🔍 Read the full analysis: Can Open-Weight Models Make On-Prem AI Practical? on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
IBM CEO Arvind Krishna argued on the company’s Oct. 21 earnings call that open-weight models can deliver about 90% of frontier performance at a fraction of the token cost. An Artificial Analysis comparison read Oct. 26 put the strongest downloadable model at 80% of its proprietary leader, while on-prem economics depend on utilization, staffing and infrastructure—not token price alone.
IBM CEO Arvind Krishna said on the company’s Oct. 21 earnings call that open-weight models can be about 90% as effective as frontier systems at a fraction of their token cost, making on-premises AI a possible option for large enterprises. A benchmark snapshot from Artificial Analysis, read Oct. 26, puts the leading downloadable model at 80% of its proprietary rival, while hardware costs and staffing needs make the business case depend on how intensively companies use their systems.
In the call’s question-and-answer session, Krishna said open-weight models could cost 10% to 25% of frontier-model token costs while being “90% as effective.” He said a large enterprise could begin with “two or three racks.” Examples he gave for tasks that may not need the most capable model included summarizing email, examining contracts and answering customer calls. He also pointed to data control: banks and traders, he said, may want AI systems that keep underlying data and logs within their control.
The benchmark comparison is more qualified. On Artificial Analysis’s Intelligence Index, which evaluates models using the same ten assessments, Xiaomi’s downloadable MiMo-V2.6-Pro scored 46.3, compared with 57.6 for proprietary Claude Opus 5.5—about 80% of the leader’s score. The index showed the open-to-proprietary ratio rising from 51% at the end of 2024 to 74% at the end of 2025 and 80% on Oct. 26, 2026. These scores provide one comparison, not a universal measure of performance across every business task.
Kimi K3.1 Pro scored 55.1 on the same index, close to the leader, but Moonshot had only promised to release its weights “by Nov. 4.” Until the weights are available, it does not qualify as a downloadable open-weight model for this comparison. The source also reports that its licence had not been published and that the model has 2.8 trillion parameters, making deployment a substantial infrastructure undertaking. Licensing terms differ among models, so a high score alone does not establish that a company can use a model on its preferred terms.
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.
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.
What Krishna said
Illustrations from the earnings call, not a named benchmark.
- Late 2024"70" vs. "20 or 30"Krishna, Q3 call, 21 Oct · no named benchmark
- Late 2025"70 to 50"Krishna, Q3 call, 21 Oct · no named benchmark
- 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.
| Model | Index | Licence |
|---|---|---|
| MiMo-V2.6-Pro Xiaomi | 46.3 | MIT |
| DeepSeek V4.1 Flash 1017 DeepSeek | 45.1 | MIT |
| GLM-5.3 Z AI | 44.8 | own commercial licence |
| Step 5 Preview StepFun | 43.7 | own commercial licence |
| Muse Spark Open Meta | 38.9 | licence 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 On-Prem Case Depends on Utilization
For enterprises, the potential benefit is not simply a lower price per token. Running a model on company-controlled infrastructure can give organizations more control over where data and logs are processed, while reducing reliance on an outside model provider. That can matter to regulated firms or companies with strict data-handling requirements. Krishna described this as a sovereignty concern, but the degree of control depends on the model, deployment, licence and operational setup.
The financial comparison is sensitive to usage. The source lists an eight-GPU H200 server from $172,643 and AWS rental of the same GPU class at $63.30 an hour. Its calculation estimates that buying the server could match rental costs after about 114 days at full utilization, or about 15 months at 25% utilization. Those estimates exclude staff, cooling and networking, so they are not a complete cost comparison. Lower token prices do not automatically make owning hardware cheaper if equipment sits idle or needs significant support.
Krishna also acknowledged a deployment constraint: enterprises need people with expertise in model kernels and networking. He said that skills gap remains outside the most sophisticated companies. An organization evaluating on-prem AI must weigh that capacity alongside hardware spending, model quality, licensing and the value of keeping data under its own control.
Products Worth Considering
DJI Osmo Mobile 8 Standard Combo Phone Gimbal Stabilizer
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Portable Waterproof Speaker, Black
As an affiliate, we earn on qualifying purchases.
IBM’s Pitch and the Benchmark Timeline
Krishna’s remarks described a rapid improvement in open-model performance: he said the gap was much larger in late 2024 and narrower by late 2025. Those figures were illustrative rather than a citation to a named benchmark. The Artificial Analysis series supplied in the source offers a separate, consistent comparison: the best downloadable model scored 7.7 against a proprietary leader’s 15.2 at the end of 2024, then 22.4 against 30.4 at the end of 2025, and 46.3 against 57.6 on Oct. 26, 2026.
The ratio has improved, but the absolute point gap in that comparison did not shrink over the period: it was 7.5 points at the end of 2024 and 11.3 points on Oct. 26, 2026. Proprietary models also advanced. The source says the ratio would reach about 96% if Kimi’s promised weights arrive, but that is a conditional comparison, not the status of a currently downloadable model.
IBM has a commercial interest in the case for on-prem deployment: it sells enterprise infrastructure, including Spyre accelerators. The source also notes that IBM claimed its llm-d project can run large models at five to ten times lower cost, but provides no independently checked measurement for that claim. The figures should therefore be treated as IBM’s statements, not verified performance results.
Costs, Licences and Deployment Still Vary
The 90% claim is not a universal finding. Krishna’s figure was not tied to a named benchmark in the call, while Artificial Analysis measures performance on its own set of evaluations. A company’s results will depend on its tasks, model choice and operating conditions. The source does not provide a full, independently audited comparison of total on-prem costs against hosted services.
Model prices and deployment rights also differ. The source says DeepSeek V4.1 Flash 1017 has MIT-licensed weights and that its hosted price is 3% to 6% of Claude Opus 5.5’s list price. It also cites a much lower listed token price for closed GPT-6 Luna. These are different service and licensing arrangements; the comparisons do not establish that self-hosting is always less expensive. Some models use custom or commercial licences, and companies must review the applicable terms.
Actual infrastructure needs remain uncertain for any particular enterprise. The server and cloud figures are list-price comparisons, and the payback estimates exclude operating costs. It is also not clear from the source how many organizations can deploy and maintain the required systems with existing staff, or how performance will compare on their own workloads.
Watch for Model Releases and Real Costs
The next near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights by its promised Nov. 4, 2026 date, and what licence accompanies them. If the weights ship, the model’s index score would make the gap with Claude Opus 5.5 substantially smaller in this particular comparison; the release, licence and practical hardware requirements still need to be checked.
For enterprises, the next step is to compare models on their own workloads and calculate total cost using realistic utilization, staffing, cooling, networking and support assumptions. Further evidence about deployments—and independently verifiable measurements of IBM’s cost claims—would help clarify whether on-prem systems are practical beyond organizations already equipped to operate large AI infrastructure.
Key Questions
Are open-weight models already as capable as the leading proprietary models?
Not across the board, based on the cited index. On Oct. 26, 2026, Artificial Analysis scored the leading downloadable model, MiMo-V2.6-Pro, at 46.3, compared with 57.6 for Claude Opus 5.5. Results depend on the benchmark and the task.
Does running AI on-prem always cost less?
No. The cited buy-versus-rent calculation changes sharply with utilization and excludes staff, cooling and networking. On-prem may make sense for sustained use or data-control needs, but the source does not establish that it is cheaper for every organization.
What did IBM say about open-model costs?
On its Oct. 21 earnings call, IBM CEO Arvind Krishna said open-weight models can cost 10% to 25% of frontier-model token costs while being “90% as effective.” This is his claim, not a result independently verified by the cited benchmark.
Why do model licences matter?
Open-weight models do not all come with the same terms. Some use permissive licences, while others impose commercial or community conditions. Companies need to check whether a model’s licence allows their intended use before deployment.
What is the next development to watch?
Moonshot had promised Kimi K3.1 Pro weights by Nov. 4, 2026. Their release and licence would determine whether the model can be treated as downloadable for the comparison; the promised date alone does not confirm that it has shipped.
Source: ThorstenMeyerAI.com
