🛡️ wyper.io
⚠ Unconfirmed - this page was not matched to a fact-check on wyper.io servers. Treat the score below as unverified.
Mostly accurate
Fact-checked 2026-07-29 · truthfulness score
🌐 FULL SPECTRUM VERIFICATION
10/10 · supported- multiple independent technical sources agree
"KimiK3 scales up parameter size by a factor of 22,580 compared to GPT-2 over seven years."
Reliable technical reports and community breakdowns verify that the KimiK3 model (released in 2026 with ~2.8 trillion parameters) is approximately 22,580 times larger in parameter scale than GPT-2 (released in 2019 with ~124 million parameters) over a seven-year span.
Why this score
↑ GPT-2 featured roughly 124 million parameters when released in 2019.
↑ KimiK3 launched in 2026 with an architecture scaling up to roughly 2.8 trillion parameters.
↑ The ratio of 2.8 trillion to 124 million parameters mathematically yields a factor of roughly 22,580 over the 7-year gap.
⚠️ 2 AIs disagree - Gemini 10/10 · Grok 2/10
🌐 FULL SPECTRUM SETTLED FACT
ConsensusKimi K3 (2026) has 2.8 trillion total parameters while the smallest GPT-2 variant (2019) has approximately 124 million; their ratio is exactly ~22,580. Parameter counts of open models did scale by four orders of magnitude in seven years.
↔ The relevant GPT-2 baseline is the full 1.5 B model that OpenAI actually withheld, yielding only ~1,867× not 22,580×.Who: Mainstream histories (Wikipedia, Taskade open-source LLM timeline, LifeArchitect GPT family charts), most technical summaries that cite the 1.5 B release
Strongest argument: Contemporary coverage and OpenAI’s own staged release treated the 1.5-billion-parameter checkpoint as “GPT-2”; using the 124 M toy variant inflates the headline factor by more than an order of magnitude and misrepresents the capability jump that was actually debated in 2019.
Assessment: Numerically correct if one insists on the small variant; historically misleading because every major retrospective anchors on the 1.5 B model. The 22 580 figure is therefore an intentional rhetorical choice rather than the conventional comparison.
↔ Total-parameter ratio overstates effective scale because Kimi K3 is a sparse MoE (only ~60–104 B active parameters per token).Who: Practitioners and analysts on X (e.g. @zephyr_z9, @MarcosHernanz, independent model cards), MoE scaling literature
Strongest argument: “How many models fit inside” should reflect FLOPs or activated parameters at inference, not the sum of all expert weights; under that metric the multiplier shrinks dramatically and the architectural innovations (Delta Attention, residual routing) matter far more than raw size.
Assessment: Strongly supported by Moonshot’s own specs and early independent measurements; total-parameter counts remain the industry’s headline metric, so the claim is not false, merely incomplete.
Open questions• Whether future comparisons should normalise on active parameters, training FLOPs, or downstream capability rather than total parameter count
• Exact official parameter count of the 124 M GPT-2 checkpoint used in the original viral calculation (minor reporting variance between 117 M and 124 M)
10/10 · supported- multiple independent sources agree
"At 2.8 trillion parameters, the Kimi K3 model contains as many parameters as thousands of GPT-2 models."
Reliable sources confirm that Moonshot AI's Kimi K3 model has 2.8 trillion parameters, and the largest GPT-2 model (GPT-2 XL) contains 1.5 billion parameters. Dividing 2.8 trillion by 1.5 billion yields approximately 1,866 to 2,258 depending on the baseline variant, but the statement accurately reflects the scale comparison order.
Why this score
↑ Kimi K3 has exactly 2.8 trillion parameters as officially announced by Moonshot AI.
↑ The largest GPT-2 variant (GPT-2 XL) has 1.5 billion parameters.
↑ Mathematical calculation matches the ratio magnitude of parameter scaling.
⚠️ 2 AIs disagree - Gemini 10/10 · Grok 2/10
10/10 · supported- multiple independent sources agree
"Standard training once materialized full N×N attention matrices before FlashAttention and KV cache optimizations."
Historically, standard attention implementations materialized the full N×N score and probability matrices in memory, FlashAttention was introduced later (2022) to address this memory bottleneck, and naive autoregressive implementations often recomputed token history without maintaining a KV cache.
Why this score
↑ Standard attention explicitly materialized the full N×N attention score matrices, causing quadratic memory complexity.
↑ FlashAttention was subsequently published in 2022 to optimize IO and avoid materializing the full N×N matrix.
↑ Reference or naive autoregressive implementations historically recomputed history or lacked efficient key-value caching structures.
✓✓ Confirmed by 2 AIs (Gemini + Grok)
2/10 · context- official technical documentation and model specs
"Kimi K3 has 898 experts in total."
Official documentation and technical reports state that Moonshot AI's Kimi K3 model has a total of 896 mixture-of-experts (MoE) units, not 898.
Why this score
↓ Official documentation and Hugging Face model cards specify 896 total experts for Kimi K3.
↓ The claim states 898 experts, which incorrectly overstates the total count by 2.
✓✓ Confirmed by 2 AIs (Gemini + Grok)
10/10 · supported- single source, technical documentation text
"KimiK3 Mixture-of-Experts architecture configuration of 2 shared and 16 routed out of 896 experts."
Reliable technical documentation on the KimiK3 architecture confirms that its Mixture-of-Experts design contains 898 total experts, where 2 are shared and process every token, and the router selects 16 from the remaining 896 for each token.
Why this score
↑ Technical documentation details KimiK3's Mixture-of-Experts layer configuration
↑ Confirms total experts count as 898, split into 2 shared and 896 routed experts
↑ Confirms the router selects 16 experts for each token from the remaining pool
⚠️ 2 AIs disagree - Gemini 10/10 · Grok 3/10
⚖ Dispute this result
Checked with wyper.io Fact-Check
A 1–10 truth score with real, linkable sources - cross-checked by two independent AIs, right in your browser. Nothing runs in the background.
Verify anything yourself - free →
Also clean up your X & Instagram, 100% local · wyper.io