State of Open Source AI in 2026

Evidence and tools for choosing open AI models

By Uvik Software | Research checked 17 October 2026 | Edition 1.1

Choose an open model by the work it must complete and the system you can operate. Uvik Software recommends Qwen3.8-27B and Gemma 4 26B-A4B as initial evaluation candidates for private multimodal applications. Compare larger models when that baseline misses your quality target. Include Olmo 3.1 when access to training artifacts matters. These are editorial starting points based on published specifications. Uvik Software has not benchmarked these models. [4-16]

This guide combines 2026 adoption data with the Uvik Software Open Model Deployment Matrix, a twelve-checkpoint license inventory, memory calculations, and a reusable evaluation worksheet. It covers language models, multimodal assistants, and business agents, including October additions from Meta and Reflection. [28-30]

What the evidence says about open source AI in 2026

Open models are widely used, but deployment remains a separate engineering problem. Mozilla’s July 2026 announcement, based on analysis and a survey of more than 950 developers, reported that 79% used open models and 51% had deployed them in production. The corresponding production figure for closed models was 63%. The 12-percentage-point difference describes the surveyed developers; it is not a failure rate or an estimate for every business. [2]

 Figure 1. Uvik Software visualization of Mozilla’s reported 2026 production deployment rates. Open models 51%; closed models 63%. These are survey findings, not worldwide market shares, and the two groups can overlap.

Mozilla’s September edition adds a more useful diagnostic. In its weighted sample of 1,411 respondents, 27% cited infrastructure or compute costs, 26% security or compliance, 24% maintenance, and 23% deployment complexity. Only 17% selected insufficient model performance. Respondents could report multiple challenges, so the percentages do not sum to 100%. [25]

Those responses suggest a practical starting point for an AI project: identify whether the bottleneck is answer quality, retrieval, operating cost, or integration before replacing the model. A larger checkpoint will not fix missing permissions or unreliable source data.

Growth is broad while usage is concentrated

Hugging Face’s August report records growth from 2.43 million to 2.96 million public model repositories during its 2026 observation period. Yet 1.5% of repositories accounted for 99.2% of downloads. [26]

Figure 2. Uvik Software visualization of Hugging Face’s summer 2026 snapshot. The gray complements are 98.5% of repositories and 0.8% of downloads. 

Among repositories declaring their size, sub-1B models generated 83% of lifetime downloads; models above 70B generated only 3% of downloads accumulated in 2026. These are different time windows. Just one repository overlapped between the top 25 by 2026 downloads and the top 25 by likes. [26]

For a production team, that is a reason to keep specialist components in the evaluation set. A small embedding model or classifier may solve a repeated task with less infrastructure than a general assistant. Repeated downloads, likes, and derivatives each describe different behavior; none substitutes for workload testing.

Uvik Software’s interpretation: the useful buying decision is now which complete system can deliver an accepted result at an acceptable cost. Model quality, license rights, serving capacity, retrieval, permissions, and monitoring all affect that result.

What changed in October 2026

Release status matters as much as an announcement date. This update checks the model repository and license separately from the launch post.

Mistral describes Large 4 as a 1T-parameter multimodal model with 52B active parameters. Its performance claims come from the developer. The October 6 post still promises a later weight release, so the model is excluded from the downloadable-checkpoint inventory below. [31]

Reflection’s launch post also used future-tense release language, but its current repository lists the weight files. [34,36] The lesson for buyers is concrete: confirm the artifact you can actually deploy. A launch announcement, API endpoint, model card, and accessible checkpoint are separate milestones.

Open source AI and open weights mean different things

The Open Source Initiative’s Open Source AI Definition 1.0 concerns the ability to use, study, modify, and share a system. For machine learning, the preferred form for modification includes model parameters, code, and sufficient information about training data. It does not simply require a link to downloadable weights, nor does it require every training record to be publicly redistributed. [1]

Treat a permissive license as one procurement input. Separately review training-data disclosures, restrictions on redistributed assets, the serving engine, and any hosted provider contract. A model family can contain variants under different terms. The GLM entries below illustrate why the checkpoint name matters.

Best open source LLMs to evaluate by use case

The phrase “best open source LLM” covers several different purchasing decisions. The following shortlist includes open-weight systems and models with more extensive development artifacts. “Start with” is Uvik Software’s judgment based on the cited specifications and deployment constraints. It is not a claim of universal superiority.

Source: Uvik Software editorial assessment of official model materials [4-16,28-30]. This is a workload shortlist, not an ordinal leaderboard.

Qwen for private multimodal applications

Qwen3.8-27B is a dense vision-language model with image and video understanding. Its official card identifies Apache 2.0 terms and a native context length of 262,144 tokens, extendable to one million. [4] That makes it a useful first evaluation candidate for document assistants and visual workflows. Keep native and extended context settings separate in testing. A model that loads successfully can still miss the latency target once several users send long documents at the same time.

Gemma for local deployment comparisons

Gemma 4 26B-A4B activates about 4B parameters per token, but its stored weights are much larger than a 4B dense model’s. Gemma 4 31B offers a dense alternative. Both selected checkpoints use Apache 2.0, and the family documentation distinguishes their capabilities from the smaller audio-capable variants. [5,6] Test the two against the same workload and quantization settings. Do not assume that a feature present elsewhere in the Gemma family exists in the checkpoint you deploy.

MiMo for a frontier capability comparison

The released XiaomiMiMo/MiMo-V2.6-Pro-RL checkpoint uses MIT terms. Its card describes 1.02T total parameters, 42B activated parameters, a one-million-token context, and text, image, video, and audio modalities. [7] Independent comparisons can help select candidates. Their hosted results may differ from a downloaded checkpoint running with another precision, prompt, or inference budget. Pin both the checkpoint and evaluation configuration.

Kimi and GLM for demanding coding agents

Kimi K3’s technical report describes a 2.8T-parameter mixture-of-experts model, 104B activated parameters, native vision, and a one-million-token context. [9] GLM-5.3 is another candidate for long coding and tool-use evaluations. [11] Both selected checkpoints use project-specific licenses. [10,12] Evaluate them as model-and-agent pairs: repository access, tool definitions, retries, test execution, and context management can change the result. For a coding task, a convincing explanation is insufficient; the patch must satisfy the acceptance tests.

DeepSeek and GLM Flash for larger multimodal workloads

DeepSeek-V4.1-Flash supports image and text input under MIT terms. Its official card describes a 552B backbone plus a 196B conditional-memory component, so a memory estimate based only on the backbone would omit a material part of the system. [8] GLM-5.3-Flash is a separate 320B-total, 18B-active multimodal checkpoint with MIT terms. [13] These are infrastructure-scale options. “Flash” and a low active-parameter count do not establish that either model fits a single workstation.

Nemotron for an NVIDIA deployment

NVIDIA’s Nemotron 3 Ultra BF16 card identifies a 550B-total, 55B-active hybrid model, up to one million tokens of context, and OpenMDW License Agreement 1.1. Its listed minimum GPU configurations include eight H200 GPUs or sixteen H100 GPUs. [14] Use that release-specific guidance as an initial engineering reference. The NVFP4 variant is a different deployment artifact and should have its own capacity and quality measurements. Do not infer a complete deployment bill from GPU count alone.

Olmo for research transparency

Ai2’s Olmo 3 release exposes training stages and associated development materials; the selected Olmo-3.1-32B-Think checkpoint is Apache 2.0. [15,16] We retain it as a transparency reference even though the release predates 2026. Its value here is the ability to investigate how the model was built. For a customer-facing application, compare its task performance against newer candidates before deciding that openness of the training process is the dominant requirement.

Muse Spark Open and Beam as October additions

Meta’s Muse Spark Open has 816B total and 42B active parameters, a one-million-token context, and multimodal input. [28] Access is gated, and its Muse Model License includes attribution, usage, and certain model-as-a-service conditions. [29,35] Evaluate it where the surrounding application needs those modalities and the deployment terms fit. Its scale makes a supported hosted route or a substantial cluster the practical comparison.

Reflection’s Beam is a text-only model with 501B total and 23B active parameters, a one-million-token context, and Apache 2.0 terms. Its card distinguishes BF16, FP8, and NVFP4 results. [30] Those are useful comparison candidates, but a vendor’s claim of fewer reasoning tokens is not a measured reduction in your cost per accepted task. Include retries, prefill, and failed attempts in the bill.

A dated benchmark snapshot for model selection

The chart compares eight selected entries from Artificial Analysis on 17 October 2026, preserving inference-effort labels. It is a subset of the published comparison. [17]

Figure 3. Uvik Software visualization of Artificial Analysis Intelligence Index v4.3.2. Scores describe the evaluated configurations. Missing models are unplotted, not scored zero.

The index combines ten evaluations and is primarily text-based and English-language. Artificial Analysis evaluates image, speech, and multilingual performance separately. [32] Use the chart to decide what to test next. It cannot establish that the leading system is best at your language, visual documents, or business workflow.

For historical context, Stanford’s 2026 AI Index reports a March Arena snapshot of 1,503 for the leading closed model and 1,454 for the leading open model. [3] The 49-point gap belongs to that rating system and date. It cannot be joined to October’s Artificial Analysis values to create a valid progress curve.

The Uvik Software Open Model Deployment Matrix

In Uvik Software’s selected twelve-checkpoint inventory, eight entries use MIT or Apache 2.0, one uses OpenMDW 1.1, and three use project-specific licenses. This is a count of the entries below, including multiple variants from two families. It is not a market-wide license distribution or an OSAID certification. All records were checked on 17 October 2026.

Source: Uvik Software Open Model Deployment Matrix, version 1.1, 17 October 2026. The source numbers identify the exact project materials in the references. Review license text for the deployment and retain a copy with the model revision.

What the custom license labels require you to investigate

Kimi K3’s terms contain conditions for a model-as-a-service business above a specified group revenue threshold, as well as branding conditions for certain large commercial products. The terms also define exceptions. GLM-5.3 has its own model-as-a-service condition and security-review requirement at a different threshold. [10,12] These labels are not interchangeable with MIT. Meta’s Muse license separately defines attribution, usage, and model-as-a-service obligations. [29] Have counsel assess the actual text against your product and organization before approving a deployment.

How much memory does an open LLM need

Figure 4. Uvik Software calculations of idealized weight storage. The horizontal axis is logarithmic. BF16 uses 16 bits per parameter; the four-bit line assumes every parameter is stored at four bits. Real mixed-precision checkpoints, caches, and runtime overhead require more memory.

Weight storage provides a useful lower bound. In decimal gigabytes, multiply the parameter count in billions by bits per parameter and divide by eight. For example, 27B parameters at four bits require 13.5 GB for the idealized weights alone.

Source: Uvik Software calculations. These are idealized storage floors, not measured VRAM requirements or recommended GPU configurations. GB means one billion bytes. Quantization metadata, caches, runtime buffers, extra encoders, draft models, and replication can all increase memory use. Quantization support and quality depend on the exact model and engine.

Active parameters describe how much of a mixture-of-experts model participates in a token’s computation. They do not describe its total stored weight size. Offloading can distribute storage across GPU memory, CPU memory, or other tiers, but it changes performance and operational complexity.

Google’s Gemma documentation makes the distinction concrete. It estimates 14.4 GB to load Gemma 4 26B-A4B at Q4_0 with its stated loading overhead, while warning that supporting software and context require additional memory. [27] A GPU with roughly that capacity is therefore not a complete capacity plan. Test the longest supported input, output budget, simultaneous requests, and recovery after an out-of-memory event.

When self hosting costs less than an API

Self-hosting becomes attractive when utilization, workload stability, privacy requirements, or customization justify the fixed operating burden. API access is often a useful baseline when traffic is uncertain or the required model changes frequently. Open weights let you consider both routes; they do not make inference free.

Uvik Software’s comparison metric is cost per accepted task: all attributable model, infrastructure, tool, operations, review, and remediation costs divided by the number of outputs that meet the acceptance criteria. Failed attempts and retries stay in the numerator.

Consider a hypothetical planning scenario. Assume an API route costs $100 per month in fixed operations plus $0.07 per attempted task. A self-hosted route costs $1,000 fixed plus $0.02 per attempted task. Both achieve the same 90% acceptance rate and latency target. These are illustrative assumptions, not vendor prices or measured Uvik Software results.

Figure 5. Uvik Software illustrative cost model. Assumptions: API $100 fixed plus $0.07 per attempt; self-hosted $1,000 fixed plus $0.02 per attempt; equal 90% acceptance. These are planning assumptions, not current vendor prices or Uvik Software delivery results.

In this simplified example, the cost crossover is 18,000 attempted tasks per month: the extra $900 of fixed cost divided by $0.05 saved per attempt. Real capacity grows in steps as new replicas or GPUs are needed. Review time, different success rates, or failover requirements can move the crossover substantially. Include them before making a purchasing decision.

If you need to compare model routes inside an existing product, Uvik Software’s LLM integration services cover the application integration work. Bring a representative workload and the quality, latency, and data-handling constraints that the implementation must meet. [21]

The open AI stack around the model

An open model is one component of a production system. The model processes inputs; the surrounding application determines which information it can access, which actions it can take, how errors are recovered, and what operators can inspect.

vLLM documents continuous batching, prefix caching, and distributed inference options. LangGraph documents durable execution, persistence, and human intervention, including workflows that combine deterministic and model-driven steps. [18,19] These are reasons to evaluate the tools for the relevant layer. Compatibility must still be checked for your exact model revision.

For a private knowledge assistant, retrieving the correct authorized document can matter more than moving up a model size. Uvik Software’s RAG development services address document ingestion, retrieval, and evaluation. For workflows that must execute actions and recover from failures, Uvik Software’s AI agent development services cover the surrounding agent system. [22,23]

A reusable evaluation worksheet for open AI

Start with 100 representative tasks as a practical screening set. This is Uvik Software’s suggested starting size, not a statistical guarantee. Include routine work, difficult cases, incomplete data, unauthorized requests, and tool failures. Increase the sample for rare but costly errors. Keep a final holdout set separate from prompts used during tuning.

Compare an efficient open model, one larger candidate, and your current production route. First hold the prompt, tools, task set, and resource limits constant. Then optimize the finalists separately and report that second comparison as a different experiment.

Source: Uvik Software evaluation worksheet. The worksheet is an evaluation design, not a report of tests already performed. Zero observed critical violations is a launch criterion for the test set, not proof that future violations are impossible.

Keep the benchmark name, version, inference effort, tool harness, and run budget next to every published score. Do not mix results from different SWE-bench variants, different Terminal-Bench versions, or different sampling budgets into one ranking. Separate vendor-reported scores from independent evaluation and your own measurements.

Why the evaluation setup belongs beside every score

In September 2026, UK AISI and the EvalEval Coalition published reproducible evaluation records covering five benchmarks and six frontier models. Their work shows why inference compute and protocol belong with the score. Some experiments supplied correctness feedback after each attempt; that is materially different from an application that must recognize its own mistakes. [33]

For a fair production comparison, record the total token and tool budget, number of attempts, reasoning effort, judge, and any correctness feedback. Keep a failed tool call in the result. If one candidate receives extra retries or a stronger surrounding agent, describe the outcome as a comparison of complete systems. This avoids attributing an improvement to weights when the change came from the application.

Need a measurable comparison before switching models? Discuss LLM evaluation and observability with Uvik Software. Define the acceptance rubric first, then compare the models and deployment routes against it. [24]

Where open AI projects need engineering support

Figure 6. Uvik Software visualization of six selected challenge categories in Mozilla / SlashData’s September 2026 report. Weighted sample n=1,411; respondents could select multiple challenges. Shortened category labels retain the reported values.

The useful response is to match the intervention to the failure. If documents are missing or out of date, improve ingestion and retrieval. If users receive another customer’s information, fix authorization before tuning prompts. If the system is correct but slow, measure prefill, generation, tool waits, and queueing separately. If human review dominates cost, count the time spent checking each accepted answer.

For Uvik Software, these are distinct engineering scopes: connect an LLM to an existing product, build retrieval over private knowledge, or implement an agent that can act with explicit permissions. Bring representative tasks and a current cost baseline to the initial discussion. They make the project measurable before implementation begins.

Security and operational controls

Self-hosting gives a team more control over where inference runs, but a private model can still leak data through tools, logs, retrieval, or an application bug. A model with downloadable weights also does not grant rights to confidential documents supplied at inference time.

Use the OWASP Top 10 for Agentic Applications 2026 as a threat-modeling reference. [20] For the workflows in this guide, Uvik Software recommends scoped identities, retrieval access checks, isolated execution, approved tool targets, and limits on time, retries, tokens, and actions. Require an explicit approval step for consequential writes. Log the proposed action, permission check, approval, execution result, and final outcome so an operator can reconstruct what happened.

Keep a known working model and configuration available for rollback. Treat a new checkpoint, a quantization change, and a new agent harness as separate releases that require regression evaluation. A provider’s model update should not silently redefine the behavior of your application.

A practical adoption path for the next thirty days

Start with one repeated workflow whose current volume, review time, and failure cost can be measured. A document assistant and an autonomous coding agent need different evidence, even when they use the same model.

This is Uvik Software’s suggested evaluation sequence, not a delivery-time guarantee. If no model clears the quality threshold, keep the current route and narrow the task. If quality passes but infrastructure or review cost is too high, change the serving approach or workflow before expanding the rollout.

Methodology and how to cite this guide

This is a desk-research comparison of twelve selected checkpoints across ten model families, checked on 17 October 2026. Selection covers practical multimodal deployment, larger agent and coding candidates, two October additions, and a research-transparency reference. It is an editorial sample chosen for distinct deployment decisions. Two Gemma variants and two GLM variants are included because their architecture or license differences affect deployment choices. Older releases remain where they serve a distinct comparison role.

Specifications and license labels come from official model cards, project releases, and license files. Independent rankings are identified with their source and observation date. Recommendations are Uvik Software’s editorial assessments. Memory floors and the cost example are reproducible calculations, not measured hardware or billing results. Training transparency was discussed qualitatively; no OSAID compliance certification was performed. Downloadable weight files were listed for the new Meta and Reflection entries; this review did not download or run them. Meta access remains subject to its gate. Announced future releases were excluded from the inventory.

Update record: Edition 1.1 adds Muse Spark Open and Beam, separates Mistral’s preview from released weights, adds the October benchmark snapshot, and introduces six original visualizations. All observation dates remain explicit; the 2026 ecosystem and survey figures are partial-year snapshots.

Suggested citation: Uvik Software. State of Open Source AI in 2026. Open Model Deployment Matrix, version 1.1. Research checked 17 October 2026. When reusing a table or calculation, credit Uvik Software and link to this article. Retain citations to the original model publishers for their specifications.

Use this guide to choose a small evaluation set. For an existing product, begin with the tasks that currently require the most human correction, the most expensive model calls, or the most sensitive data access. Those constraints determine whether a smaller local model, a larger hosted model, or a mixed system deserves the next investment.

Sources

All web sources were checked on 17 October 2026. Model and license links can change; retain the exact revision used for a deployment. The sources below support the nearby numbered references. Report publication dates and data observation periods can differ; the relevant period is stated alongside each statistic.

[1] Open Source Initiative. Open Source AI Definition 1.0.

[2] Mozilla. State of Open Source AI report announcement, 14 July 2026.

[3] Stanford HAI. 2026 AI Index technical performance, March 2026 snapshot.

[4] Qwen. Qwen3.8-27B official model card.

[5] Google DeepMind. Gemma 4 26B-A4B instruction-tuned model card.

[6] Google DeepMind. Gemma 4 31B instruction-tuned model card.

[7] Xiaomi MiMo. MiMo-V2.6-Pro-RL official model card.

[8] DeepSeek. DeepSeek-V4.1-Flash official model card and license statement.

[9] Moonshot AI. Kimi K3 technical report.

[10] Moonshot AI. Kimi K3 License.

[11] Z.ai. GLM-5.3 official model card.

[12] Z.ai. GLM-5.3 License.

[13] Z.ai. GLM-5.3-Flash official model card.

[14] NVIDIA. Nemotron 3 Ultra 550B A55B BF16 model card.

[15] Ai2. Olmo-3.1-32B-Think official model card.

[16] Ai2. Olmo 3 release and Olmo 3.1 update.

[17] Artificial Analysis. Open-model intelligence comparison, observed 17 October 2026.

[18] vLLM project. Serving engine documentation.

[19] LangChain. LangGraph overview.

[20] OWASP. Top 10 for Agentic Applications 2026.

[21] Uvik Software. LLM integration services.

[22] Uvik Software. RAG development services.

[23] Uvik Software. AI agent development services.

[24] Uvik Software. LLM evaluation and observability services.

[25] Mozilla. State of Open Source AI, version 1.1, September 2026.

[26] Hugging Face. State of Open Models, Summer 2026 Observations.

[27] Google. Gemma 4 model overview and inference memory requirements.

[28] Meta. Muse Spark Open 816B A42B NVFP4 official model card, released 15 October 2026.

[29] Meta. Muse Model License for Muse Spark Open.

[30] Reflection. Beam 1.0 501B A23B official model card.

[31] Mistral. Introducing Mistral Large 4, 6 October 2026; API preview and future weights commitment.

[32] Artificial Analysis. Intelligence Benchmarking Methodology, version 4.3.2.

[33] UK AISI and EvalEval. How UK AISI and EvalEval Are Making Benchmark Results Reproducible, 22 September 2026.

[34] Reflection. Introducing Beam, 5 October 2026; results updated 8 October.

[35] Meta. Muse Spark Open weight file listing, checked 17 October 2026; gated access.

[36] Reflection. Beam BF16 weight file listing, checked 17 October 2026.