AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: IBM’s CEO On Open-Weight AI And The On-Premises Volume Bet on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

IBM CEO Arvind Krishna said open-weight models can cost 10% to 25% as much as frontier models while performing about 90% as well, and argued that enterprises can run them on their own infrastructure. Artificial Analysis data reviewed October 26 puts the leading downloadable model at 80% of the top proprietary model; the closer comparison depends on weights that have not yet shipped. On-prem economics also hinge on hardware utilization, staffing and operating costs.

IBM CEO Arvind Krishna told investors on the company’s October 21 earnings call that open-weight AI models can cost 10% to 25% as much as frontier models while being “90% as effective,” making on-premises deployment attractive for some enterprise work. An independent model scoreboard reviewed October 26 shows the leading downloadable model at 80% of the leading proprietary model, while the closer comparison cited in the source depends on weights that have not yet been released.

Krishna described the cost-performance trade-off during the call’s question-and-answer session, saying a business might choose a model at “10% of the cost” if it is “90% as good as the best.” He pointed to routine tasks such as summarizing email, examining contracts and answering customer calls as work that may not require the most capable model. His cost figures are a company executive’s characterization, not a measurement tied to a named benchmark in the call.

Krishna also said open-weight models give companies permission and intellectual-property rights to run them on on-premises infrastructure or private clouds. He cited data sovereignty concerns among banks and traders, saying they want to keep underlying data and logs under their control. IBM has a commercial interest in the argument: it sells infrastructure and expects demand for on-prem AI to support products including its Spyre cards.

Artificial Analysis’s Intelligence Index, as read October 26, scores the best downloadable model listed, Xiaomi’s MiMo-V2.6-Pro, at 46.3, against Claude Opus 5.5 at 57.6. That makes the open model about 80% of the proprietary leader’s score. Kimi K3.1 Pro scores 55.1, but its maker, Moonshot, had only promised to release the weights “by Nov 4”; without the weights, it remains a closed model. The source says the model’s licence had not been published.

The index comparison tracks a changing field rather than a fixed gap. At the end of 2024, its best open model scored 7.7 against the proprietary leader’s 15.2, or 51%. At the end of 2025, the comparison was 22.4 against 30.4, or 74%. In October 2026, the gap in score points is 11.3, compared with 7.5 at the end of 2024, even as the open model’s score rose as a share of the leader’s. Both model categories advanced.

At a glance
reportWhen: Earnings call October 21, 2026; benchma…
The developmentOn IBM’s October 21 earnings call, CEO Arvind Krishna made a cost-and-performance case for open-weight models running on enterprise infrastructure, a pitch supported by a narrowing but still material benchmark gap.
Thorsten Meyer AI
Open vs. closed · Oct 2026

IBM's 90% claim, checked

Open weights reach 80% of the leader

Krishna says open-weight models are "90% as effective". On an independent index the best model you can download today reaches 80% of the best closed one, up from about half two years ago.

80%best open vs. best closed model, 26 Oct
11.3points gap (46.3 vs. 57.6)
96%if Kimi K3.1 Pro's weights ship
51%the same ratio at the end of 2024

Best open model as a share of the best closed

Artificial Analysis Intelligence Index, current version, best downloadable vs. best proprietary model at each date.

  • End of 2024Qwen2.5 72B 7.7 vs. o1 15.251%
  • End of 2025MiMo-V2-Flash 22.4 vs. GPT-5.2 30.474%
  • End of June 2026GLM-5.2 33.7 vs. Claude Fable 5 49.668%
  • October 26, 2026MiMo-V2.6-Pro 46.3 vs. Claude Opus 5.5 57.680%
  • If Kimi K3.1 Pro ships weights55.1 vs. 57.6 · promised “by Nov 4”96%

What Krishna said

Illustrations from the earnings call, not a named benchmark.

  1. Late 2024"70" vs. "20 or 30"Krishna, Q3 call, 21 Oct · no named benchmark
  2. Late 2025"70 to 50"Krishna, Q3 call, 21 Oct · no named benchmark
  3. Now"90% as effective"Krishna, Q3 call, 21 Oct · no named benchmark

Top downloadable models today

Intelligence Index score and licence type. Read the licence before the leaderboard.

ModelIndexLicence
MiMo-V2.6-Pro
Xiaomi
46.3MIT
DeepSeek V4.1 Flash 1017
DeepSeek
45.1MIT
GLM-5.3
Z AI
44.8own commercial licence
Step 5 Preview
StepFun
43.7own commercial licence
Muse Spark Open
Meta
38.9licence with conditions

In points the gap did not shrink: 7.5 points at the end of 2024, 11.3 today. Closed models moved too.

Sources: Artificial Analysis, Intelligence Index v4.3.2 (read 26 Oct 2026), IBM Q3 2026 earnings call transcript (21 Oct 2026), Artificial Analysis on Kimi K3.1 Pro (23 Oct 2026).
Powered by Thorsten Meyer AI

The Economics of Running AI In-House

The practical question for enterprises is not just whether an open model scores near a proprietary one. It is whether the savings on model use outweigh the cost of purchasing and operating the computing systems needed to run it. The source’s price examples illustrate why utilization matters: an eight-GPU H200 server lists from $172,643, while renting the same GPU class on AWS is listed at $63.30 an hour. At full utilization, the purchase-price comparison reaches parity in about 114 days; at 25% utilization, it takes about 15 months.

Those figures are a simplified comparison, not a full cost-of-ownership estimate. They do not include staff, cooling or networking, and the source does not specify every assumption behind the calculation. Krishna himself named a skills constraint, asking where enterprises would find kernel engineers and people to manage the underlying network. He said that gap likely remains outside the most sophisticated enterprises. Low token prices alone do not make on-prem deployment economical.

The choice also involves control and licensing, not only price. An open-weight licence can permit local use, but terms vary by model; downloadable weights do not automatically mean unrestricted commercial use. Meanwhile, the source’s examples show that hosted open models can be far cheaper than a leading proprietary model, but also that closed services may have low prices of their own. Enterprises need to compare the actual task, licence, operating requirements and current service prices rather than treating “open” as a guarantee of savings.

Products Worth Considering

Apple MacBook Air 15-inch (2026) with M5 chip, 16GB memory, 512GB SSD — Midnight

Apple MacBook Air 15-inch (2026) with M5 chip, 16GB memory, 512GB SSD — Midnight

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, Black

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, Black

As an affiliate, we earn on qualifying purchases.

What Krishna Said on the Call

Krishna framed the recent change as a narrowing performance gap. He recalled a late-2024 comparison in which a leading model might score 70 while an open model scored 20 or 30, followed by a late-2025 comparison of 70 to 50. He then made the widely cited “90% as good” and lower-cost argument. The figures were examples in his answer, not results from the Artificial Analysis index or another benchmark he named.

IBM also made a separate claim in prepared remarks that its llm-d project can run large models at five to 10 times lower cost. The supplied source says it could not independently check that claim. Krishna’s pitch comes as IBM has a stake in enterprise infrastructure spending; the source reports that IBM mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle, citing The Register. The earnings-call argument is therefore relevant both as an AI deployment view and as a commercial position from a hardware supplier.

The source also notes that model quality is only one part of the decision. Some available models use MIT licences, while others have custom or commercial terms. It identifies MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 as MIT-licensed, and says other models carry their own conditions. Companies must check the licence attached to the specific weights, as well as whether those weights are actually available.

The Open-Model Gap Is Still Moving

The benchmark figures are a snapshot of Artificial Analysis’s index as reviewed October 26, 2026. They do not establish that every open model performs at 80% of every proprietary model, or that the index reflects performance on a particular company’s workloads. Krishna’s “90%” statement was not tied to a named evaluation, so the two figures are not directly equivalent measures.

Kimi K3.1 Pro could narrow the measured gap if Moonshot releases the weights as promised, but they were not available in the source account and the model’s licence had not been published. Its reported 2.8-trillion-parameter size also raises deployment questions; the source describes it as a data-centre-scale project. The final model, licence and release timing remain unconfirmed in the material provided.

The cost examples also leave open how many hours a server would actually run, what workload it would serve, and what expenses beyond listed hardware or rental rates would apply. IBM’s llm-d savings statement remains an attributed company claim rather than an independently verified result here. Nvidia had not commented on the separate GeForce firmware issue described in the source, which is unrelated to the enterprise cost comparison.

Model Releases and Deployment Tests

The next near-term benchmark question is whether Moonshot releases Kimi K3.1 Pro’s weights by the promised November 4 date and publishes licence terms. If the weights arrive, independent evaluators and prospective users can assess the model as downloadable rather than relying on a hosted version. Until then, the index’s leading open-weight comparison remains MiMo-V2.6-Pro against Claude Opus 5.5.

For IBM, the next evidence needed is practical deployment data: how its cost claims perform on specified workloads, what infrastructure they require and how those results compare with hosted alternatives. Enterprises weighing on-premises systems will also need to estimate utilization and account for engineering staff, power and networking. The earnings-call remarks set out IBM’s case, but the supplied material does not provide a confirmed schedule for customer deployments or a full independently audited cost comparison.

Key Questions

What did IBM’s CEO claim about open-weight AI?

Arvind Krishna said open-weight models can cost 25% to 10% of frontier-model token costs while being “90% as effective.” He gave examples of routine business tasks, but did not tie the 90% figure to a named benchmark in the cited remarks.

Do independent scores confirm the 90% figure?

Not on the benchmark snapshot provided. Artificial Analysis’s October 26 index puts MiMo-V2.6-Pro at 46.3 versus Claude Opus 5.5 at 57.6, or about 80% of the leading proprietary score. Kimi K3.1 Pro is closer, at 55.1, but its weights had not yet been released.

Why would a company run a model on its own infrastructure?

Krishna cited data sovereignty and control over underlying data and logs. On-premises deployment may also reduce costs at high utilization, but requires suitable hardware, engineering expertise and operating capacity.

Does buying a server always cost less than renting GPUs?

No. The source’s simplified comparison says an eight-GPU H200 server’s listed purchase price equals about 114 days of AWS rental at full use, but about 15 months at 25% use. Staffing, cooling and networking are excluded, so actual costs can differ.

Are all downloadable AI models free for commercial use?

No. Model licences vary, and downloadable weights do not by themselves establish unrestricted commercial rights. The source identifies some MIT-licensed models and others with custom or commercial terms; users should review the licence for the specific model.

Source: ThorstenMeyerAI.com

You May Also Like

AI Automation Software For Small Businesses: A Labor Day sales Guide

Discover how small businesses can leverage AI automation software to save time, cut costs, and improve customer service with real-world examples and practical tips.

One upload in. A whole channel’s worth of content out.

ChannelHelm’s latest v1.5 update automates content repurposing, creating a full suite of posts from a single upload, saving creators time and effort.

The Ultimate Multi-Channel Reseller Listing Solution For Facebook Sellers

A new crosslisting solution for Facebook resellers is now in testing, aiming to streamline multi-channel listings directly from Facebook Marketplace.

Trade and supply-chain operations signal monitor: MEPs urge FIFA to investigate chief Infantino over Trump peace prize

European MEPs have called for FIFA to investigate President Gianni Infantino regarding allegations linked to the Trump peace prize, amid geopolitical tensions.