GLM-5.3 Flash / Ox Alpha
GLM-5.3 Flash / Ox Alpha
Vintage: 2026-08. Primary evidence is one lab-adjacent X post by @jietang dated 2026-08-27. Not a US frontier-lab official account. Photo attachment. Numbers in this post only: "AA = 57", "1/100 frontier price", "nearly 20% weekly token share (no. 1) on OpenRouter", "Powered by pure Chinese chips." Do not invent parameter counts from that post. 2026-09-04: unsloth issuer docs describe 320B / 18B / 1M — Unsloth-described, not a Z.ai model card. Overnight omen-alpha is attached as unconfirmed stealth only — do not rename Omen to this SKU.
One-line summary: @jietang wrote that Ox Alpha is GLM-5.3 Flash (AA = 57, 1/100 frontier price, OpenRouter weekly #1 claim). Sep 4: Unsloth claims 3.3× faster local GGUF. Omen Alpha is a separate OpenCode stealth name, identity unconfirmed.
What it is
A named GLM Flash-tier model equated in the post to "Ox Alpha." Treat this as a lab-adjacent X claim (Z.ai / GLM-associated account per the clipping qualify note), not an official @GoogleDeepMind / @OpenAI / @AnthropicAI post and not a fetched model card.
Why it matters to this thread
Open-weight / Chinese-lab model economics are in-scope (chinese-open-weight-frontier-parity, open-vs-closed-source-model-economics). This is the first dated X claim in the thread that a GLM-5.3 Flash SKU is taking OpenRouter weekly token share. It does not rewrite the June GLM 5.2 / July Kimi K3 podcast evidence on those pages.
Key facts (from 2026-08-27-x-ai-news-27-aug-2026-deepmind-double-blind-evals)
The bullets are one X post (fetch_method: x-mcp). Permalink lives on the source page.
- Identity (@jietang, 2026-08-27): "Ox Alpha = GLM-5.3 Flash"
- Numbers in this post only: "AA = 57 ," / "1/100 frontier price," / "Delivered nearly 20% weekly token share (no. 1) on OpenRouter."
- Silicon claim in this post: "Powered by pure Chinese chips."
- Photo attachment. Do not invent parameter counts from news clusters.
2026-09-04 — Unsloth local speedup (issuer docs, not a Z.ai card)
Chronological attach from 2026-09-04-x-overnight-opencode-omen-alpha-unsloth-glm-5-3-flash. Does not rewrite the Aug 27 @jietang numbers.
- From the same source (verified @UnslothAI, 2026-09-04 12:32 UTC): GLM-5.3-Flash now runs “3.3x faster locally”; “Local GGUF inference… 1.6–3.4× faster with optimized decoding and bonus multi-token prediction.”
- From the same source (fetched unsloth.ai/docs/models/glm-5.3-flash): banner “Sep 4: GLM-5.3-Flash now runs with 3.3x faster inference”; also known as
ox-alpha; 320B total / 18B active multimodal open model; 1M context; hardware table (1-bit ~100 GB through BF16 ~650 GB); tok/s example tg32 @ 65536: 20.66 → 48.99 before MTP n=2 examples. GGUF: unsloth/GLM-5.3-Flash-GGUF. - These specs are Unsloth-described. They were not in the Aug 27 @jietang post. Do not back-fill them onto that post. Do not treat them as a Z.ai model card. Full issuer page: unsloth.
- From the same source (@BAI_AGI): claimed #1 on B.AI with “2.41 trillion” cumulative tokens and restated 320B/18B / 1M — platform marketing, video not transcribed. Not a Z.ai card.
2026-09-04 — Omen Alpha as unconfirmed stealth (do not rename)
From the same source: OpenCode announced omen-alpha (Go exclusive; $100 usage for $10). Community posts and secondary press route it toward Zhipu/GLM the same way Ox Alpha was later named GLM-5.3-Flash. Z.ai posted nothing overnight confirming Omen. This page does not silently rename Omen to a GLM SKU. Identity stays on omen-alpha.
- Empty opencode.ai/data/zhipu/omen-alpha vs community screenshots — not reconciled.
- Lookonchain
$0.20/$0.66per million in/out is secondary. - @SPAC89 benchmark videos vs GLM-5.3 Flash — not transcribed; no numbers promoted.
2026-09-07 — AA Index v4.3 42 (not a rewrite of “AA = 57”)
From 2026-09-07-artificial-analysis-intelligence-index-v4-3. Chronological color only.
- From the same source (open-weights post + cost post): Intelligence Index v4.3 = 42; $0.25/Index-task vs GPT-5.6 Terra (max) $1.40 at the same Index band. Named on AA’s intelligence-vs-cost frontier with Fable 5.1 (53) and MiMo-V2.5-Pro (26).
- This is not the Aug 27 “AA = 57.” Different date, different (or unspecified) instrument. Leave both. Full treatment: artificial-analysis-intelligence-index.
What this source does not establish
- Not a US frontier-lab official post. Qualify in the clipping: Z.ai / GLM-associated.
- No parameter count, context length, or MoE shape in the Aug 27 post. Do not write cluster-summary specs from that post. Sep 4 Unsloth docs now state 320B/18B / 1M — Unsloth-described only.
- No fetched OpenRouter dashboard, AA methodology, or price table. "AA = 57" and "1/100 frontier price" are as written.
- Not a re-file of Kimi K3 or GLM 5.2. Those stay on chinese-open-weight-frontier-parity.
- @karpathy had zero original posts in the Aug 27 window. Official lab accounts were also silent in the 2026-09-04 overnight window.
- Omen is not this SKU. Do not collapse omen-alpha into GLM-5.3-Flash / GLM 5.4.
Open questions
- What does "AA = 57" measure (which Artificial Analysis index, which date)? The post does not say. Evening v4.3 print is 42 — not treated as the same number.
- Is "1/100 frontier price" a serving-price claim, and against which frontier SKU? The post does not say.
- Do Z.ai weights independently state 320B / 18B / 1M, or does that remain Unsloth-described?
- Is overnight omen-alpha a later GLM drop, or a different lab/SKU? Unconfirmed.
Sources
- 2026-08-27-x-ai-news-27-aug-2026-deepmind-double-blind-evals
- 2026-09-04-x-overnight-opencode-omen-alpha-unsloth-glm-5-3-flash
- 2026-09-07-artificial-analysis-intelligence-index-v4-3 — Index v4.3 42 / $0.25 per task; not flattened into Aug 27 “AA = 57”