🔍 Read the full analysis: Open-Weight Models Are Close: IBM’s CEO On AI At The Edge on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
IBM CEO Arvind Krishna said on the company’s October 21 earnings call that open-weight models can be about 90% as effective as frontier models at a fraction of their token cost. An Artificial Analysis comparison read October 26 put the strongest downloadable model at 80% of the leading proprietary model; a model that could narrow the gap has not yet released its weights.
IBM CEO Arvind Krishna said on the company’s October 21 earnings call that open-weight AI models can be “90% as effective” as frontier models at between 10% and 25% of their token cost. An Artificial Analysis comparison read October 26 put the best model with downloadable weights at 80% of the leading proprietary model, showing that the performance gap has narrowed but remains.
In the call’s question-and-answer session, Krishna argued that many routine enterprise tasks do not require the most capable available model. He cited email summaries, contract review and customer calls as examples, and said a large enterprise could begin running models on two or three racks. He also said open weights give organizations permission and intellectual-property rights to run models on their own infrastructure or private clouds.
Krishna’s cost and performance figures are his characterization, not results from a named benchmark. Artificial Analysis’s index, which applies the same 10 evaluations to models and identifies whether weights can be downloaded, scored Xiaomi’s MiMo-V2.6-Pro at 46.3 and proprietary Claude Opus 5.5 at 57.6 on its October 26 reading. That makes the downloadable model’s score about 80% of the leader’s; the figures do not establish that every open model is 80% as capable across all tasks.
A possible exception is Kimi K3.1 Pro, which scored 55.1 on the same index, or 96% of Claude Opus 5.5. But its maker, Moonshot, had only promised to release the weights by November 4. Until they are available, the model is not open-weight. The source also cautions that its reported 2.8-trillion-parameter scale would make running it a substantial infrastructure undertaking.
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 Cost of Running Enterprise AI
Krishna’s pitch puts the focus on a practical choice for companies: whether to pay for hosted model use or invest in infrastructure to run models themselves. Lower token prices can matter at high volumes, and keeping systems in a company’s own environment may appeal to organizations with strict data-control requirements. Krishna specifically cited banks and traders seeking AI systems that are “sovereign” and keep underlying data and logs under their control.
The comparison is not simply model price versus hardware price. The source’s listed example puts an eight-GPU H200 server at $172,643, compared with $63.30 an hour to rent the same GPU class on AWS. Its arithmetic estimates that buying could match rental costs after about 114 days at full use, but about 15 months at 25% use, before staffing, cooling and networking. Those estimates depend on the stated prices and utilization assumptions; they are not a general guarantee of savings.
That operational burden matters because Krishna acknowledged a shortage of people who can manage kernels and underlying networks outside the most sophisticated enterprises. He described frontier cloud services as the easier option, though more expensive. IBM also has a commercial interest in on-premises AI: it sells infrastructure and expects demand to support its Spyre cards. Its recommendation should be read with that interest in mind.
Products Worth Considering
Apple MacBook Pro 14-inch with M5 Max Chip (18-Core CPU, 20-Core GPU), 48GB RAM, 2TB SSD, Space Black
Apple MacBook Pro 16-inch Laptop with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 2023 14 with M3 Pro Chip, 18 GB RAM, 2 TB SSD, Space Black
As an affiliate, we earn on qualifying purchases.
How the Model Gap Has Changed
The Artificial Analysis figures show the strongest downloadable model’s score rising from 7.7 against a proprietary leader at 15.2 at the end of 2024, to 22.4 against 30.4 at the end of 2025, and then to 46.3 against 57.6 in October 2026. Expressed as a share of the proprietary score, that is about 51%, 74% and 80%, respectively. These are index readings at the listed dates, not a universal measure of performance on every business task.
The current point gap has not narrowed in the same way: the source lists it as 7.5 points at the end of 2024 and 11.3 points on October 26, 2026. The share of the leader’s score increased because the open model’s score grew relative to the leader, but proprietary models also continued to improve.
Costs and access vary by model and provider. The source says hosted DeepSeek V4.1 Flash 1017 costs 3% to 6% of Claude Opus 5.5’s list price, while OpenAI’s closed GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. It also cites differing model licences, including MIT and custom commercial terms. Downloadable weights do not mean identical usage rights; organizations need to check each licence as well as benchmark scores.
Performance, Licensing and Payback
The available figures do not establish that open-weight models are 90% as effective across all workloads. Krishna did not name a benchmark for that claim, while the Artificial Analysis index is one comparison using its own evaluations. Model rankings can also change as new systems and index versions appear.
Kimi K3.1 Pro’s weights had not shipped as of the source’s October 26 reading, and the source reports no published licence. Its high score therefore cannot yet be treated as evidence that enterprises can download and run it. For other models, licences may restrict commercial use or impose conditions, and the listed token prices do not account for every deployment cost.
The server-versus-rental calculation depends heavily on utilization and excludes staff, cooling and networking. The source does not provide enough information to treat the payback estimates as a full cost-of-ownership analysis. IBM’s prepared remarks also said its llm-d project could run large models at five to 10 times lower cost, but the source says it could not independently check that claim. Nvidia had not commented on the reported GeForce speed-limit workaround described in the source.
Watch for Weights and Deployment Data
The next near-term checkpoint identified in the source is Moonshot’s promised November 4 release of Kimi K3.1 Pro’s weights. Whether the release occurs, what licence accompanies it and what hardware it requires will determine how relevant its benchmark score is to organizations considering self-hosting.
For enterprises weighing deployments, the practical next steps are to test models on their own workloads, check licence terms and estimate utilization alongside hardware, staffing and hosting costs. Further public benchmark readings and concrete deployment data may show whether the narrowing performance gap translates into savings outside the index. IBM’s claims about Spyre and llm-d may also draw scrutiny as customers assess on-premises options.
Key Questions
What did IBM CEO Arvind Krishna claim about open-weight models?
On IBM’s October 21 earnings call, Krishna said open-weight models can be “90% as effective” as frontier models at 10% to 25% of their token cost. He did not tie that specific comparison to a named benchmark.
How close are downloadable models to the leading proprietary model?
In the Artificial Analysis reading dated October 26, 2026, Xiaomi’s MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5. That is about 80% of the proprietary model’s score on that index, not a finding about every task or model.
Is Kimi K3.1 Pro an open-weight model now?
Not according to the source’s October 26 status. Moonshot had promised to release its weights by November 4, but they had not yet shipped. The source also said no licence had been published.
Does running an open-weight model on-premises always cost less?
No. The economics depend on factors such as hardware cost, utilization, staffing, cooling and networking. The source’s rent-versus-buy example shows a much longer estimated payback at 25% utilization than at full load, and excludes several operating costs.
Source: ThorstenMeyerAI.com
