Comparisons

Best Open-Source AI Models Compared: Llama, Qwen, DeepSeek

James Carter

James Carter

March 13, 2026

Best Open-Source AI Models Compared: Llama, Qwen, DeepSeek

Disclosure: Some links in this article are affiliate links. If you sign up or purchase through them, we may earn a commission at no extra cost to you. This helps support our independent reviews.

Updated November 7, 2026. The first version of this guide compared Llama 3, Mistral 7B, Gemma 2 and Phi-3. All of them have been replaced since, so we rewrote it.

When we first published this comparison, a 70B dense model was the ceiling of what you could download and Meta's Llama set the pace. That world is gone. The strongest open-weight models today are trillion-parameter mixture-of-experts systems from Moonshot, Z.ai, DeepSeek and Alibaba, Meta has moved on from Llama to a family called Muse, and Google ships Gemma under Apache 2.0.

We have not run these models through a lab of our own. What follows comes from the model cards on Hugging Face, the vendors' release notes, and two independent scoreboards, the Artificial Analysis Intelligence Index and the Arena text leaderboard, both read on November 7, 2026. Where a number is the vendor's own, we say so. Most of these are open weight rather than open source, and we spell out the revenue clauses because that is the detail that bites a business later.

The short list

You want Our pick Why
The strongest model you can download Kimi K3.1 Pro Top open-weight score on Artificial Analysis (46); needs a GPU cluster
Coding and terminal agents GLM-5.3, or GLM-5.3-Flash on a budget 88.2 on Terminal-Bench 2.1 (vendor figure); Flash is MIT
Cheap frontier-class API, open weights DeepSeek V4.1-Flash MIT, 1M context, 16B active parameters
One 24 GB graphics card Qwen3.8-27B or Gemma 4 31B 4-bit builds fit in 16 to 20 GB; both read images
16 GB of VRAM gpt-oss-20b or Gemma 4 26B A4B 14 GB download; the Gemma needs 16 to 18 GB
8 GB of VRAM Gemma 4 12B or Qwen3.5-9B 7 to 8 GB at 4-bit; 6.5 GB for the Qwen
Many languages Qwen3.5 and Qwen3.8 Alibaba claims 201 languages and dialects
The cleanest licence Gemma 4, Qwen3.8-27B, gpt-oss, Muse Glimmer Plain Apache 2.0, no clauses

Comparison table: parameters, licence, context, VRAM

Memory figures are the RAM plus VRAM that Unsloth or Ollama publish for a 4-bit build, not our measurements. "n.p." means not published.

Model (developer) Parameters, total / active Licence Context Memory, 4-bit local Best at
Kimi K3.1 Pro (Moonshot) 2.8T / 104B Kimi K3.1 Pro licence (custom) 1M Cluster; 1.5 TB native Agents, research
GLM-5.3 (Z.ai) about 753B / n.p. GLM-5.3 licence (custom) 1M in Z.ai's evals Cluster Coding agents
GLM-5.3-Flash (Z.ai) 320B / 18B MIT n.p. Workstation Coding on a budget
DeepSeek V4-Pro-0813 1.6T / 49B MIT 1M Cluster Reasoning, price
DeepSeek V4.1-Flash-1017 552B / 16B at decode MIT 1M Workstation (V4-Flash: 162 GB) Cheap API
Qwen3.5-397B-A17B (Alibaba) 397B / 17B Apache 2.0 262K, 1M with YaRN 214 GB Multilingual, vision
Qwen3.8-27B (Alibaba) 27B dense Apache 2.0 262K 16 to 19 GB A 24 GB card
Gemma 4 31B (Google) 31B dense Apache 2.0 256K 17 to 20 GB All-round, vision
Gemma 4 26B A4B (Google) 26B / 3.8B Apache 2.0 256K 16 to 18 GB Speed on 16 GB
Muse Spark Open (Meta) 816B / 42B Muse Model licence 1M 700 GB VRAM at NVFP4 US-built frontier
Muse Glimmer 30B (Meta) 30B dense Apache 2.0 131K 17 GB quant, 24 GB card Local agents
Llama 4 Scout (Meta) 109B / 17B Llama 4 Community 10M 66 GB Long context (2025)
Mistral Large 4 1T / 52B Mistral Ecosystem Open-Weight (modified MIT) 1M 8-GPU node Coding, security
gpt-oss-120b / 20b (OpenAI) 117B / 5.1B and 21B / 3.6B Apache 2.0 128K 65 GB and 14 GB Reasoning, one GPU

Kimi K3 and K3.1 Pro: the open-weight frontier

Moonshot AI's Kimi K3 is the largest thing you can download: 2.8 trillion parameters, 104 billion active per token, a one-million-token context (model card). The weights went public on July 27, 2026 (release table on Wikipedia). K3.1 Pro, posted to Hugging Face in early November, keeps the size and swaps full attention for sparse attention to make long contexts cheaper (K3.1 Pro card). It scores 46 on the Artificial Analysis index, the best of any open model. Arena voters put K3 at rank 19 overall, again the top open entry.

The licence is the catch. It reads like MIT until section 2: sell the model as a hosted service and pass 20 million dollars of group revenue in any twelve months, and you need a separate agreement with Moonshot; above 100 million monthly users your product must display "Kimi K3.1-Pro" (licence text). Home use is out too, since Unsloth's smallest 1-bit cut still wants 610 GB of memory (Unsloth). Our verdict: rent it through an API. Do not plan to own it.

GLM-5.3

Z.ai, formerly Zhipu, opened 2026 with GLM-5, a 744B mixture of experts with 40B active under plain MIT, released in February (card). GLM-5.3 reached Hugging Face in late August with roughly 753B parameters and a vendor table that puts it at 88.2 on Terminal-Bench 2.1 and 84.5 on CyberGym (card). Artificial Analysis gives it 32; Arena ranks it 26th overall, second open model after Kimi.

GLM-5.3 moved from MIT to a custom "GLM-5.3 License", so we read it. The only extra condition is a Z.ai security review for hosted-model businesses above 10 billion dollars of revenue, which is to say almost nobody (licence). GLM-5.3-Flash, 320B total and 18B active, stays MIT and scores 84.3 on the same test (card). For the kind of work in our AI coding assistants comparison, start here, Flash first if you pay for the GPUs.

DeepSeek V4

DeepSeek skipped the R2 everyone expected and shipped V4 on April 24, 2026 in two sizes: V4-Pro at 1.6T total and 49B active, V4-Flash at 284B and 13B active, both MIT, both with a 1M context (DeepSeek on Wikipedia; V4-Pro card). V4.1-Flash followed on September 10 and the 1017 build in October. V4.1 moves to a 552B backbone with 8B parameters active while reading and 16B while writing, which cuts the KV cache to a quarter of V4-Flash's (card). Artificial Analysis scores it 33.

Locally, V4-Flash's 4-bit build needs about 162 GB of RAM plus VRAM (Unsloth). Workstation territory. Our verdict: the open model we would wire into a product where price per token matters most, with the least ambiguous licence of the trillion-class group.

Qwen3.5, 3.6 and 3.8: a size for every budget

Alibaba releases more open models than anyone. Qwen3.5 landed on February 16, 2026 (CNBC) with a 397B flagship, 17B active, under Apache 2.0, taking image and video input, with a 262K context and a claimed 201 languages (card). Smaller 3.5 sizes followed within weeks, from 122B-A10B down to 0.8B. Qwen3.6-35B-A3B in April reports 73.4 on SWE-bench Verified from a model that runs in 22 GB (card; Unsloth memory table).

August brought the 3.8 generation and a split. Qwen3.8-27B is dense, Apache 2.0, reads images, and posts 90.3 on LiveCodeBench v6 and 89.2 on GPQA Diamond in Alibaba's table (card). At 4-bit it needs 16 to 19 GB, so an RTX 4090, a 5080 or a 24 GB Mac (Unsloth). The 2.4-trillion-parameter Qwen3.8-2.4T-A95B is open weight too, but text only and under a "Qwen3.8-Max" licence: hosted-model or AI-assistant businesses above 50 million dollars of revenue in twelve months need a commercial licence (licence). Our verdict: Qwen3.8-27B is our default for a 24 GB card.

Gemma 4

Google fixed the one thing people complained about: Gemma 4 ships under Apache 2.0 instead of the old Gemma terms (Google's announcement, April 2, 2026). Five sizes exist: E2B and E4B for phones and single-board computers, a 12B added in May, a 26B mixture of experts with 3.8B active, and a 31B dense model. All take images and video, the three smallest audio too. Edge models get a 128K context, the rest 256K (model card).

The card's numbers are strong for the size: the 31B scores 85.2 on MMLU Pro, 80.0 on LiveCodeBench v6 and 84.3 on GPQA Diamond, and the 26B A4B lands within a few points while running far faster. Unsloth puts the 4-bit builds at 4 GB for E2B, 5.5 to 6 GB for E4B, 7 to 8 GB for the 12B, 16 to 18 GB for the 26B A4B and 17 to 20 GB for the 31B (Unsloth). Our verdict: the best family at 8 GB of VRAM, and our joint pick with Qwen3.8-27B at 24 GB.

Meta: from Llama 4 to Muse

Llama 4 Scout and Maverick, released April 5, 2025, were Meta's last Llama weights: 109B and 400B total, 17B active each, a 10M-token context on Scout, and the Llama 4 Community License with its 700-million-monthly-user clause (Scout card). Behemoth never shipped. Scout's 4-bit build still needs about 66 GB (Unsloth).

Meta's new line is Muse. Muse Spark, in April 2026, was closed; two open releases followed. Muse Glimmer 30B, August 2026, is a dense model with a built-in vision encoder under plain Apache 2.0; Meta's card claims 76.0 on SWE-bench Verified and says the 17 GB quant fits a 24 GB card (Glimmer card; Ollama). Muse Spark Open, October 15, 2026, is the big one: 816B total, 42B active, text, image, video and audio in, a 1M context, and 39 on the Artificial Analysis index at launch, the top US-built open model (Spark Open card; The Batch). Its licence asks for a "Built with Muse Spark Open" notice and a separate agreement for hosted-model businesses above 20 million dollars in twelve months. Our verdict: Glimmer is a real rival to Qwen3.8-27B for local agents; Spark Open is for companies that own the hardware.

Mistral

Mistral 3, on December 2, 2025, put Mistral Large 3 (675B, 41B active) and the Ministral 3 trio of 3B, 8B and 14B under Apache 2.0 (announcement); Ministral 3 14B reads images and fits in 24 GB at FP8 (card). A week later Devstral Small 2 (24B, Apache 2.0, 68.0 on SWE-bench Verified) targeted coding agents on one GPU (announcement). Mistral Small 4, March 16, 2026, is a 119B mixture of experts with 6B active, Apache 2.0, sized for a four-H100 server (announcement).

Mistral Large 4 was previewed on October 6 and its weights went public on October 29, 2026 (Fortune). It is a 1-trillion-parameter model with 52B active, a 1M context and image input, under a modified MIT that requires a commercial licence only for hosted services (card). Mistral reports 48 percent on the Artificial Analysis coding-agent index, tied with GLM-5.3 at 47 percent fewer tokens per task, and an Intelligence Index score of 26. Our verdict: the European choice when procurement cares where a model was trained; Ministral 3 and Devstral Small 2 are the parts a solo developer will run.

gpt-oss

OpenAI's gpt-oss pair from August 2025 has not been refreshed, and it still earns its place. gpt-oss-120b (117B total, 5.1B active) runs on one 80 GB GPU, and gpt-oss-20b (21B, 3.6B active) runs in 16 GB, because the expert weights ship already quantised to MXFP4 (card). Both are Apache 2.0 with a 128K context; Ollama's files are 65 GB and 14 GB (Ollama). Text only. Our verdict: still the simplest reasoning model to fit on a 16 GB card.

Phi

Microsoft's Phi line stayed small. Phi-4 (14B, December 2024) spawned Phi-4-reasoning and two 3.8B variants; the newer, Phi-4-mini-flash-reasoning, carries a 64K context (card). The 2026 entry is Phi-4-reasoning-vision-15B, March 4, 2026, MIT, with a 16K context and 88.2 on the ScreenSpot v2 GUI-grounding test (card). There is no Phi-5. Our verdict: pick Phi for a maths or screenshot task on a tiny box; for general chat the small Gemma 4 and Qwen models have moved past it.

Others worth a look: Nemotron 3, MiMo, OLMo

NVIDIA's Nemotron 3 family, hybrid Mamba-Transformer mixtures of experts, runs from the 30B Nano (December 2025) through Super 120B-A12B (March 2026) to Ultra 550B-A55B (June 2026), all under the NVIDIA Nemotron Open Model License (Nano card; Super card). Xiaomi's MiMo-V2.6-Pro (September 2026, 1.02T total, 42B active, MIT) scores 30 on Artificial Analysis (card). For open in the full sense, training data and code included, Ai2's OLMo 3 (7B and 32B, November 2025, Apache 2.0) is the one to study (Ai2).

Running one of these on your own machine

Ollama remains the fastest path; every tag below exists today.

ollama run gemma4          # Gemma 4 E4B, about 7 GB
ollama run gpt-oss:20b     # 14 GB, wants 16 GB of memory
ollama run qwen3.8         # Qwen3.8-27B, 18 GB, reads images
ollama run muse-glimmer    # Meta Muse Glimmer 30B, 18 GB

Unsloth's rule is the one to remember: RAM and VRAM together should exceed the size of the file you download, or llama.cpp pages from disk and crawls. A 24 GB card with 64 GB of system RAM covers the 27B to 35B tier with room for context. If you are shopping for the machine, we keep a list of laptops for AI development and a cheaper set of budget picks for AI and coding.

How we pick, and what we do not do

We did not benchmark these models. We read every model card and licence linked above, took release dates from the vendors or Hugging Face timestamps, and used Artificial Analysis and Arena as independent references. We weighted four things: licence terms, because they decide whether you can ship; memory at 4-bit, because that decides whether you can run it; context length; and published coding and agent results. Raw chat quality got little weight, which is why Kimi K3 takes the top spot without being our pick for most people. Both leaderboards shift weekly; treat the numbers as a November 2026 snapshot.

Frequently Asked Questions

What is the best open-source AI model right now?

On the Artificial Analysis index as of November 7, 2026, Kimi K3.1 Pro at 46, then DeepSeek V4.1-Flash and GLM-5.3 in the low thirties. Those need a cluster. For a model you run yourself, Qwen3.8-27B and Gemma 4 31B.

Which open model can I run on 8 GB of VRAM?

Gemma 4 12B at 4-bit needs 7 to 8 GB of combined memory and takes images and audio; Qwen3.5-9B needs about 6.5 GB; Gemma 4 E4B gets down to 5.5 GB.

Are these models really open source?

Mostly no. Gemma 4, Qwen3.8-27B, gpt-oss, Muse Glimmer and OLMo 3 are Apache 2.0; DeepSeek V4 and GLM-5.3-Flash are MIT. Kimi K3, Qwen3.8-2.4T, Muse Spark Open, GLM-5.3 and Mistral Large 4 use custom licences with revenue triggers for companies reselling the model as a service.

Is Llama still worth using?

Llama 4 Scout and Maverick still work, and Scout's 10M context remains unusual. But Meta has published no Llama weights since April 2025 and its 2026 open models carry the Muse name. For a new project we would start with Muse Glimmer 30B.

How far behind the closed models are they?

Fourteen points on the Artificial Analysis index, 46 against 60 for Gemini 4 Carbon, and rank 19 on Arena. If that gap matters for your task, our ChatGPT vs Claude vs Gemini comparison covers the paid side.

Which open model is best for coding?

For an agent in a terminal, GLM-5.3 and Kimi K3.1 Pro post the highest published Terminal-Bench results, with GLM-5.3-Flash as the affordable version. On a single card, Qwen3.8-27B, Muse Glimmer and Devstral Small 2 are the three we would try first.

You might also like