The best open-source AI models in 2026, November update

The best open-source AI models right now

Kimi K3 from Moonshot AI is the largest open-weight release so far: a 2.8-trillion-parameter mixture-of-experts model with native vision and a 1-million-token context window, launched July 16, 2026 with full weights on Hugging Face eleven days later. Moonshot's own Terminal-Bench 2.1 result of 88.3 sits within half a point of GPT-5.6 Sol. The catches: a custom Kimi K3 License rather than MIT or Apache, an API at $3 per million input tokens and $15 output, and a checkpoint around 1.4 TB that makes self-hosting a data-center job. Pick it when raw capability matters more than cost.

GLM-5.3 from Z.ai is the agentic coding flagship. Released August 14, 2026, it keeps the GLM-5.2 base and gains everything through post-training: Terminal-Bench 3.0 jumps from 4.6 to 28.3, and it scores 84.5 on CyberGym for vulnerability discovery, ahead of the reported 83.8 for Anthropic's Mythos 5. The family has cheaper tiers: GLM-5.3 Flash is a 320-billion-parameter model with 18 billion active, priced at $0.15 per million input tokens and $0.50 output. Z.ai gates its most sensitive cyber functions behind a trusted-access program, which tells you where the capability frontier sits.

DeepSeek V4.1-Flash is the efficiency play. A 552-billion-parameter mixture-of-experts design with only 8 to 16 billion parameters active per token, it serves a 1-million-token context at $0.30 per million input tokens and $1.20 output at peak, with off-peak rates at half that. The October 17 checkpoint refresh (V4.1-Flash-1017) added five more weeks of reinforcement-learning post-training, and DeepSeek is phasing out V4-Pro in its favor. If volume is high and budget is tight, this is where the math starts.

Qwen3.8 from Alibaba is the workstation option. Secondary reports tie the 27B dense checkpoint to an Apache 2.0 license, making it the cleanest legal and hardware fit in this list: one good GPU, no custom license review. The larger reported Qwen3.8 open checkpoints (2.4 trillion parameters, 95 billion active) carry a custom license, so verify the model card before building on them.

Ultramarine 1 Flash is the niche pick that keeps beating generalists. Paper Instruments post-trained GLM-5.3 Flash with a rank-64 LoRA on professional knowledge work, and it now outscores Claude Fable 5 on finance, law and operations benchmarks at $0.13 per completed task against Fable's $8.19. We covered the full Ultramarine 1 Flash release and its RL method when it launched in October.

One to watch: Meituan's LongCat-2.5, a 1.6-trillion-parameter multimodal model released November 1, is API-only for now at $0.30 per million input tokens, with weights expected but not yet published. If the weights land, it goes straight into the top three.

Best open-source AI models for coding

For coding specifically, the ordering tightens. GLM-5.3 holds open-source state of the art on Terminal Bench 3.0 and DeepSWE in Z.ai's testing, and its Flash tier is the cheapest coding-grade weights you can download. Kimi K3 matches it on Terminal-Bench 2.1 (88.3 vs 88.2) and leads on long-context repository work thanks to the 1M window. DeepSeek V4.1-Flash is the budget default for high-volume agent loops, and it now ships inside a full open agent harness, which we covered in the DeepSeek harness note. For most teams the practical answer is GLM-5.3 Flash for daily driver work and Kimi K3 when the context gets long.

Best open-source AI models to run locally

Running locally is where total parameters stop mattering and active parameters take over. DeepSeek V4.1-Flash's 8 to 16 billion active parameters explain its serving cost, but 552 billion total parameters still need hundreds of gigabytes of memory, so "local" means a server, not a laptop. The same applies to Kimi K3's 1.4 TB checkpoint and GLM-5.3's flagship tier. Realistic single-machine options are Qwen3.8 27B and the smaller GLM Flash variants quantized. Ultramarine 1 Flash needs about four B300-class GPUs at FP8 in Paper Instruments' own serving setup. If your hardware is a workstation, download Qwen; if it is a cluster, pick on benchmark fit.

Open-source AI models vs closed in 2026

The gap is now a routing question, not a capability question. On vendor-run benchmarks the best open weights sit within a few points of GPT-5.6 Sol and Claude Fable 5 on coding and agentic suites, and the open side wins every price comparison by an order of magnitude: GLM-5.3 Flash at $0.15 input versus Claude Opus 5.5 territory at $14.30 per Terminal-Bench task. Where closed models still lead is the hardest long-horizon agentic work and the convenience of a single managed endpoint. The pattern that works in practice: route routine volume to open weights, reserve frontier closed spend for the hard tail.

One caution on the label. Almost none of these models are open source in the strict sense. You get weights and a model card, not training data or full training code, and the license decides what a business can ship. Kimi K3 uses a custom license, GLM-5.3's flagship terms differ from GLM-5.2's MIT, and only the smaller tiers and the Qwen 27B carry clearly permissive terms. Send the license text to legal before the first commercial pilot.

FAQ

Sources

1. 01Kimi K3: 2.8T Open Model for Coding and Knowledge Work - Moonshot AI (opens in a new tab)
2. 02GLM-5.3: Frontier Coding with Emergent Cyber Capabilities - Z.ai (opens in a new tab)
3. 03Introducing DeepSeek-V4.1-Flash - DeepSeek (opens in a new tab)
4. 04Best Open Source LLM: 5 Models From $0.15 Input - Shattered (opens in a new tab)
5. 05Ultramarine 1 Flash - Paper Instruments (opens in a new tab)
6. 06Introducing LongCat-2.5 - Meituan (opens in a new tab)

Elsewhere on this site

The selected work catalogue lists the SaaS products and dashboards behind notes like this one.