🔍 Read the full analysis: Could Open-Weight Models Shift AI On-Prem? IBM’s CEO On The Bet on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
IBM CEO Arvind Krishna said open-weight models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, making on-premises AI worth considering for some enterprise workloads. An Artificial Analysis comparison cited in the source puts the leading downloadable model at 80% of the leading proprietary model’s score; a higher-scoring model is not yet available with its promised weights.
IBM CEO Arvind Krishna used the company’s October 21 earnings call to argue that enterprises can run open-weight AI models on their own infrastructure at a fraction of frontier-model token costs. A comparison cited from Artificial Analysis supports the direction of his argument—the performance gap has narrowed—but places the best currently downloadable model at 80% of the top proprietary model’s score, not 90%.
In the call’s question-and-answer session, Krishna said open-weight models cost “25% to 10% of the token cost of the frontier models” while being “90% as effective.” He suggested that a large enterprise could begin with “two or three racks.” Those figures are Krishna’s characterization, not results from a benchmark he named. IBM sells infrastructure for enterprise AI, making it an interested party in the case for on-premises deployments.
Krishna framed the trade-off around routine tasks such as summarizing email, examining contracts and answering customer calls, where he argued a company may not need the strongest available model. He also cited data control: banks and traders, he said, may want AI running in a “sovereign” environment where underlying data and logs are not exposed to outside providers. Open weights can allow models to run on company hardware or private clouds, subject to the model’s licence and technical requirements.
The independent comparison cited in the source gives MiMo-V2.6-Pro a score of 46.3 and Claude Opus 5.5 a score of 57.6 on the Artificial Analysis Intelligence Index, or about 80% of the proprietary model’s score. The comparison is a snapshot read October 26, 2026, and does not establish that every model or task follows the same pattern. Kimi K3.1 Pro scores 55.1, about 96% of Opus 5.5, but Moonshot had only promised to release its weights “by Nov 4.” Until they are available, it is not a downloadable open-weight model.
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 Economics of Running Models In-House
The practical question for companies is not only whether an open model is capable enough, but whether owning and operating the required infrastructure costs less than paying for hosted inference. The answer depends on workload volume, hardware use and the cost of staff, networking and facilities. Lower per-token charges alone do not establish that an on-premises system will save money.
The source compares an eight-GPU H200 server listed from $172,643 with renting the same GPU class on AWS for $63.30 an hour. On those list prices, it estimates that a purchase takes about 114 days to match rental costs if the hardware runs at full load, or about 15 months at 25% utilization. These calculations exclude operating expenses such as cooling, networking and personnel, so they are not a complete cost comparison or a forecast of savings.
There is also an implementation constraint. Krishna himself pointed to shortages of “kernel engineers” and staff who understand the underlying network, saying that capability gap remains for all but the most sophisticated enterprises. For organizations without that expertise, a cloud service can be easier to adopt even if its listed inference price is higher.
Products Worth Considering
Canon IVY 2 Mini Photo Printer Bundle, Pure White
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How the Model Gap Has Changed
Krishna described model performance as converging, using illustrative scores: a top model at 70 versus an open model at 20 or 30 in late 2024, then 70 versus 50 in late 2025. Those examples were not tied to a named benchmark. The Artificial Analysis series cited in the source offers a separate, consistent comparison: open models rose from 51% of the proprietary leader’s score at the end of 2024 to 74% at the end of 2025 and 80% in the October 2026 snapshot.
The score gap has not narrowed in absolute points in that series: it was 7.5 points at the end of 2024 and 11.3 points in October 2026, because proprietary models also improved. The source says the comparison uses the current index version, read October 26, 2026. It also notes that models differ in licence terms: some use MIT licences, while others impose separate commercial or community conditions. Availability and licence rights matter alongside benchmark scores when a company considers deploying a model.
Krishna also said IBM’s llm-d project helps run large models at “5 to 10 times lower cost.” That remains an IBM claim in the source material; no independent measurement is provided there. The article source also points to IBM’s commercial interest in on-premises infrastructure, including its Spyre cards, and reports that IBM mainframe revenue fell 20% in the quarter. That business context does not by itself confirm or disprove the merits of Krishna’s model-cost argument.
Costs, Licences and Weights Still Matter
Actual deployment costs remain uncertain. The server-versus-rental calculation uses listed prices and specified utilization assumptions, but excludes staffing, cooling and networking. The source does not provide a full cost model for a real enterprise workload, nor does it show how token prices and hardware utilization would compare for a particular company.
The model scores also do not establish that open models match proprietary ones across every task. The comparison is an index snapshot, and Kimi K3.1 Pro’s higher score cannot be treated as evidence of available open weights until the promised release happens. The source says no licence had been published for that model and notes its 2.8-trillion-parameter scale would itself require substantial infrastructure.
Finally, the value of on-premises deployment depends on more than model access: firms must check each licence, have the staff to operate the system and determine whether their data-control requirements justify the additional work. It is not clear from the cited material how many enterprises will meet those conditions or adopt the approach.
Watch for Kimi’s Promised Weight Release
The nearest milestone in the source is Moonshot’s stated plan to release Kimi K3.1 Pro’s weights by November 4. Whether the files arrive, under what licence, and what hardware they require will determine whether its near-frontier score can inform a real on-premises comparison. The date is a stated promise, not confirmation that the release has occurred.
For enterprise buyers, the next step is a workload-specific comparison: test candidate models on the tasks they need, review licence terms, and measure the costs of hosting against cloud alternatives at realistic utilization. IBM’s cost claims and its llm-d statement may also attract scrutiny, but the supplied source does not identify an independent validation or provide further measurement details.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight models?
On IBM’s October 21 earnings call, Krishna said open-weight models could cost 10% to 25% as much per token as frontier models while being “90% as effective.” These are his claims, not figures tied in the source to a named benchmark.
How close is the best downloadable model to the leading proprietary model?
The October 26 Artificial Analysis snapshot cited in the source scores MiMo-V2.6-Pro at 46.3 and Claude Opus 5.5 at 57.6. That makes the open model’s score about 80% of the proprietary leader’s score in this comparison.
Is Kimi K3.1 Pro available as an open-weight model?
Not according to the source’s October 26 status. Moonshot had promised to release the weights by November 4, but until they are released, the model cannot be counted as downloadable open weights.
Does running an open model on-premises guarantee lower costs?
No. The cost depends on utilization, hardware, staffing, cooling, networking and other expenses. The source’s server-versus-rental comparison excludes several operating costs, so it is an illustration rather than a complete estimate for any particular enterprise.
Source: ThorstenMeyerAI.com
