Best Open Source LLMs 2026: Open Source AI Models Ranked and Tracked

Updated October 26, 2026: This tracker now opens with a ranked list of the best open source LLMs by independent benchmark score, adds a guide to choosing an open-weight model, and logs every major open release from July through October 2026, including Meta’s return to open weights with Muse Spark Open. The original analysis below is unchanged.

The Pulse

On April 8, 2026, Meta shipped Muse Spark, the first model from its newly formed Meta Superintelligence Labs, as a fully closed, proprietary system available only through Meta’s own apps. No downloadable weights, no open license, no Hugging Face release. For the company that spent three years positioning itself as the open-source alternative to OpenAI and Anthropic, that was a reversal significant enough that multiple tech outlets independently described it as the end of an era. [The New Stack Meta abandons open-source Llama for Muse Spark]

Meta has not discontinued Llama, and existing Llama models remain downloadable, but the company’s frontier development effort and its heaviest investment have moved to the closed Muse Spark line. That leaves a meaningful gap at the top of the open-weight leaderboard that Chinese labs, and a resurgent Mistral and Google, have moved quickly to fill.

The result is an open-source AI landscape in mid-2026 that looks structurally different from a year earlier: Apache 2.0 has become the dominant permissive license among the labs still releasing openly, several open models now genuinely compete with closed frontier systems on real benchmarks, and the single most prominent original champion of open-weight AI has stepped back from the category entirely at the frontier level.

October 2026 update: Meta partly reversed course on October 15, 2026, releasing the weights for Muse Spark Open, an 816-billion-parameter open-weight version of Muse Spark, under a custom Muse Model License with what Meta calls limited commercial terms. [Meta AI Muse Spark Open announcement] It is now the highest-scoring US open-weight model on the Artificial Analysis Intelligence Index, but it ranks only ninth among open models. The top of the open-weight leaderboard belongs to Xiaomi, DeepSeek, Z.ai, StepFun and Moonshot.

Best Open Source LLMs in 2026: The Ranked List (October Update)

The table below ranks open-weight models, meaning models whose weights can be downloaded, fine-tuned and self-hosted, by the Artificial Analysis Intelligence Index v4.3.2. The index is an independent composite of 10 evaluations covering agentic work, coding, reasoning, knowledge and long-context tasks. Scores were read on October 26, 2026. Each model appears once, at its best-scoring setting, and models with equal scores share a rank. License and parameter figures come from each model’s Artificial Analysis page and official model card.

Rank	Model (lab)	Intelligence Index	License	Parameters (total / active)	Context	Weights released
1	MiMo-V2.6-Pro (Xiaomi)	46	MIT (commercial use allowed)	1.0T / 42B	1M	Sep 2026
=2	DeepSeek V4.1 Flash 1017 (DeepSeek)	45	MIT (commercial use allowed)	552B / 16B	1M	Oct 17, 2026
=2	GLM-5.3 (Z.ai)	45	GLM-5.3 License (commercial use with restrictions)	753B / 40B	1M	Aug 28, 2026
=4	Step 5 Preview (StepFun)	44	StepFun Model License 1.0 (commercial use with restrictions)	600B / 27B	1M	Oct 2026
=4	Kimi K3 (Moonshot AI)	44	Kimi K3 License (commercial use with restrictions)	2.8T / 104B	1M	Jul 27, 2026
6	GLM-5.3-Flash (Z.ai)	42	MIT (commercial use allowed)	320B / 18B	1M	Aug 2026
=7	Qwen3.8-Flash-Next (Alibaba)	40	Qwen Community License 1.0 (commercial use with restrictions)	180B / 6B	256K	Aug 2026
=7	Qwen3.8-2.4T-A95B (Alibaba)	40	Qwen3.8-Max License (commercial use with restrictions)	2.4T / 95B	262K	Aug 2026
9	Muse Spark Open (Meta)	39	Muse Model License (commercial use with restrictions)	816B / 42B	1M	Oct 15, 2026
10	MiMo-V2.6-Flash (Xiaomi)	38	MIT (commercial use allowed)	309B / 15B	1M	Sep 2026
11	DeepSeek V4 Pro 0813 (DeepSeek)	36	MIT (commercial use allowed)	1.6T / 49B	1M	Aug 2026
12	Qwen3.8-27B (Alibaba)	34	Apache 2.0 (commercial use allowed)	27B dense	256K	Aug 2026
13	Beam 1.0 (Reflection AI)	30	Apache 2.0 (commercial use allowed)	501B / 23B	1M	Oct 16, 2026

Two readings matter more than the order itself. The first is the gap: the best open-weight score, 46, trails the best proprietary model, Claude Opus 5.5 at 58, by 12 points. The second is the pending leader. Moonshot’s Kimi K3.1 Pro launched on October 23 at 55, fourth overall, but its weights had not been published at the time of writing. Artificial Analysis says it will become the highest-scoring open-weights model once they are, which would cut the open-closed gap to 3 points. [Artificial Analysis Kimi K3.1 Pro benchmarks]

Below the top tier the picture changes. Alibaba’s Qwen3.8-27B scores 34 with just 27 billion parameters under Apache 2.0, well ahead of Google’s Gemma 4 31B (15) and OpenAI’s gpt-oss-120b (12), both also Apache 2.0. Geography is just as lopsided: nine of the top ten open models come from China-based labs. Meta’s Muse Spark Open (39) is the highest-scoring US entry, followed by Reflection AI’s Beam (30).

Core Significance

Why it matters:

• Licensing terms, not just capability, now genuinely separate the open-weight field into distinct usability tiers:  OSI-approved permissive licenses like Apache 2.0 and MIT, used by DeepSeek, Mistral, Gemma, and the open Qwen line, allow unrestricted commercial use and redistribution. Source-available licenses like Meta’s Llama Community License permit commercial use but carry field-of-use restrictions, including a monthly active user cap that excludes the very largest companies without a separate agreement, a distinction the Open Source Initiative has publicly disputed as inconsistent with the term open source. [TechJack Solutions best open-source AI models 2026]
• Chinese labs have become the default reference point for permissively licensed frontier-adjacent open weights:  Alibaba’s Qwen line reportedly captured more than half of global open-source model downloads by April 2026, while DeepSeek’s V4 series, released as a public preview on April 24, 2026 under an MIT license, offers a 1 million token context window trained on more than 32 trillion tokens, positioning both labs as the practical default for developers who want a permissively licensed model without negotiating Meta’s usage restrictions. [AIUnpacking open source AI models 2026 complete guide]
• Meta’s pivot to closed frontier development did not happen in isolation, and every other major lab’s 2026 release cadence should be read against it:  With Alibaba separately flipping its own top-tier Qwen3.7 Max model closed and API-only in the same window, and with Google, Mistral, and a cluster of Chinese labs including Z.ai’s GLM series and Xiaomi’s MiMo continuing to ship openly, the open-weight field in 2026 is simultaneously losing its most prominent Western champion and gaining new entrants willing to compete on openness specifically. [CodersEra open-source LLMs landscape May 2026]

Deep Context: Why Meta Walked Away from the Frontier of Open Weights

Meta’s public rationale for Muse Spark’s closed release centers on cost and competitive positioning rather than a rejection of open source as a concept. The company’s 2026 AI capital expenditure guidance sits between 115 and 135 billion dollars, nearly double the prior year, following a roughly 14.3 billion dollar deal in mid-2025 that brought Scale AI co-founder Alexandr Wang to Meta as chief AI officer leading the newly formed Meta Superintelligence Labs. [Miraflow Meta ended Llama built Muse Spark changes everything]

At that level of spending per training run, giving away the resulting model’s weights for free became a harder economic case to justify internally, particularly after Meta’s own Llama 4 release in 2025 was widely seen as falling short of closed frontier competitors despite the company’s compute advantage. Muse Spark represents a bet that closed, tightly integrated distribution through Meta’s own consumer apps, reaching billions of existing users directly, generates more value than open community adoption did.

As covered in our OpenAI vs Anthropic enterprise comparison, the broader 2026 pattern across frontier labs is that closed, tightly governed model access increasingly correlates with the heaviest compute investment, while openness has become more associated with labs using open releases as a distribution and ecosystem strategy rather than a monetization one. Meta’s pivot is the clearest single data point yet that even a company that built its AI brand on openness will abandon it at the frontier once the economics shift.

The Gap Meta Left Is Being Filled Fastest by Non-US Labs

Hugging Face download data cited by industry analysts shows Chinese open-weight labs, primarily DeepSeek and Alibaba’s Qwen, accounting for roughly 41% of downloads by late 2025, even before Meta’s April 2026 pivot removed the most prominent Western open-weight alternative from frontier contention entirely. [Pebblous Meta drops open weights Muse Spark EU AI Act]

That timing means the shift toward Chinese-led open weights was already underway before Muse Spark launched, and Meta’s move away from open frontier development has likely accelerated a trend rather than started one. For any enterprise or developer whose AI infrastructure strategy assumed a steady stream of frontier-competitive open weights from a US lab, 2026 has been the year that assumption stopped holding.

Data Insights

By the Numbers:

Figures below are drawn from named industry outlets and are attributed individually given how quickly benchmark and download figures shift in this category.

• Meta’s Llama ecosystem had reached approximately 1.2 billion cumulative downloads by early 2026, averaging roughly 1 million downloads per day, before the company’s frontier strategy shifted toward the closed Muse Spark line:  That download base illustrates the scale of the existing developer ecosystem Meta is asking to wait for an uncertain future open release, since the company has stated future Muse-family models may eventually ship with open weights but has not committed to a timeline.[StartupHub Zuckerberg dual-track AI Llama 5 open Muse Spark closed]
• Meta’s 2026 AI capital expenditure guidance of 115 to 135 billion dollars is nearly double its prior-year spending, a scale that reframes the economics of open releases industry-wide:  At that level of per-model training cost, the calculation facing any lab, not just Meta, over whether to release weights openly increasingly depends on whether open distribution itself, rather than direct monetization, is the more valuable strategic outcome for that specific company’s position.
• Google’s Gemma 4, released under Apache 2.0 in April 2026, and DeepSeek’s V4 series, released under MIT the same month, illustrate that fully permissive licensing remains the default choice for labs still competing on openness:  Both releases avoid the field-of-use restrictions present in Meta’s Llama Community License, reinforcing that Apache 2.0 and MIT have become the practical industry standard for any lab using open weights as a genuine adoption strategy rather than a limited or conditional release.
• The best open-weight model trailed the best proprietary model by 12 points on the Artificial Analysis Intelligence Index in late October 2026, 46 versus 58: Artificial Analysis expects that gap to fall to 3 points once Moonshot publishes the weights for Kimi K3.1 Pro, which scored 55 at launch on October 23, 2026. For buyers, that means the capability case for waiting on closed models is narrowing faster than the licensing case is being resolved.

Table 1: Major 2026 Open-Weight Model Releases and Licensing

Model	Lab	Release window	License	Notable spec
DeepSeek V4	DeepSeek	Public preview, April 24, 2026	MIT	1M token context, trained on 32T-plus tokens
Gemma 4	Google	April 2026	Apache 2.0	Native audio support, 256K context
Qwen 3.5 and successors	Alibaba	Rolling releases through 2026	Apache 2.0, open line	Reported over 50% of global open-source downloads
Mistral Large 3	Mistral	2026	Apache 2.0	Sparse MoE, European data residency
Muse Spark	Meta Superintelligence Labs	April 8, 2026	Closed, proprietary	First Meta model with no open weights since 2023
Inkling	Thinking Machines Lab	July 2026	Apache 2.0	975B total / 41B active parameters, 1M context
Kimi K3	Moonshot AI	API July 16; weights July 27, 2026	Kimi K3 License (custom)	2.8T total / 104B active, 1M context; Moonshot calls it the first open 3T-class model
Qwen3.8-2.4T-A95B	Alibaba	Weights August 2026	Qwen3.8-Max License (custom)	2.4T / 95B active; first Qwen Max-class model with downloadable weights
Qwen3.8-27B	Alibaba	August 2026	Apache 2.0	27B dense, natively multimodal, 262K context extensible to 1M
GLM-5.3-Flash	Z.ai	August 2026	MIT	320B / 18B active, 1M context
GLM-5.3	Z.ai	API August 14; weights August 28, 2026	GLM-5.3 License (custom)	753B / 40B active; weights held about two weeks for safety evaluation
DeepSeek V4 Pro 0813	DeepSeek	August 2026	MIT	1.6T / 49B active, 1M context
DeepSeek V4.1 Flash (1017 update)	DeepSeek	September 10, 2026; update October 17	MIT	552B / 16B active in decode, 1M context; DeepSeek is phasing out V4-Pro in its API
MiMo-V2.6-Pro and MiMo-V2.6-Flash	Xiaomi	Weights September 21 to 22, 2026	MIT	Pro 1.02T / 42B active, Flash 309B / 15B active, 1M context; top open-weight model on Artificial Analysis (46)
Step 5 Preview	StepFun	API September; weights October 2026	StepFun Model License 1.0 (Apache-style, extra conditions)	600B / 27B active, 1M context
Beam 1.0	Reflection AI	Announced October 5; weights October 16, 2026	Apache 2.0	501B / 23B active, 1M context
Muse Spark Open	Meta Superintelligence Labs	October 15, 2026	Muse Model License (limited commercial terms)	816B / 42B active, 1M context; runs on two M5 Ultra Mac Studios
Nemotron 3.5 Super VL	NVIDIA	October 2026	OpenMDW-1.1	120B / 12B active, 1M context
Kimi K3.1 Pro	Moonshot AI	API October 23, 2026	Weights not yet published	Scores 55 on Artificial Analysis, fourth overall; expected to lead open models once weights ship
Mistral Large 4 Preview	Mistral	October 2026	API preview; listed as proprietary until weights publish	524K context

Rows from Inkling onward were added in October 2026 and cover July through October releases. License, parameter and context figures are taken from each model’s Artificial Analysis page and official sources: Kimi K3, Qwen3.8-Max, Qwen3.8-27B, GLM-5.3, DeepSeek V4.1 Flash 1017, MiMo-V2.6, Step 5 Preview, Beam and Muse Spark Open.

Table 2: Open Versus Closed Strategy by Lab, Mid-2026

Lab	Frontier strategy	Open-weight commitment
Meta	Shifted to closed with Muse Spark	Existing Llama models remain available; future open releases uncertain
DeepSeek	Open by default	MIT license maintained across V4 series
Alibaba	Split strategy	Qwen open line continues; top Qwen3.7 Max model flipped closed and API-only
Google, Mistral	Open by default for mid-tier models	Apache 2.0 maintained for Gemma and Mistral open releases

October 2026 update to Table 2: Meta now publishes open weights again through Muse Spark Open, under a custom license rather than an OSI-approved one. Alibaba released Max-class weights for the first time with Qwen3.8-2.4T-A95B, also under a custom license, while keeping Qwen3.8-27B on Apache 2.0. DeepSeek and Xiaomi stayed on MIT. Two newer US labs entered the field with Apache 2.0 releases: Thinking Machines with Inkling and Reflection AI with Beam. Mistral’s Large 4 is so far an API preview without published weights.

The Business Case: What Meta’s Pivot Means for Enterprises Building on Open Weights

Enterprises with existing production deployments on Llama face no immediate disruption. Meta has stated existing Llama models remain available and will continue receiving some maintenance support, and the models already downloaded and deployed do not stop functioning because Meta’s frontier strategy has shifted. The practical risk is forward-looking: any roadmap that assumed continued frontier-competitive open releases from Meta specifically now needs a contingency plan.

The more defensible enterprise posture in mid-2026 is treating open-weight model selection the way most FinOps-mature organizations already treat cloud vendor selection, as a portfolio decision rather than a single-vendor commitment. DeepSeek’s MIT-licensed releases, Alibaba’s open Qwen line, and Google’s Apache 2.0 Gemma releases each offer genuinely permissive commercial terms without the field-of-use restrictions Meta’s Community License carries, giving enterprises real alternatives that did not require nearly the same diligence eighteen months ago.

As covered in our Hardware-as-a-Service report, the same portfolio logic that applies to GPU infrastructure procurement, avoiding single-vendor lock-in given how quickly hardware generations turn over, applies directly to open-weight model selection now that the open-weight leaderboard itself is turning over across labs and licenses faster than most enterprise procurement cycles can track.

Expert Nuance: Benchmark Competitiveness No Longer Tracks Neatly With License Permissiveness

A pattern worth flagging for anyone evaluating open models by license alone is that the most permissively licensed models are not automatically the weakest, nor are the most restricted the strongest. Independent comparison data shows Apache 2.0 and MIT licensed models, including entries from Qwen, DeepSeek, and the GLM family, now leading specific benchmark categories including coding and long-context tasks outright, not merely trailing acceptably behind more restricted alternatives.[ComputingForGeeks open source LLM comparison table 2026]

That matters because it removes a assumption that used to simplify model selection, that tighter licensing correlated with stronger performance because labs with more resources tended to license more conservatively. In 2026, some of the best-performing open models by category are also the most permissively licensed, meaning license terms and technical capability have become genuinely independent variables that both need separate evaluation rather than a single combined judgment call.

The practical consequence for technical teams is that license review can no longer be treated as a formality to complete after a benchmark-driven model choice is already made. With performance and licensing decoupled, the license terms deserve equal weight in the initial evaluation, not a secondary check applied only once a technical winner has already been selected.

How to Choose an Open Source LLM

A leaderboard rank is where the decision starts, not where it ends. Teams that choose well tend to work through the same filters in roughly this order.

1. Read the license before the benchmark: MIT and Apache 2.0 models, including MiMo-V2.6, DeepSeek V4.1 Flash, GLM-5.3-Flash, Qwen3.8-27B and Beam, allow commercial use, modification and redistribution without the extra commercial conditions found in custom licenses. The custom licenses on GLM-5.3, Kimi K3, Step 5 Preview, Qwen3.8’s Max tier and Muse Spark Open allow commercial use with restrictions, so legal review belongs at the start of the evaluation, not the end.
2. Size the hardware from total and active parameters: Total parameters set memory requirements; active parameters set per-token compute. Muse Spark Open, at 816 billion total and 42 billion active, needs about a terabyte of memory, which Meta reaches with two 512GB M5 Ultra Mac Studios. [Meta AI Muse Spark Open announcement] A 27-billion-parameter dense model such as Qwen3.8-27B is a different class of deployment: Alibaba says it runs on consumer-grade hardware and, quantized, on a laptop. [Alibaba Cloud Qwen3.8-27B announcement]
3. Match the context window to the job: Most leading open models now offer 1 million tokens, while the Qwen3.8 models ship with 256K to 262K natively. Long-codebase agents and large document sets need the former; most chat and retrieval workloads do not.
4. Test on your own workload: Composite scores hide category gaps, and vendor benchmark tables are claims until an independent evaluator repeats them. A few hundred prompts drawn from real tasks will tell you more than a one-point difference on a leaderboard.
5. Price the whole deployment, not the weights: Free weights still cost GPUs, utilization and engineering time to run. Hosted access to open models can be dramatically cheaper than closed frontier models: Artificial Analysis measures MiMo-V2.6-Pro at $0.13 per Intelligence Index task against $5.98 for Claude Opus 5.5 at its maximum setting. [Artificial Analysis LLM leaderboard]
6. Check provenance against your own policies: Some organizations restrict Chinese-developed models. US-developed options exist, including Muse Spark Open, Beam, NVIDIA Nemotron, gpt-oss and Gemma, but in October 2026 they score below the leading Chinese releases.
7. Plan to re-evaluate every quarter: The top open-weight spot changed hands between July and October 2026 and is likely to change again when Kimi K3.1 Pro’s weights ship. Treat the choice as a portfolio decision with a review date, as the business case above argues.

Strategic Outlook

1. Watch whether Meta actually ships an open-weight successor to Llama, since the company has publicly left that door open without committing to a timeline:  Meta Superintelligence Labs leadership has stated that bigger models are already in development with plans to eventually open-source future versions, but until a specific model and license actually ships, that commitment remains a stated intention rather than a confirmed roadmap item enterprises should plan around.[AI News Meta sacrifice open-source identity competitive AI model] Update, October 2026: Meta did ship an open-weight successor, Muse Spark Open, on October 15, 2026, under the Muse Model License with limited commercial terms rather than an OSI-approved license. The door has reopened, but on Meta’s terms, and its benchmark standing (39 on the Artificial Analysis index) sits below the leading Chinese open models.
2. Expect the Chinese open-weight labs’ download share to keep growing in the near term simply due to reduced Western competition at the frontier tier:  With Meta’s frontier effort now closed and no other major US lab positioning a frontier-class model as fully open, DeepSeek, Qwen, and any new entrants from Chinese labs face structurally less competition for developers specifically seeking a frontier-competitive, permissively licensed model.
3. Watch for the Open Source Initiative’s ongoing dispute over what qualifies as genuinely open source AI to gain more practical weight in enterprise procurement:  As the field-of-use restrictions in licenses like Meta’s Community License become more consequential to enterprises that hit usage caps, expect procurement and legal teams to weight OSI-approved licensing status more heavily as an explicit selection criterion rather than treating all models marketed as open source interchangeably.

Key Question Answered

What is the state of open source AI model releases in 2026, and why did Meta step back from the category?

The open-weight AI landscape in 2026 has been reshaped by one dominant event: Meta’s April 8, 2026 launch of Muse Spark, its first fully closed frontier model, built by the newly formed Meta Superintelligence Labs and available only through Meta’s own consumer apps. Meta has not discontinued Llama, but its heaviest investment and frontier development effort have shifted to the closed Muse Spark line, driven primarily by the economics of a 115 to 135 billion dollar 2026 AI capital expenditure budget and a desire to capture more value from tightly integrated distribution across Meta’s billions of existing app users.

That shift has not slowed the broader open-weight ecosystem, it has redirected it. DeepSeek’s MIT-licensed V4 series, Alibaba’s Apache 2.0 Qwen line, and Google’s Apache 2.0 Gemma 4 have all shipped genuinely competitive open models in 2026, with Chinese labs specifically capturing a growing majority of global open-source download share. The practical result for developers and enterprises is a more geographically diverse and more license-fragmented open-weight landscape than existed a year earlier, one where license terms and technical capability now need to be evaluated as independent factors rather than assumed to move together.

FAQ: Best Open Source LLMs

What is the best open source LLM right now?

As of October 26, 2026, Xiaomi’s MIT-licensed MiMo-V2.6-Pro is the highest-scoring model with published weights on the Artificial Analysis Intelligence Index, at 46. DeepSeek V4.1 Flash 1017 and Z.ai’s GLM-5.3 follow at 45. Moonshot’s Kimi K3.1 Pro scored 55 at launch and is expected to take the top spot once its weights are released.

Which open source LLMs can be used commercially without restrictions?

Models under MIT or Apache 2.0 carry no extra commercial conditions. In the ranked table above that includes MiMo-V2.6-Pro and MiMo-V2.6-Flash, DeepSeek V4.1 Flash and V4 Pro, GLM-5.3-Flash, Qwen3.8-27B and Beam, along with Google’s Gemma 4 and OpenAI’s gpt-oss. GLM-5.3, Kimi K3, Step 5 Preview, Qwen3.8’s Max tier and Muse Spark Open allow commercial use but add conditions in custom licenses.

What is the best open source LLM for smaller hardware?

Qwen3.8-27B is the strongest small model in the ranking. It scores 34 on the Artificial Analysis index with 27 billion dense parameters, a 262K native context window and an Apache 2.0 license, more than double the score of Gemma 4 31B.

Are open source AI models the same as open-weight models?

Not quite. Almost every model in this tracker is open-weight: the trained weights can be downloaded and run, but training data and full training code are usually withheld. The Open Source Initiative’s definition of open source AI is stricter, which is why license terms deserve as much scrutiny as the label a lab puts on a release.

The Takeaway

Meta’s pivot to closed frontier development with Muse Spark is the single most consequential open source AI story of 2026, not because it ends open-weight AI as a category, but because it removes the assumption that a well-resourced Western lab would keep releasing frontier-competitive weights openly as a matter of course. That assumption underpinned a meaningful share of enterprise and developer infrastructure planning, and 2026 is the year it stopped being safe to make.

The labs filling that gap, DeepSeek, Alibaba’s Qwen line, Google’s Gemma, and a growing cluster of Chinese entrants including GLM and MiMo, are proving that frontier-adjacent capability and genuinely permissive licensing can still coexist. But the geographic and licensing landscape those labs represent looks meaningfully different from the one Meta’s Llama dominated through 2023 and 2025, and enterprises whose AI infrastructure strategy has not accounted for that shift are working from an outdated map.

For any organization building on open weights, the practical mandate for the remainder of 2026 is treating model and license selection as a portfolio decision reviewed on a recurring basis, not a one-time choice. The open-weight leaderboard changed meaningfully in a single April week this year, and there is no strong reason to assume it will move more slowly for the rest of 2026 than it just did.