Local AI models ranked: an A to F tier list for businesses

The short answer first: among the AI models that can be run locally, Google’s Gemma 4 leads. Gemma 4 31B scores 18 of 20 points; together with Gemma 4 26B A4B, Qwen3.8 27B and GLM-5.3 Flash (16 points each) it forms Tier A. By the size of their weights, three of these four models fit on a single graphics card with 80 gigabytes (GB) of memory; GLM-5.3 Flash requires an eight-card server[1]. Kimi K3 and GLM-5.3 have the highest Arena ratings, yet land only in Tier C, because they need more than one server and carry a custom licence with conditions[2][3][4]. Llama 4 Maverick and Command A fall into Tier F because of their licences[5][6]. The basis is the Arena rating with style control, the size of the weight files and the vendors’ licence texts, as of 30 October 2026.

Disclosure: With Alwisr Company OS, Avolang itself operates a local AI agent and therefore has a commercial interest in the topic of local AI. The tier list says nothing about which models are used in Alwisr and contains no purchase recommendation for a product. The scoring scheme, measurements and sources are disclosed; Avolang’s own calculations and assessments are marked as such. All information is as of 30 October 2026. It does not replace legal advice, in particular not a review of licence terms in an individual case.

The tier list at a glance

Exhibit 1 shows the result. Each of the 21 models receives up to 20 points from three criteria: performance (Arena rating), operation (memory requirement of the weights) and licence. The score appears in brackets after the model name. How it comes about is explained in the section on method; all individual values are in Exhibits 3 and 4.

Exhibit 1: Tier list of local AI models from Tier A to F, as of 30 October 2026

Sources: Arena rating with style control[2], weight files on Hugging Face[1] and the vendors’ licence texts; points and tiers are Avolang’s own calculations. Command A arithmetically scores 8 points and is placed in Tier F because of its non-commercial licence.

Local AI models: what “local” and “open” mean

A tier list of local AI models first needs clear terms. Artificial intelligence (AI) in a company rests on language models, known as large language models (LLMs): trained computational models that understand and generate text. An AI assistant or AI agent is the product that surrounds such a model with memory, access rights and approvals. The tier list scores only the models, not the products around them.

“Local” here refers to location and allocation, not to an individual workstation. A local AI runs on servers in Germany, separate from other customers, and not at an AI provider abroad. Companies that want to run local AI therefore do not have to install it on their own premises. Avolang’s Alwisr Company OS offers three operating options, all of which run in Germany[7]:

• Hosted by Avolang. Agent and models run in a German data centre, in a separate area dedicated to the company.
• Dedicated server. Agent and models run on a server in Germany that is reserved for the company alone.
• On site on the company’s own hardware. Agent and models run on hardware at the company.

How much hardware a model requires determines the cost under each of these operating options (Avolang assessment). That is why memory requirement is a criterion of its own in the tier list.

The models in the tier list are “open” in the sense of published weights: the parameters of the trained model are available as files and can be run on your own hardware. That is not the same as open-source AI. Under version 1.0 of the Open Source Initiative’s definition, the preferred form for making modifications includes, besides the parameters, the code and information about the training data[8]. Whether each of the 21 models meets that definition was not checked. This article therefore speaks of open models, not of open-source AI.

How local models compare with ChatGPT and Claude is covered in the article ChatGPT, Claude vs. a local AI agent in Germany. When running your own infrastructure makes sense overall is described in the article on local IT infrastructure.

Comparing AI models: data and method

The tier list combines three measures that can be checked publicly. The selection of models is Avolang’s own: it comprises the current open models from 14 vendors and a few older models that are still under discussion in companies. The prerequisites were an unambiguous overall entry in the Arena and a published standard checkpoint whose size can be determined on Hugging Face. The list is not a full survey: in the Arena, 227 of 419 entries are not labelled proprietary (Avolang’s own count)[2].

Performance: the Arena rating with style control

Performance is measured by the Arena (arena.ai, described in a 2024 paper as “Chatbot Arena”). It collects votes from many users who rate two models against each other in pairwise comparisons[9]. From the votes, the Arena computes a rating based on the Bradley-Terry model, with a lower and an upper confidence bound for each value[2]. The interval between the two bounds shows how reliable a value is.

This article uses the rating with style control. It takes response length and the number of headings, bold text and lists into account as influencing factors[10], so that the presentation of an answer is not mistaken for its quality. Since 16 May 2025, style control has been the default view of the text rating[2]. The overall rating (“overall”) and the category “german” were used, in the release of 28 October 2026, retrieved on 30 October 2026[2].

The overall Arena rating is translated into one of seven performance classes (0 to 6). One class is 30 points wide and therefore wider than the confidence interval of each of the 21 models (5 to 21 points; Avolang’s own calculation).

Operation: memory requirement of the weights

The size of the weight files of the standard checkpoint, that is, the version the vendor provides as its main model, serves as the measure of operating effort[1]. The file sizes were summed in gigabytes (decimal) and compared with cards of 80 GB each; an H100 card has 80 or 94 GB depending on the variant[11]. This yields four operating classes: up to 80 GB (one card), up to 320 GB (two to four cards), up to 640 GB (one eight-card server) and above (more than one server).

Licence: three gradations and one exclusion

The vendor’s licence text is decisive, not the short note on a league-table or model page. Two points go to models with a standard licence: Apache 2.0[12], MIT[13] or the Linux Foundation’s OpenMDW-1.1[14]. One point goes to models with a custom licence that attaches conditions to use: Kimi K3[3] and GLM-5.3[4]. Zero points and an exclusion (“licence knock-out”) apply where the licence rules out commercial use or carves companies in the EU out of the usage rights: Command A[6] and Llama 4 Maverick[5].

Points and tiers

The total score is 2 × performance class + 2 × operating class + licence points, at most 20 points. The weights of 60, 30 and 10 percent are Avolang’s own choice (Avolang assessment): performance counts most, because a model that does not solve the task is of no use even on small hardware. Operation counts for 30 percent, because the memory requirement determines how much hardware a model needs. The licence counts for only 10 percent, but acts as an exclusion criterion. One performance class and one operating class are each worth two points, which is exactly one tier; Tiers B to E are two points wide.

Exhibit 2: Scoring scheme of the tier list

The scoring scheme is Avolang’s own choice (Avolang assessment). The performance-class boundaries are 1,290 plus multiples of 30.

LLM ranking in detail: tier by tier

The following sections place the results tier by tier. All measurements are in Exhibits 3 and 4 at the end of this section.

Tier A: Gemma 4, Qwen3.8 and GLM-5.3 Flash

Google’s Gemma 4 31B leads the tier list with 18 of 20 points. The Arena rating of 1,452 (interval 1,445 to 1,460) means performance class 5; the weights come to 62.5 GB and fit on a single card with 80 GB; the licence is Apache 2.0. Gemma 4 26B A4B (1,438) and Qwen3.8 27B (1,436) follow with 16 points each, likewise on one card and under Apache 2.0.

Gemma 4 26B A4B is a so-called mixture-of-experts model: it consists of 25.2 billion parameters, of which only 3.8 billion are active per word fragment (token)[1]. All parameters still count towards the memory requirement, so the weights come to 51.6 GB.

Z.ai’s GLM-5.3 Flash has the highest Arena rating in Tier A, 1,477, and reaches performance class 6. At 328.3 GB, however, it requires an eight-card server and receives only operating class 1 of 3; together with the MIT licence, this yields 16 points.

In the Arena, the four models trail the top rating of all 419 entries (1,528, “Gemini 4 Carbon (High)”, proprietary) by 51 to 92 points[2]. Anyone looking for the highest Arena rating will find it among proprietary models.

At around 6,100 votes, the Gemma 4 models have clearly fewer votes than Qwen3.8 27B (27,779) and GLM-5.3 Flash (41,642); their intervals, at 15 points, are correspondingly wider than those of the other two (9 points)[2]. The interval of Gemma 4 31B lies entirely within one performance class; for Gemma 4 26B A4B and Qwen3.8 27B, it touches the class boundary of 1,440.

Tier B: DeepSeek V4 Flash and the Qwen3.5 models

DeepSeek V4 Flash reaches the same Arena rating as Qwen3.8 27B (1,436), but at 159.6 GB it needs two to four cards and therefore scores 14 points. Qwen3.5 122B A10B (1,416, 250.2 GB) is in the same hardware class. At 71.9 GB, Qwen3.5 35B A3B still fits on one card, but at 1,394 it falls into performance class 3. On a card with 80 GB, a model of this size leaves only about 8 GB for the context and intermediate values (Avolang’s own calculation), so running it on one card is tight. The three models are the obvious alternatives if the Tier A models do not suffice for your own tasks (Avolang assessment).

Tier C: two groups with opposite profiles

Kimi K3 (1,488) and GLM-5.3 (1,481) have the highest Arena ratings of the 21 models; Kimi K3 trails the top rating of all entries by 40 points. Both receive zero operating points: with 1,560.9 GB of weights, Kimi K3 needs at least three eight-card servers with 80 GB cards, and with 755.6 GB, GLM-5.3 needs two (Avolang’s own calculation). In each case a one-point deduction for the custom licence applies, so both end up at 13 points. Inkling from Thinking Machines (1,442) faces the same problem with 1,904.8 GB of weights, needs at least three eight-card servers and reaches 12 points.

The second group is the reverse. Mistral Small 3.2 (1,356, 48.0 GB) and gpt-oss-120b (1,352, 65.2 GB) run on one card and are under Apache 2.0, but reach only performance class 2. For gpt-oss-120b, OpenAI itself states that it fits on a single H100 graphics processor[15].

NVIDIA’s Nemotron 3 Ultra (1,427, 352.3 GB) sits in between: performance class 4, an eight-card server and the OpenMDW-1.1 licence. That makes 12 points.

Tier D: when hardware or performance do not fit

Mistral Large 3 (1,414) has the highest performance in this tier and, at 681.5 GB, is just above the 640 GB boundary. Were the weights under 640 GB, for example in a compressed version, it would score 12 points and land in Tier C (Avolang’s own calculation); how compression affects quality has not been measured here. Inkling Small (1,406, 531.9 GB) needs an eight-card server. IBM’s Granite 4.2 30B (1,340) and Allen AI’s Olmo 3.1 32B (1,329) run on one card but reach only performance class 1.

Tier E: very frugal, but clearly weaker

At 13.8 GB, gpt-oss-20b (1,317) needs only a fraction of a card; Microsoft’s Phi-4 (1,256) comes to 29.3 GB. Both receive full marks for operation and licence, but only performance class 0: Phi-4 trails Gemma 4 31B by 196 points in the Arena, gpt-oss-20b by 135 points. Whether such a model suffices for a narrowly defined task can only be shown by a test with your own tasks (Avolang assessment).

Tier F: the licence rules out use

Cohere’s Command A (1,354, 222.1 GB) arithmetically scores 8 points, but is under the CC BY-NC 4.0 licence and is therefore available for non-commercial use only. Meta’s Llama 4 Maverick (1,327, 803.2 GB) comes to 2 points; here the acceptable use policy carves companies with a principal place of business in the EU out of the usage rights. The section on licence pitfalls covers both cases.

All measurements and points

Exhibit 3 contains the measurements of all 21 models, Exhibit 4 the allocation of points. The brackets in Exhibit 3 show the confidence interval of the Arena.

Exhibit 3: Measurements by model: Arena rating with style control, size of the weights and licence

Sources: Arena, release of 28 October 2026[2]; file sizes and licence information on Hugging Face[1]; sizes in decimal gigabytes, Avolang’s own calculation.

Exhibit 4: Points awarded and tier by model

Avolang’s own calculation using the scoring scheme in Exhibit 2. Performance class and operating class count double.

Hardware: what the models demand in memory

Exhibit 5 sorts the models by the memory requirement of their weights. The classes are based on cards of 80 GB each[11].

Exhibit 5: Models by memory class of the weights

Avolang’s own calculation from the file sizes of the standard checkpoints on Hugging Face[1].

10 of the 21 models fit on one card, 3 need two to four, 3 an eight-card server and 5 more than one server. A single card is already enough to reach Tier A. The best single-card model, Gemma 4 31B (1,452), trails the best rating among the 21 models (Kimi K3, 1,488) by 36 Arena points; that model needs at least three eight-card servers, and the intervals do not overlap. Whether this difference is noticeable on a specific task can only be shown by a test of your own (Avolang assessment).

The classes are lower bounds: memory for the context and for concurrent users comes on top (Avolang assessment). At 352.3 GB, the model card of Nemotron 3 Ultra lists eight H100 cards, among other things, as the minimum configuration[1], which corresponds to the eight-card server class. Compressed versions (quantisation) need less memory but can change quality; that, too, has not been measured here.

Licence pitfalls: where an open model is not free

“Open” does not mean “free of conditions”. Five points should be checked before use. Assessing an individual case is a matter for legal advice.

1. Llama 4: exclusion for companies in the EU.Under Meta’s acceptable use policy, the rights under section 1(a) for multimodal Llama 4 models are not granted to individuals domiciled in, or companies with a principal place of business in, the EU. The restriction does not apply to end users of a product or service that incorporates such a model[5]. Llama 4 Maverick is natively multimodal[16]. On redistribution, the licence also requires the notice “Built with Llama” and a copy of the licence[17].
2. Command A: non-commercial only.The CC BY-NC 4.0 licence defines “NonCommercial” as not primarily intended for or directed towards commercial advantage or monetary compensation[6]. Use in business operations therefore needs a separate agreement with the vendor (Avolang assessment).
3. Kimi K3: custom licence with two conditions.The Kimi K3 License follows the pattern of the MIT licence. If the licensee or an affiliate runs a “Model as a Service” business (third parties are given access to inference or fine-tuning with substantial control over inputs, parameters or training data, for example via an application programming interface (API)) and the combined revenue of both exceeds US$20 million in twelve consecutive months, a separate agreement with Moonshot AI is required before any commercial use. Commercial products with more than 100 million monthly active users or more than US$20 million in monthly revenue must display “Kimi K3”. Neither obligation applies to internal use[3]. The one licence point deducted still decides the tier: with two points, Kimi K3 would reach 14 points and Tier B (Avolang’s own calculation).
4. GLM-5.3: security review for very large providers.The GLM-5.3 License likewise follows the pattern of the MIT licence. If the licensee or an affiliate runs a “Model as a Service” business and the combined revenue of both exceeds US$10 billion in twelve consecutive months, the licensee must pass Z.ai’s security review before any commercial use[4]. The Arena lists GLM-5.3 with the licence entry “MIT”[2]; the licence text in the model’s repository is decisive. With two licence points, GLM-5.3 would likewise reach 14 points (Avolang’s own calculation).
5. Standard licences also carry obligations.Apache 2.0 and MIT require the licence and the copyright notices on redistribution[12][13]. OpenMDW-1.1 requires a copy of the licence and the notices on redistribution and places no restrictions on the model’s outputs[14]. In purely internal operation, a model is generally not redistributed (Avolang assessment); anyone building it into a product for customers should check the obligations.

German: how the models fare in the German-language rating

Besides the overall rating, the Arena maintains a category “german” with 320 entries. A rating is available there for 18 of the 21 models; it is missing for Gemma 4 31B, Gemma 4 26B A4B and Granite 4.2 30B[2].

The German rating rests on clearly fewer votes. At the median of the 320 entries, the number of German votes equals 2.2 percent of the votes in the overall rating (Avolang’s own calculation); for the 18 models with a German rating, 212 to 1,446 votes stand behind the German value, against 3,389 to 88,325 for the overall rating. The intervals are correspondingly wide: 33 to 82 points against 5 to 21 points in the overall rating. For 15 of 18 models, the overall value lies within the German interval. The three exceptions have a lower value in German: Qwen3.5 35B A3B (1,394 overall, 1,360 German), gpt-oss-120b (1,352 against 1,328) and gpt-oss-20b (1,317 against 1,274).

The order is largely the same in both ratings. Spearman’s rank correlation coefficient, where a value of 1 means an identical order, is 0.990 across the 154 entries not labelled proprietary that have a rating in both categories, and 0.971 across the 99 of them with an overall value of 1,300 or higher (Avolang’s own calculation)[2]. For the order, the overall rating is thus a usable approximation, but not for fine differences: for the 18 models, the German value deviates from the overall value by minus 43 to plus 23 points (Avolang’s own calculation).

Vendor statements on language coverage, for example “35+ languages” for Gemma 4[1], say nothing about the quality of the German. For texts from your own company, a test of your own with German-language tasks makes sense (Avolang assessment).

Limits of the tier list and models not scored

• Conversational preference, not task performance. The Arena measures which answer users prefer in open conversations. It does not show how well a model summarises contracts, checks invoices or handles tools (Avolang assessment).
• Measurement uncertainty. For 8 of 21 models, the confidence interval touches a class boundary: Gemma 4 26B A4B, Qwen3.8 27B, DeepSeek V4 Flash, Inkling, gpt-oss-120b, Inkling Small, Granite 4.2 30B, gpt-oss-20b (Avolang’s own calculation). If the performance class were one class lower for all models, GLM-5.3 Flash, Gemma 4 26B A4B and Qwen3.8 27B would slip from Tier A to B; Gemma 4 31B would remain in Tier A (Avolang’s own calculation).
• Data basis. With the raw values without style control, Nemotron 3 Ultra would be in Tier B, Inkling Small and Granite 4.2 30B in Tier C, Mistral Small 3.2 in Tier D and Olmo 3.1 32B in Tier E; Tier A would remain unchanged (Avolang’s own calculation from the configurations “text” and “text_style_control”)[2].
• Weighting. The weights are Avolang’s own choice. Across 57 weightings that were calculated (performance 30 to 80 percent, operation 10 to 50 percent, licence 0 to 30 percent, in steps of five percentage points and summing to 100 percent), Gemma 4 31B ranks first or second among the 19 models without a licence exclusion every time, and first in 48 of the 57 weightings. Gemma 4 26B A4B and Qwen3.8 27B rank between second and sixth. Of the four Tier A models, GLM-5.3 Flash reacts most strongly (rank 1 to 13): anyone who weights hardware more heavily will downgrade it (Avolang’s own calculation).
• Assignment of models. The Arena lists Kimi K3 and GLM-5.3 in “Max” mode. The tier list assumes that this mode corresponds to the published weights; this is not documented. For Nemotron 3 Ultra, the Arena entry is the compressed NVFP4 version, which matches the file size that was checked[2][1].
• Operation by weights only. Throughput, concurrent users and context length are not taken into account (Avolang assessment).
• Licence and origin. Supplementary acceptable use policies of the vendors were not checked for every model. The origin of the vendors and of the training data was not assessed; whether it matters for a company belongs in its own AI policy (Avolang assessment).
• A league table is one view. Different league tables order open models differently. A post on DEV Community, written by a single author, compared two league tables on nine matching open models and found a Kendall rank correlation of 0.61: similar, but not the same[18]. The Arena, too, is only one view of performance.

Models were not scored where the Arena entry, the weights or the licence could not be matched unambiguously:

• Mistral Large 4: The weights (1,060.6 GB) are under the Mistral Ecosystem Open-Weight License, which requires a separate commercial licence from providers of “Model as a Service” without a revenue threshold[19][20]. The Arena lists “Mistral Large 4” as proprietary (1,432); a match to the published weights is not documented.
• Mistral Small 4 119B: Apache 2.0, 120.9 GB[19], but no Arena entry was found.
• DeepSeek V4.1 Flash: The Arena entry “DeepSeek-V4.1-Flash-20260910 (Max)” (1,476) and the weights “DeepSeek-V4.1-Flash-1017” (510.3 GB, MIT) carry different dates; the match is unclear[2][19].
• MiMo-V2.6: The best Arena entry, MiMo-V2.6-Pro (1,480, 4,037 votes), has no German rating; no unambiguously matching standard checkpoint was found[2].
• MiniMax-M3: The Arena names the “MiniMax Community License” (1,439); the licence text was not read[2].

Other entries without a proprietary label, such as DeepSeek V4 Pro or MiMo-V2.5-Pro, were not checked. The tier list does not claim to be complete.

Five questions before choosing a model

1. Which tasks should the model solve?The tier list does not replace a test. Five to ten recurring tasks with fixed quality criteria are a sound basis.
2. Which data does it process?Personal and confidential data argue for operation in Germany, separate from other customers. The data classes belong in the AI policy.
3. How many people use it at the same time?The size of the weights is only the lower bound. Concurrent users and long contexts increase the memory requirement (Avolang assessment).
4. Which hardware is available or obtainable?Exhibit 5 shows which tier is reachable in which memory class. Which operating option fits is described on the page on Alwisr Company OS.
5. Which licence and which origin are acceptable?The licence pitfalls show where an open model attaches conditions. The origin of the vendors is a matter for your own AI governance.

Four steps to choosing a model

The tier list narrows the field; the decision is made on your own use case. Four steps are a workable approach (Avolang assessment):

Frequently asked questions about local AI models

Conclusion

With local AI models, the decisive factor is not the highest Arena rating but the interplay of performance, hardware and licence. The best rating among the 21 models, Kimi K3, requires several servers and a licence with conditions. The best overall balance, Gemma 4 31B, runs on one card under Apache 2.0 and trails it by 36 Arena points. Companies that want to introduce local AI models should test two or three candidates from Tiers A and B on their own tasks and only then decide on hardware or operating option (Avolang assessment).