The Best Open Source LLMs in 2026

The best open source LLMs in 2026 are Kimi K3, Qwen3.8, DeepSeek-V4, GLM-5.2, and MiMo-V2.5-Pro. All of them reach frontier-adjacent performance on coding and agent benchmarks, and can be downloaded and self-hosted today.

The more useful finding is what separates them. This guide compares top open source LLMs on parameters, context, license, price, and more. It also tells you when to skip each model, which most model roundups leave out.

Best open source LLMs at a glance

Model	Org	Params (total / active)	Context	License	Output per 1M
Kimi K3	Moonshot AI	2.8T / 104B	1M	Kimi K3 License	$15.00
Qwen3.8	Alibaba	2.4T / 95B	1M (262K native)	Qwen3.8-Max License	$6.00
DeepSeek-V4-Pro-0813	DeepSeek	1.6T / 49B	1M	MIT	$1.98
GLM-5.2	Zhipu AI	753B / 40B	1M	MIT	$4.40
MiMo-V2.5-Pro	Xiaomi	1.02T / 42B	1M	MIT	$0.87

Prices are official API rates per 1 million tokens as of August 16, 2026. The DeepSeek figure is the off-peak rate. DeepSeek moved to peak and off-peak pricing on August 16, and the peak rate is double, at $3.96.

Three patterns stand out:

• Most of them use MIT. Permissive licensing is now the norm at the frontier of open weights, not the exception.
• Every model reaches a 1M-token context window. A million tokens has become the default for this tier rather than a headline feature. Qwen3.8 is the one that needs configuring: 262K natively, extensible to 1M.
• Price does not track parameter count. DeepSeek-V4-Pro-0813 has 57% of the total parameters of Kimi K3 and costs about one eighth as much per output token off-peak, while trading wins with it on agent benchmarks.

Open weights vs open source: what the labels mean

Most models in this guide are open weights, not open source in the strict sense. The distinction matters when you are reading a license.

Open weights means the model parameters are published and free to download, but the license may not meet the Open Source Initiative definition of open source. Some carry commercial-use limits, attribution requirements, or redistribution conditions.

Feature	Open weights	Open source
Weights and biases	Released	Released
Training code	Not shared	Fully shared
Intermediate checkpoints	Withheld	Nice to have
Training dataset	Not shared or disclosed	Released when legally allowed
Training data composition	Partially disclosed or not disclosed	Fully disclosed

Both let you self-host, inspect behavior, and fine-tune. The differences are licensing freedom and how much of the training pipeline is disclosed.

This guide does not go deeper into license taxonomy. The practical bar used here is simple: can you download the weights and run them on your own hardware.

Kimi K3

Kimi K3 is the latest open source flagship model from Moonshot AI, and a significant jump in scale rather than an incremental update over the K2 line. It is positioned for long-horizon coding, agentic knowledge work, and reasoning, and Moonshot bills it as the first open 3T-class model in the world.

Architecturally, K3 is a 2.8T-parameter MoE model with 104B active parameters per token, built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), paired with a Stable LatentMoE framework that activates 16 of 896 experts. This design yields roughly a 2.5× improvement in overall scaling efficiency over Kimi K2. Vision is native via a MoonViT-V2 encoder (~401M parameters), and the context window is 1M tokens.

Kimi K3 is released under the custom Kimi K3 License. It reads like MIT for most of the content (use, modify, fine-tune, distribute, sell), but adds two commercial gates:

• Model-as-a-Service providers whose revenue passes $20M over any consecutive 12 months must sign a separate agreement with Moonshot
• Any commercial product above 100M monthly active users or $20M monthly revenue must display “Kimi K3” in the product interface.

Purely internal use, where neither the model nor the outputs it produces ever reach a third party, is exempt from both.

Why use Kimi K3:

• Frontier-level long-horizon coding. K3 is the strongest open model on end-to-end software engineering today, with results competitive with Claude Fable 5 and GPT-5.6 Sol across coding benchmarks. More telling than the scores are the case studies:

  • A 24-hour autonomous GPU kernel optimization run, a from-scratch Triton-like compiler (MiniTriton) that matches or beats Triton on some roofline benchmarks
  • A 48-hour run that designed and verified a chip using open-source EDA tools.

• Agentic knowledge work and native multimodality. Text, image, and video are understood within the same model rather than bolted on, which shows up in deep research, document-heavy workflows, and interactive dashboards (e.g., BrowseComp, OfficeQA Pro, SpreadsheetBench 2, OmniDocBench, Video-MME). It also handles motion design and video editing, which is unusual territory for an open source model.

• Serving economics that are better than the parameter count suggests. K3 applies quantization-aware training from the SFT stage onward, shipping MXFP4 weights with MXFP8 activations for broad hardware compatibility. Moonshot also contributed a KDA-compatible prefix caching implementation to vLLM, since KDA breaks conventional prefix caching.

Note that:

• K3 is trained in preserved thinking history mode, so your harness must pass back the complete assistant message, including reasoning_content and tool_calls, on every turn. Dropping it, or switching an in-flight session from another model to K3, makes generation quality highly unstable. I suggest you use a verified harness such as Kimi Code CLI.
• Kimi K3 provides configurable reasoning efforts. Thinking is always on, with a top-level reasoning_effort field supporting low, high, and max (default max).
• Because K3 was trained heavily on long-horizon, open-ended tasks, it tends toward excessive proactiveness. When it hits an ambiguous instruction or a minor blocker mid-task, it will often make a decision on your behalf rather than stopping to ask. If you are wiring K3 into an agentic framework and need the agent to stay inside well-defined boundaries, put explicit behavioral constraints in the system prompt or AGENTS.md.

Full specs, benchmarks, and pricing for Kimi K3

Qwen3.8

Qwen3.8 is the largest model Alibaba has ever released at 2.4 trillion parameters, with 95 billion active per token. Alibaba announced Qwen3.8-Max on August 3, 2026 and published the weights on August 12 as Qwen3.8-2.4T-A95B, the first Qwen model at Max scale to be opened.

The model uses a fine-grained sparse MoE architecture with hybrid attention. Across 92 layers, three Gated DeltaNet linear-attention layers run before every full attention layer, so only a quarter of the layers accumulate a growing KV cache.

Qwen3.8 is released under the custom Qwen3.8-Max License. Like the Kimi K3 License, it permits use, modification, fine-tuning, and redistribution, then adds two commercial gates:

• Model-as-a-Service or AI Work Assistant businesses whose revenue passes $50M over any consecutive 12 months must obtain a separate license from Qwen
• Any commercial product above 100M monthly active users or $20M monthly revenue must display the model name prominently in the product interface.

Why use Qwen3.8:

• Improved coding and professional work. Handed nothing but a research paper and a set of GPUs, Qwen3.8 wrote the pipeline, training code, and evaluation from scratch in about 125 hours, reproduced all six findings, then invented a method that beats the original by 2.7 points on AIME24. The professional track is just as strong, surfacing 1,284 relevant clauses from hundreds of contracts in under an hour.
• Excellent agent execution. Given an empty RTL workspace and no reference design, it cut a cryptographic accelerator from 8,298 gates to 678 across roughly 500 turns. Execution also survives a change of scaffold, staying comparable across QwenWork, Claude Code, Codex, OpenClaw, and Hermes, where most models are tuned to one harness and drop elsewhere.

Note that:

• The open weights are not the same product as the hosted Qwen3.8-Max API. The download is text-only, and thinking cannot be disabled: every response opens with a <think> block. The API keeps vision input, a non-thinking mode, 1M context by default, and built-in tools.
• The published benchmark scores describe the hosted system, so the vision and 1M-context rows are not reproducible on the download.

Skip Qwen3.8 if you need vision or a non-thinking mode, or you can’t afford the required hardware for self-hosting. BF16 weights run about 4.45 TiB, which requires three 8× B300 nodes before you account for quantization.

Full specs, benchmarks, and pricing for Qwen3.8

DeepSeek-V4

DeepSeek-V4 offers the best cost-to-performance ratio in this guide, and it ships in two sizes under the MIT license. Both support a 1M-token context and were pre-trained on more than 32 trillion tokens.

• DeepSeek-V4-Pro-0813 (1.6T total, 49B active) is the flagship for reasoning and coding, and the official build since August 2026.
• DeepSeek-V4-Flash-0731 (284B total, 13B active) is the official Flash build, retrained for agents in July 2026.

The architecture interleaves Compressed Sparse Attention (CSA) with Heavily Compressed Attention (HCA). CSA compresses blocks of the KV cache and attends only to the most relevant entries. HCA folds much larger spans into single representations for a cheap global view. A sliding window of recent tokens stays uncompressed to preserve local accuracy.

Why use DeepSeek-V4:

• Post-training is doing the heavy lifting. Both official builds kept the architecture of the April previews and changed only post-training, and both jumped on agent work. V4-Pro-0813 went from 72.1 to 87.9 on Terminal Bench 2.1 and from 12.8 to 62.7 on DeepSWE against the preview it replaced.
• The flagship earns the premium again, narrowly. For a few weeks in July, V4-Flash-0731 outscored the V4-Pro preview on every agent benchmark DeepSeek published. The August V4-Pro-0813 build restored the expected order, leading Flash on all eleven, but the margins are modest: 87.9 against 82.7 on Terminal Bench 2.1, 62.7 against 54.4 on DeepSWE. V4-Pro costs three times more per output token, so Flash remains the better default for high-volume work.
• Long context without the memory bill. At 1M tokens, V4-Pro uses 27% of the per-token inference FLOPs and 10% of the KV-cache memory of DeepSeek-V3.2.
• Still the cheapest way into this tier. Off-peak, V4-Flash-0731 costs $0.22 per 1M cache-miss input tokens and $0.66 per 1M output tokens. A cache hit drops input to $0.007, a 97% discount.
• Codex support out of the box. Both official builds natively speak the Responses API format, so you can register DeepSeek as a model provider directly. DeepSeek also open sourced an agent harness of its own, DeepSeek Harness, under the MIT license.

One caveat on those prices: they went up, a lot. On August 16, 2026, DeepSeek replaced flat rates with peak and off-peak tiers. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, and off-peak is half the peak rate. V4-Pro output moved from $0.87 to $3.96 at peak and $1.98 off-peak. V4-Flash moved from $0.28 to $1.32 and $0.66. Measured on cached input the largest increase runs above 1,100%, so re-check the DeepSeek pricing page against your own traffic pattern before budgeting.

Skip DeepSeek-V4 if you need multimodal input. Both variants are text only. If your workload involves images, video, or audio, Kimi K3 or MiMo-V2.5-Pro are the better fits. The open Qwen3.8 weights are text-only too.

Full specs for DeepSeek-V4-Pro-0813 and DeepSeek-V4-Flash-0731

GLM-5.2

GLM-5.2 is the latest flagship LLM from Zhipu AI, built for agentic engineering, software development, and long-horizon reasoning tasks. It is shipped under the MIT license at 753B total parameters with 40B active and a 1M-token context.

GLM-5.2 keeps the GLM-5 backbone and DeepSeek Sparse Attention (DSA). The new addition is IndexShare, a long-context optimization that reuses one sparse-attention indexer across each group of four sparse-attention layers instead of recomputing it per layer. Zhipu reports this cuts per-token compute by roughly 2.9x at the full 1M-token context.

Why use GLM-5.2:

• Benchmark leadership on software engineering. GLM-5.2 leads on SWE-Bench Pro, Terminal-Bench 2.1, and FrontierSWE, in some cases ahead of GPT-5.5 and Claude Opus 4.8. It allows you to select different thinking levels to balance performance and latency.
• Repository-scale context that is actually usable. The 1M window lets coding agents hold a mid-sized repository in context instead of constantly compacting.

Skip GLM-5.2 if price per token is the deciding factor, though that argument weakened in August 2026. At $1.40 input and $4.40 output per 1M tokens, GLM-5.2 used to cost about 5x DeepSeek-V4-Pro on output. After the DeepSeek price increase it is about 2.2x the off-peak rate and only 1.1x the peak rate, and the two are close on many coding tasks.

MiMo-V2.5-Pro

MiMo-V2.5-Pro is the most token-efficient model in this guide, which matters more than raw speed on long agent runs. Xiaomi released the 1.02T-parameter MoE with 42B active per token under the MIT license, trained on 27T tokens in FP8 mixed precision.

The architectural centerpiece is long-context efficiency. MiMo-V2.5-Pro interleaves sliding-window attention and global attention at a 6:1 ratio with a 128-token window. That cuts KV-cache storage by nearly 7x while preserving long-context performance.

Why use MiMo-V2.5-Pro:

• Fewer tokens for the same result. On ClawEval, MiMo-V2.5-Pro matches Claude Opus 4.6 while using roughly 40 to 60% fewer tokens per trajectory. Across long agentic runs that difference compounds into real money.
• Long-context performance that holds. The hybrid SWA/GA architecture is purpose-built for long-context tasks. On the GraphWalks benchmark from OpenAI, MiMo-V2.5-Pro maintains strong performance well past 512k tokens. By contrast, the previous V2-Pro collapsed to 0 at that length.

Skip MiMo-V2.5-Pro if your infrastructure is modest, or if you need frontier capability.

• A 1.02T-parameter model needs a two-node 8× H200 deployment, which is a meaningful step up from the single-node setups. The smaller MiMo-V2.5 at 310B total and 15B active is the multimodal sibling and a far lighter lift.
• Age is another consideration. Both MiMo-V2.5 models shipped in April 2026, which puts them among the oldest releases in this guide, and three months is a long time at the current pace. Kimi K3 arrived in July 2026 with a clearly higher ceiling on coding and agentic work, so reach for K3 when you need the frontier and keep MiMo-V2.5-Pro when token efficiency is what you are optimizing for.

Frequently asked questions

What is the best open source LLM right now?

There is no single best open source LLM. The right choice depends on your workload, budget, and hardware. Here are commonly recommended open source LLMs for different use cases.

• Price-performance: DeepSeek-V4-Flash-0731
• Coding: Kimi K3
• Agentic engineering: GLM-5.2
• General chat: DeepSeek-V4-Pro-0813
• Story writing & creative tasks: DeepSeek-V4-Pro-0813

Treat these as starting points, not final answers. The open source LLM space moves fast enough that a model released this quarter can reset the list. The timeline of major milestones gives a sense of the pace.

How big is the gap between open source and proprietary LLMs?

The gap between open-source and proprietary LLMs has narrowed dramatically, but it is not uniform across all capabilities.

Use case	Gap	Notes
Coding assistants and agents	Small	GLM-5.2 and Kimi K3 are already competitive
Math and reasoning	Small	DeepSeek-V4 reaches Opus-tier performance
General chat	Small	Open models increasingly match GPT-5-tier quality
Multimodal image and video	Moderate to large	Closed models still lead on performance and polish
Extreme long context with high reliability	Moderate	Proprietary models hold up better at scale

Why choose open source LLMs over proprietary ones?

Open source LLMs win on four dimensions:

• Customization. You can fine-tune on your own data and apply inference optimizations like speculative decoding, prefix caching, and prefill-decode disaggregation for your specific SLA goals. None of that is possible behind a proprietary API.
• Data control. Models run inside your own infrastructure, so sensitive data never leaves your boundary.
• Cost at scale. Infrastructure requires investment, but recurring per-token API costs disappear. With inference optimization the price-performance ratio often beats commercial APIs.
• No vendor lock-in. You do not depend on the roadmap, pricing, or uptime of a single provider.

Proprietary models still make sense for low-volume prototypes, workloads needing the absolute frontier on multimodal tasks, and teams without infrastructure capacity.

How do I optimize open source LLM inference?

Start with a serving engine that implements the major optimizations for you. vLLM and SGLang both ship continuous batching, prefix caching, and speculative decoding.

As models grow, single-node tuning stops being enough. The KV cache expands, GPU memory becomes the bottleneck, and long-context agent runs stretch past what one node holds. Distributed techniques then matter: prefix-aware routing, KV cache offloading, tensor and data parallelism, and prefill-decode disaggregation. The inference optimization page covers the major techniques and which engines implement them.

What should I consider when deploying open source LLMs in production?

• Model size. Smaller models respond faster and cost less to serve. Benchmark against your own workload before committing.
• GPUs. These workloads are bound by GPU memory and bandwidth. Common choices are NVIDIA H200 and B200 or AMD MI325X and MI355X. Benchmark on the hardware you intend to buy.
• Scalability. Autoscaling needs fast cold starts, or user experience suffers during traffic spikes.
• Observability. Track Time to First Token (TTFT), Inter-Token Latency (ITL), and token throughput alongside standard monitoring.
• Context length. A 1M-token window costs real memory. Long sequences can require close to 1TB of GPU memory once weights, KV cache, and activations are counted.

Final thoughts

The open source tier in 2026 is no longer a budget alternative. DeepSeek-V4-Pro-0813 scores 87.9 on Terminal Bench 2.1, ahead of Claude Opus 4.8, at $1.98 per 1M output tokens off-peak. The budget framing is fading in another sense too: DeepSeek raised API prices several times over in August, so open weights now compete on capability rather than on price alone.

The practical takeaway: do not marry a model. Pick a serving stack that makes swapping models cheap, then re-evaluate every quarter.

For more numbers on the ecosystem, see the open source LLM statistics and the DeepSeek statistics roundups.