: output ceiling and first-token latency.
The claim that does not survive is any comparison involving GLM-5.5. There is no model card, no identifier, no rate and no weight repository behind the name, and the four-month gap since GLM-5.3 has now outlasted the cadence that made the rumour plausible in the first place. If you find a page pricing it, that page is describing a version number.
Compared in this article3
Detected from this article · Benchmarks: Artificial Analysis · updated daily
