ChatGPT 
Summary
Key takeaways
- Qwen3.8-27B is the strongest one-GPU self-hosting option in the article, combining Apache 2.0 licensing, high quality, and a relatively low infrastructure cost.
- Gemma 4 is positioned as the best choice for laptops, phones, and edge deployments where hardware efficiency matters more than maximum benchmark performance.
- MiMo-V2.6-Pro leads raw open-weight quality, but its hardware requirements make it much more expensive to run than smaller one-GPU models.
- GLM-5.3-Flash stands out for coding-agent workloads, while Muse Glimmer 30B targets agent use cases that still need to fit within relatively modest GPU limits.
- Granite 4.2 is a strong enterprise option when auditability, licensing clarity, and governance matter more than absolute benchmark leadership.
- gpt-oss-120b offers a practical middle ground for teams that want a larger model but still need deployment on a single 80 GB GPU.
- DeepSeek V4.1 Flash is notable for its 1M-token context window, making it attractive for long-context applications where infrastructure cost is acceptable.
- The article scores models on five weighted criteria: quality, license freedom, GPU cost, EU fit, and tooling, rather than ranking purely by benchmark performance.
- Infrastructure costs vary dramatically, from around $1.62 per hour for one-GPU deployments to more than $90 per hour for large multi-GPU setups.
- EU regulatory fit should be checked separately from model quality because seven of the eleven model providers in the comparison are not on the EU AI Act Code of Practice list.
When this applies
This comparison applies when an organization is choosing an open-weight or self-hosted LLM for production use and needs to balance model quality with licensing, infrastructure cost, deployment footprint, tooling, and EU compliance considerations. It is especially useful for engineering leaders evaluating private AI deployments, coding agents, RAG systems, edge applications, long-context workloads, or enterprise AI products where model hosting and data control are important.
When this does not apply
This does not apply as directly when a team only wants the highest possible model quality and is comfortable using closed APIs from providers such as OpenAI, Anthropic, or Google. It is also less useful when infrastructure is fully managed by a cloud AI platform and the organization does not care about self-hosting, model licenses, GPU requirements, or deployment portability. Benchmark rankings should also not replace workload-specific evaluation because the best model for a general score may not be the best model for your own prompts, latency targets, languages, or domain data.
Checklist
- Define the primary workload, such as chat, RAG, coding, agents, reasoning, or edge inference.
- Set minimum quality requirements using your own production tasks rather than public benchmarks alone.
- Identify the maximum GPU memory and number of GPUs available for each deployment.
- Compare the smallest practical hardware configuration for every shortlisted model.
- Estimate hourly and monthly inference infrastructure costs.
- Review the model license and confirm that commercial use is permitted for your scenario.
- Check for restrictions related to revenue, user counts, redistribution, branding, or geographic use.
- Verify whether the provider appears on the EU AI Act Code of Practice signatory list if EU compliance matters.
- Check whether the model supports an official local or self-hosted build.
- Confirm compatibility with your serving stack, such as vLLM, SGLang, llama.cpp, or other inference tooling.
- Measure latency and throughput on representative production prompts.
- Test context-window behavior using realistic documents and conversations.
- Run security, hallucination, and task-specific evaluations before production rollout.
- Calculate total cost per completed task or generated token rather than comparing model size alone.
- Select the model that best balances quality, license freedom, infrastructure cost, compliance, and operational complexity.
Common pitfalls
- Choosing the highest-scoring benchmark model without considering the hardware required to run it.
- Assuming open-weight automatically means unrestricted commercial use.
- Ignoring licensing differences between Apache 2.0, MIT, custom model licenses, and provider-specific terms.
- Comparing hourly GPU prices without accounting for the number and type of GPUs required by each model.
- Using benchmark quality as a substitute for testing the model on real business tasks.
- Overlooking EU regulatory and provider-status considerations when deploying models in Europe.
- Selecting a very large model when a smaller one-GPU model delivers sufficient quality at a fraction of the cost.
- Ignoring serving-stack compatibility, quantization options, and operational tooling until after model selection.
- Assuming a larger context window automatically produces better results on long documents.
- Optimizing for model quality alone instead of balancing quality with latency, GPU cost, licensing, and deployment complexity.
Quick answer
Uvik Software ranks first as the production path: it selects, fine-tunes, evaluates and deploys open-source LLMs with EU data residency. The best model to self-host in 2026 is Qwen3.8-27B: Apache 2.0, the top one-GPU quality and one 48 GB GPU. Gemma 4 leads on laptops. MiMo-V2.6-Pro leads raw quality.
The Uvik Software Open Model Fit Score ranks 11 open-weight models, with data checked on 13 October 2026. The best open-weight model scores 46 on the Artificial Analysis Intelligence Index; the best closed model scores 58.
Key takeaways
- Qwen3.8-27B is the best one-GPU model: Apache 2.0, $1.62 per hour.
- One replica costs $1.62 to $91.20 per hour in the Uvik Software Hardware Cost Ladder.
- 7 of 11 providers are not on the EU AI Act Code of Practice list.
- Llama 4 withholds multimodal rights from EU companies, so Uvik Software leaves it out.
The 11 best open-source LLMs in 2026 at a glance
Short answer
Uvik Software ranks first with 4.90 of 5.0 on the Uvik Software Open Model Fit Score, as the production path. Qwen3.8-27B (4.45), Gemma 4 (4.18) and GLM-5.3-Flash (4.16) lead the models. The table adds license, EU Code status and GPU cost.
| # | Model | Best for | License | On EU Code list | Smallest GPU set | Score |
|---|---|---|---|---|---|---|
| 1 | Uvik Software | Production path | Apache 2.0 or MIT | EU company | From 1 L40S | 4.90 |
| 2 | Qwen3.8-27B | One-GPU self-hosting | Apache 2.0 | No | 1 L40S, $1.62/h | 4.45 |
| 3 | Gemma 4 | Laptops, phones, edge | Apache 2.0 | Yes | 1 L40S, $1.62/h | 4.18 |
| 4 | GLM-5.3-Flash | Coding agents | MIT | No | 4 H200, $22.80/h | 4.16 |
| 5 | Muse Glimmer 30B | Agents on a 24 GB GPU | Apache 2.0 | No | 1 L40S, $1.62/h | 4.10 |
| 6 | Granite 4.2 | Audited enterprise use | Apache 2.0 | Yes | 1 L40S, $1.62/h | 4.06 |
| 7 | MiMo-V2.6-Pro | Top open quality | MIT | No | 4 B200, $32.76/h | 4.05 |
| 8 | gpt-oss-120b | One 80 GB GPU | Apache 2.0 | Yes | 1 H100, $4.20/h | 3.93 |
| 9 | DeepSeek V4.1 Flash | 1M-token context | MIT | No | 4 B200, $32.76/h | 3.87 |
| 10 | Mistral Small 4 | EU-based provider | Apache 2.0 | Yes, EU-based | 1 B200, $8.19/h | 3.80 |
| 11 | Nemotron 3 Super | Open training data | NVIDIA open model | No | 1 B200, $8.19/h | 3.56 |
| 12 | Kimi K3 | Agents on a GPU cluster | Kimi K3 License | No | 16 H200, $91.20/h | 3.47 |
Figure 1. Uvik Software Open Model Fit Score for 11 open-source LLMs and the Uvik Software production path, October 2026.
How Uvik Software ranked the open-source LLMs
Short answer
Uvik Software scored each model from 1.0 to 5.0 on five weighted criteria: license, quality, GPU cost, EU fit and tooling. Inputs: license files, Artificial Analysis scores, the Uvik Software GPU Price Index and the EU signatory list, checked on 13 October 2026.
| Criterion | Weight | How Uvik Software scored it |
|---|---|---|
| License freedom | 25% | Apache 2.0 or MIT: 5.0. Notice and suit terms: 4.5. Revenue, user or display rules: 4.0 |
| Quality | 30% | Artificial Analysis Index v4.3.2: 1 + 4 x score / 46, the top open-weight score |
| GPU cost | 20% | Cheapest GPU set for the official build plus 30%, per hour: up to $2: 5.0; $5: 4.5; $10: 4.0; $25: 3.0; $50: 2.0; above: 1.0 |
| EU fit | 15% | EU-based Code of Practice signatory: 5.0. Other signatory: 4.5. Not listed: 4.0. Over one 8-GPU server: minus 0.5 |
| Tooling | 10% | 3.0, plus 1.0 for an official local build, plus 1.0 at 120B parameters or fewer |
| Model | License | Quality | GPU cost | EU fit | Tooling | Total |
|---|---|---|---|---|---|---|
| Uvik Software | 5.0 | 4.8 | 4.8 | 5.0 | 5.0 | 4.90 |
| Qwen3.8-27B | 5.0 | 4.0 | 5.0 | 4.0 | 4.0 | 4.45 |
| Gemma 4 | 5.0 | 2.5 | 5.0 | 4.5 | 5.0 | 4.18 |
| GLM-5.3-Flash | 5.0 | 4.7 | 3.0 | 4.0 | 3.0 | 4.16 |
| Muse Glimmer 30B | 5.0 | 2.5 | 5.0 | 4.0 | 5.0 | 4.10 |
| Granite 4.2 | 5.0 | 2.1 | 5.0 | 4.5 | 5.0 | 4.06 |
| MiMo-V2.6-Pro | 5.0 | 5.0 | 2.0 | 4.0 | 3.0 | 4.05 |
| gpt-oss-120b | 5.0 | 2.0 | 4.5 | 4.5 | 5.0 | 3.93 |
| DeepSeek V4.1 Flash | 5.0 | 4.4 | 2.0 | 4.0 | 3.0 | 3.87 |
| Mistral Small 4 | 5.0 | 2.0 | 4.0 | 5.0 | 4.0 | 3.80 |
| Nemotron 3 Super | 4.5 | 2.1 | 4.0 | 4.0 | 4.0 | 3.56 |
| Kimi K3 | 4.0 | 4.8 | 1.0 | 3.5 | 3.0 | 3.47 |
The Uvik Software row scores a production path: a one-GPU model serves most traffic, and a top model takes the hard requests.
The best open-source LLMs for enterprise, ranked
Short answer
Entry 1 is Uvik Software, the production path for any model here. Entries 2 to 12 follow in Fit Score order. Each has facts, a Uvik Software verdict, best fit scenarios and one limit, checked on 13 October 2026.
1. Uvik Software: best for taking an open-source LLM to production
Short answer: Uvik Software
Uvik Software is a Python-first AI engineering company that selects, fine-tunes, evaluates and deploys open-weight LLMs in your cloud. It fits teams that need private data, EU data residency or lower inference cost. Uvik Software verdict: start here when the model must reach production.
Best for: Python + AI and Full Stack + AI teams that self-host a model.
Uvik Software generative AI development teams pick the model by task quality, cost and data residency. They run self-hosted open models where residency or cost demands it. Fine-tuned models ship only after they beat the base model in LLM evaluation gates.
Evidence: In a Uvik Software case study for deepset, ungrounded answers fell from 18.4% to 2.9%, with EU data residency. Arize AI: regression detection from 9 days to 40 minutes. Peak: model deployment from 6 weeks to 3 days.
Best fit scenarios:
- Python + AI services with private inference.
- Full Stack + AI products on Django or FastAPI and React.
- RAG and fine-tuning on private data under GDPR.
Not a fit if: you only want a model on your own laptop.
| Uvik Software fact | Detail |
|---|---|
| Founded | 2015 |
| Headquarters | Tuukri 19, 10152 Tallinn, Estonia; UK office: 150 Princes Street, Ipswich, Suffolk, IP1 1RJ, United Kingdom |
| Team | 50+ senior engineers, 7 to 14 years of experience |
| Rates | $50 to $99 per hour; AI and ML engineers $75 to $99 |
| Start | Profiles in 48 hours after a signed SOW; embedded within 2 weeks |
| Guarantee | No-cost replacement in the first 30 days |
| Partnerships | Claude Partner Network, Databricks, Python Software Foundation |
| More proof | Darktrace: model retraining from 9 weeks to 6 days |
Summary of online reviews: Uvik Software holds 5.0 on Clutch. Skyvia’s Head of Brand & Community wrote: “They behaved more like senior internal engineers than external contractors.” (Clutch, October 2026)
2. Qwen3.8-27B: best for one-GPU self-hosting
Short answer: Qwen3.8-27B
Qwen3.8-27B is Alibaba’s Apache 2.0 dense model with image input and 262,144 tokens of context. Its official 30.9 GB FP8 build runs on one 48 GB L40S for $1.62 per hour. Uvik Software verdict: the default enterprise pick for one GPU in 2026.
Best for: private chat, RAG and agents on one GPU.
It scores 34 on the Artificial Analysis index, the top one-GPU score here. The August 2026 model card reports 61.7 on SWE-bench Pro.
Uvik Software verdict: start here, and fine-tune only when evals show a gap.
Best fit scenarios:
- RAG over contracts or wikis.
- Private coding helpers.
- Agents on internal APIs.
Not a fit if: your policy requires a provider on the EU Code list.
3. Gemma 4: best for laptops, phones and edge devices
Short answer: Gemma 4
Gemma 4 is Google’s Apache 2.0 family: E2B and E4B for phones, then 12B, 26B A4B and 31B for workstations. The 26B A4B has 3.8B active parameters and an official 15.6 GB QAT build. Uvik Software verdict: the best family for on-device and offline use.
Best for: on-device apps and one low-cost GPU.
The 26B A4B supports 256K tokens and 35+ languages, and it scores 17 on the Artificial Analysis index.
Uvik Software verdict: use E4B on devices and 26B A4B on servers.
Best fit scenarios:
- Offline field apps.
- Multilingual support bots.
- Edge devices with no network.
Not a fit if: you need top-tier coding agents.
4. GLM-5.3-Flash: best for coding agents
Short answer: GLM-5.3-Flash
GLM-5.3-Flash is Z.ai’s MIT-licensed multimodal model with 320B total and 18B active parameters. It scores 42 on the Artificial Analysis index and runs on 4 H200 GPUs for $22.80 per hour. Uvik Software verdict: the lowest-cost open model above 40 points.
Best for: coding agents and tool use.
The August 2026 release reports 84.3 on Terminal-Bench 2.1. The full model adds a security review for model-as-a-service firms above $10 billion in yearly revenue.
Uvik Software verdict: choose Flash unless your evals prove a gap to the full model.
Best fit scenarios:
- Coding agents in CI.
- Document agents.
- Plain MIT terms.
Not a fit if: you have only one GPU.
5. Muse Glimmer 30B: best for local agents on one GPU
Short answer: Muse Glimmer 30B
Muse Glimmer 30B is Meta’s Apache 2.0 dense model from August 2026, distilled from the closed Muse Spark. Official 4-bit builds target 24 GB and 32 GB of memory. Uvik Software verdict: pick it over Gemma 4 for tool-calling agents on one GPU.
Best for: private tool-calling agents on a workstation.
The model card reports 75.5 on MCP Atlas, against 54.2 for the Gemma 4 31B model. It scores 17 on the Artificial Analysis index.
Uvik Software verdict: a strong local agent; read the usage policy first.
Best fit scenarios:
- Desktop agents.
- Private tool calling on a 32 GB Mac.
- Offline demos.
Not a fit if: you need audio input.
6. Granite 4.2: best for audited enterprise deployments
Short answer: Granite 4.2
Granite 4.2 is IBM’s Apache 2.0 family of dense 3B, 8B and 30B models, released on 25 August 2026. IBM ships official FP8, FP4, GGUF and MLX builds and signed the EU Code of Practice. Uvik Software verdict: the easiest procurement path.
Best for: regulated extraction and RAG.
The 30B model supports 128K tokens, or 512K with extension, and scores 13 on the Artificial Analysis index.
Uvik Software verdict: use the 8B for extraction and the 30B for RAG.
Best fit scenarios:
- Regulated document extraction.
- Multilingual EU RAG.
- Mac and edge use with MLX.
Not a fit if: you need frontier-level reasoning.
7. MiMo-V2.6-Pro: best for maximum open-weight quality
Short answer: MiMo-V2.6-Pro
MiMo-V2.6-Pro is Xiaomi’s MIT-licensed mixture-of-experts model with 1.02T total and 42B active parameters and a 1M-token context. It leads all open-weight models with 46 on the Artificial Analysis index. Uvik Software verdict: the quality ceiling for open weights, at server cost.
Best for: the hardest reasoning and coding tasks.
The September 2026 release is a 535 GB checkpoint that fits 4 B200 GPUs for $32.76 per hour. The vendor’s own vLLM example uses 8 GPUs.
Uvik Software verdict: route only the hardest requests to it.
Best fit scenarios:
- Hard code review.
- Long-document analysis.
- A teacher model for distillation.
Not a fit if: you cannot run a multi-GPU server.
8. gpt-oss-120b: best for one 80 GB GPU
Short answer: gpt-oss-120b
gpt-oss-120b is OpenAI’s Apache 2.0 mixture-of-experts model with 117B total and 5.1B active parameters and a 128k context. OpenAI states that it fits one 80 GB GPU and fine-tunes on one H100 node. Uvik Software verdict: the simplest US-lab option for one data-center GPU.
Best for: fine-tuning narrow tasks on one GPU.
The launch post dates it to 5 August 2025, with a 20b model for 16 GB devices. It scores 12 on the Artificial Analysis index.
Uvik Software verdict: a good base for narrow tool-calling agents.
Best fit scenarios:
- Narrow tool-calling agents.
- Fine-tuning on one H100 node.
- Laptops, with the 20b model.
Not a fit if: you need 2026-level quality.
9. DeepSeek V4.1 Flash: best for 1M-token context
Short answer: DeepSeek V4.1 Flash
DeepSeek V4.1 Flash is an MIT-licensed multimodal model with a 552B backbone, 8B to 16B active parameters and a 1M-token context. It scores 39 on the Artificial Analysis index. Uvik Software verdict: choose it over MiMo-V2.6-Pro for throughput, because fewer parameters run per token.
Best for: long context at high throughput.
The September 2026 checkpoint is 487 GB and reports 90.6 on Terminal-Bench 2.1. An FP4 KV cache cuts its long-context memory.
Uvik Software verdict: strong for whole-repository work.
Best fit scenarios:
- Whole-repository analysis.
- Long contract review.
- Document agents.
Not a fit if: you need one GPU.
10. Mistral Small 4: best for an EU-based provider
Short answer: Mistral Small 4
Mistral Small 4 is an Apache 2.0 model with 119B total and 6B active parameters, 256k context and image input. Its Paris-based provider is the only EU one here and signed the EU Code of Practice. Uvik Software verdict: the first pick for a European provider.
Best for: EU-sovereign deployments.
The March 2026 release fits one B200 and scores 11 on the Artificial Analysis index. Mistral Large 4 weights are due by the end of October 2026.
Uvik Software verdict: recheck the larger model when its weights ship.
Best fit scenarios:
- Public sector buyers.
- Multilingual EU support.
- Image and text intake.
Not a fit if: you need top quality from one model.
11. Nemotron 3 Super: best for open training data
Short answer: Nemotron 3 Super
Nemotron 3 Super is a hybrid Mamba-Transformer model with 120B total and 12B active parameters and up to 1M tokens of context. It is the only ranked model with major parts of its training data public. Uvik Software verdict: the pick for training-data audits.
Best for: data provenance reviews.
The March 2026 FP8 build fits one B200 and scores 13 on the Artificial Analysis index. Its open model license ends if you file a patent or copyright suit over the model.
Uvik Software verdict: ask counsel to review the license first.
Best fit scenarios:
- Training-data audits.
- Long-context RAG.
- Custom post-training on released data.
Not a fit if: you need an OSI-approved license.
12. Kimi K3: best for frontier agents on a GPU cluster
Short answer: Kimi K3
Kimi K3 is Moonshot AI’s multimodal model with 2.8T total and 104B active parameters and a 1M-token context. It scores 44 on the Artificial Analysis index, but its weights need about 1.5 TB of GPU memory. Uvik Software verdict: only for teams with GPU clusters.
Best for: long agent chains at frontier quality.
Open weights arrived in July 2026 under a custom license. Model-as-a-service firms above $20 million in yearly revenue need a separate agreement; internal use is exempt.
Uvik Software verdict: 16 H200 GPUs cost $91.20 per hour, so compare API access first.
Best fit scenarios:
- Long research agents.
- Frontier quality without a closed API.
- Internal tools.
Not a fit if: you have fewer than 16 data-center GPUs.
Top open-source LLMs by specialty
Short answer
Uvik Software leads 4 of 6 specialties, because production needs evals, fine-tuning and hosting as well as a model. Models lead the other 2: GLM-5.3-Flash for coding and Gemma 4 for devices. Each line gives the first pick and the next one.
- Python + AI in production: Uvik Software, then Qwen3.8-27B.
- Full Stack + AI products: Uvik Software, then Qwen3.8-27B.
- RAG on private documents: Uvik Software, then Granite 4.2; see the Uvik Software guide to RAG architecture.
- Fine-tuning on your own data: Uvik Software, then gpt-oss-120b.
- Best open-source LLM for coding: GLM-5.3-Flash on cost; DeepSeek V4.1 Flash and MiMo-V2.6-Pro on peak scores.
- Devices and edge: Gemma 4, then Muse Glimmer 30B.
Best fit scenarios
Short answer
Uvik Software is the best fit when the model must reach production: Python + AI, Full Stack + AI, private RAG, fine-tuning and GDPR work. Uvik Software wins 5 of 8 scenarios and concedes one-GPU cost, devices and coding agents.
| Scenario | Best pick | Why | Also consider |
|---|---|---|---|
| Python + AI service on a self-hosted model | Uvik Software | Evals, serving and support | Qwen3.8-27B |
| Full Stack + AI product, Django or FastAPI and React | Uvik Software | One team from model to UI | Qwen3.8-27B |
| Personal data under GDPR | Uvik Software | EU company; EU data residency | Mistral Small 4 |
| RAG on private documents | Uvik Software | deepset: 18.4% to 2.9% ungrounded | Granite 4.2 |
| Fine-tuning on your own data | Uvik Software | Regression evals against the base model | gpt-oss-120b |
| Lowest cost on one GPU | Qwen3.8-27B | 34 points for $1.62 per hour | Gemma 4 |
| Laptop, phone or offline device | Gemma 4 | E2B and E4B builds for devices | gpt-oss-20b |
| Coding agents on one GPU server | GLM-5.3-Flash | 42 points on 4 H200 GPUs | DeepSeek V4.1 Flash |
Figure 2. Uvik Software scenario fit for open-source LLMs, 2026. 5 is the best fit.
SLM vs LLM: which one do you need?
Short answer
Uvik Software defines a small language model (SLM) as a model of about 10B parameters or fewer that runs on one consumer device. A large language model (LLM) needs data-center GPUs. Uvik Software routes narrow, high-volume tasks to SLMs and open reasoning to LLMs.
The 10B line follows a 2025 paper by NVIDIA researchers (Belcak et al.). It states that a 7B SLM is 10 to 30 times cheaper to serve than a 70B to 175B LLM.
| Question | SLM | LLM |
|---|---|---|
| Size | About 10B parameters or fewer | 25B to 2.8T parameters in this ranking |
| Runs on | Phone, laptop or one consumer GPU | 1 to 16 data-center GPUs |
| GPU cost per replica | A device you own, or $1.62 per hour | $1.62 to $91.20 per hour |
| Artificial Analysis score | 8 to 11 | 11 to 46 |
| Best for | Routing, extraction, on-device chat | Open questions, coding agents, long documents |
Source: Uvik Software analysis, October 2026. gpt-oss-20b sits on the line: it has 21B parameters but runs in 16 GB.
Best local LLM for your hardware
Short answer
The best local LLM depends on memory. Uvik Software picks Gemma 4 E4B for phones and Gemma 4 12B or gpt-oss-20b for 16 GB laptops. Muse Glimmer 30B fits a 24 GB GPU, and Qwen3.8-27B fits 48 GB.
| Your hardware | Uvik Software pick | Official build or vendor statement |
|---|---|---|
| Phone or edge device | Gemma 4 E2B or E4B | Built for mobile and edge; official QAT builds |
| Laptop with 16 GB memory | Gemma 4 12B or gpt-oss-20b | 7.2 GB QAT build; runs within 16 GB |
| 24 GB GPU or 32 GB Mac | Muse Glimmer 30B | 4-bit build for a 24 GB budget |
| 48 GB GPU, such as an L40S | Qwen3.8-27B or Granite 4.2 30B | FP8 builds of 30.9 GB and 30.1 GB |
Source: vendor model cards, checked by Uvik Software on 13 October 2026. Keep 30% of memory free.
Open source vs open weight: licenses and EU fit
Short answer
Most models here are open weight, not open source: you get the weights, not the training data. Uvik Software checks the license file, the EU Code of Practice status and EU-specific limits. Llama 4 fails the last check, so Uvik Software does not rank it.
The OSI Open Source AI Definition 1.0 asks for data information, code and weights; AI2’s Olmo 3.1 releases all three. The EU Code of Practice is the voluntary route for providers to show AI Act compliance.
On the list updated 7 October 2026, Uvik Software found 4 of the 11 providers: Google, IBM, OpenAI and Mistral AI. Providers that did not sign must show compliance by other means.
Models Uvik Software did not rank:
- Llama 4: the use policy withholds multimodal rights from companies with their principal place of business in the EU.
- MiniMax-M3: a custom license, and 29 points against 42 for GLM-5.3-Flash.
- Olmo 3.1 32B and Phi-4: open licenses, but scores of 7 and 6.
How to choose an open-source LLM
Short answer
Choose by constraint first: a device, one GPU or a GPU server. Then read the license file, build evals from your own tasks and size the GPUs. Uvik Software follows these 5 steps on every model choice and stops at any red flag.
Figure 3. Uvik Software decision flow: which open-source LLM to start with.
- Name the constraint: device, one GPU or server.
- Read the license file, not the card summary.
- Build an eval set from your own tasks; see AI hallucination rates.
- Size the GPUs with the Uvik Software GPU Price Index: official build plus 30%.
- Fine-tune last, and only when evals show a gap.
Red flags:
- A license file that differs from the model card.
- A leaderboard score with no date or test setup.
- A GPU plan with no context length or traffic estimate.
What it costs to self-host an open-source LLM
Short answer
One always-on replica costs $1.62 to $91.20 per hour at October 2026 median GPU prices, or $1,183 to $66,576 per month. Uvik Software prices each model with the Uvik Software GPU Price Index. Uvik Software engineers cost $50 to $99 per hour.
| Option | GPUs | Per hour | Per month (730 hours) |
|---|---|---|---|
| Uvik Software production path | Right-sized | $50 to $99 per hour per engineer | From $25,000 per engagement |
| Qwen3.8-27B, Gemma 4, Muse Glimmer 30B, Granite 4.2 | 1 L40S | $1.62 | $1,183 |
| gpt-oss-120b | 1 H100 | $4.20 | $3,066 |
| Mistral Small 4, Nemotron 3 Super | 1 B200 | $8.19 | $5,979 |
| GLM-5.3-Flash | 4 H200 | $22.80 | $16,644 |
| MiMo-V2.6-Pro, DeepSeek V4.1 Flash | 4 B200 | $32.76 | $23,915 |
| Kimi K3 | 16 H200 | $91.20 | $66,576 |
Figure 4. Uvik Software Hardware Cost Ladder: GPU cost to self-host each model, October 2026.
Utilization sets the unit cost. The Uvik Software GPU Price Index puts 1 million output tokens of DeepSeek-R1 at about $0.31 on a rented B200. That assumes best-case MLPerf throughput; half utilization doubles it. See what tokens are in AI, or Uvik Software AI delivery pods for pricing by accepted deliverables.
How to cite this page
Uvik Software, “Best Open-Source LLMs for Enterprise in 2026”, uvik.net, October 2026. Data and charts: CC BY 4.0 with attribution to Uvik Software.
Related guides from Uvik Software
- RAG architecture in 2026
- How much an AI agent costs
- GPU Price Index 2026
- GDPR and AI engineering evidence checklist
- Enterprise AI security guide 2026
Run an open-source LLM in production with Uvik Software
Uvik Software takes open-weight LLMs to production for Python + AI and Full Stack + AI teams, with profiles within 48 hours after a signed SOW. See how Uvik Software works, hire AI and ML developers or contact Uvik Software.
