Best Open-Source LLM in 2026: Qwen3.8 vs Gemma 4 vs GLM-5.3

Choose Qwen3.8-27B for a practical local assistant with an Apache-2.0 license; choose Gemma 4 when you need smaller options. For a server deployment, shortlist DeepSeek V4.1 Flash, GLM-5.3 and Kimi K3 against your own tasks.

“Best open-source LLM” needs a qualification: this comparison covers downloadable open-weight models. A custom weight license is not interchangeable with Apache-2.0 or MIT, and an API subscription does not provide model weights.

The recommendations below prioritize deployability, licensing and workload fit. They are editorial picks, not a claim that one model wins every benchmark. Start with the smallest candidate that meets your acceptance criteria.

Comparison table

Model links lead to the publishers' Hugging Face cards. Context is the card's stated capacity, not a promise that a minimum-memory installation can use the whole window. B means billion parameters; T means trillion.

The hardware estimates assume short context and one request. Qwen's Ollama Q4_K_M package is 18 GB; Google's Q4_0 loading estimates are 14.4 GB and 17.5 GB for Gemma's larger models. Packages, memory guide.

These figures leave additional runtime and context allocations to measure. A file size, a loading estimate and a tested deployment requirement describe different things; none establishes a universal maximum prompt length.

DeepSeek: V4 and V4.1 Flash

Both V4 variants have downloadable weights. The Flash/Pro distinction matters: the card specifies 284B/13B and 1.6T/49B total/active parameters, respectively, with mixed FP4/FP8 instruct weights. V4 card.

V4.1 Flash accepts images and text and produces text. Its card separates a 552B backbone from 196B of Engram conditional memory; Hugging Face reports a 763B repository inventory. These are different accounting scopes.

Its active count also changes between prompt processing and generation. Size a deployment from the actual tensors and inference recipe, not the “Flash” name. V4.1 card.

My recommendation is to evaluate it for agents that repeatedly consume documents or images. Keep the model's prompt encoding and tool protocol intact, and measure a full task rather than treating a benchmark headline as a purchase decision.

Qwen: distinguish weights from hosted products

Qwen3.8-27B is my first workstation candidate because its size and license suit that role. Its native context is 262,144 tokens; reaching one million requires extension settings. Model card.

The 2.4T release uses the Qwen3.8-Max license, while Flash-Next uses Qwen Community 1.0. Do not infer a family-wide Apache license from the 27B model. Large-model card, Flash-Next card.

Flash-Next's embeddings and prediction module add storage beyond its backbone. Keep those components in the memory budget even when only a small fraction participates in each token's computation.

Qwen3.7 Max, Plus and Flash are documented as hosted services; downloadable counterparts were not verified. Treat those products as API-only for this shortlist, not as open-source downloads. Max, Plus, Flash.

Z.ai: GLM-5.2 and GLM-5.3

Both releases have weights on the publisher's account. GLM-5.3 uses the same base model as GLM-5.2, with changes from post-training. Their Hub inventories show 753B parameters; active counts are not published in those cards.

GLM-5.2 is MIT-licensed; GLM-5.3 has its own license with conditions for certain large model-service businesses. 5.2 card, 5.3 card, license.

The downloadable GLM-5.3 configuration specifies 1,048,576 positions. Use that for this comparison; a hosted provider's larger window does not establish the local checkpoint's supported configuration. Configuration.

GLM-5.3-Flash is another downloadable model: its card specifies 320B total and 18B active parameters under MIT. It deserves a separate deployment evaluation, rather than inheriting assumptions from GLM-5.3. Flash card.

Its downloadable configuration sets a 1,048,576-token context. Size that checkpoint separately from the main GLM-5.3 release. Flash configuration.

Moonshot: Kimi K2.6 and K3

Kimi K3 has released weights, rather than merely a hosted endpoint. Its 2.8T total parameters and MXFP4 weight format make the inference recipe essential reading before provisioning. K3 card.

K2.6 is also downloadable, with a smaller total count and a shorter context. It remains worth testing when an existing deployment already supports its architecture. K2.6 card.

K3's custom license includes separate-agreement conditions for qualifying model-service businesses and attribution conditions. Read it for your deployment; “open weights” alone does not settle commercial terms. K3 license.

Mistral: Small 4 and Medium 3.5

Both models have downloadable weights. Small 4 is a 119B MoE model, despite “Small” in its name; Medium 3.5 is dense. Neither name is a consumer-memory sizing guide. Small card, Medium card.

Small 4 uses Apache-2.0. Medium 3.5 uses modified MIT with a company-revenue restriction, so it should not be described as unrestricted MIT or Apache-2.0. Medium license.

Choose between them by testing your reasoning and coding workload alongside the license requirements. A dense and a sparse model with similar total counts can have very different serving characteristics.

Meta: Llama 4

Scout and Maverick are downloadable multimodal models under the Llama 4 Community license. Scout's advertised 10M context is a reason to investigate it for long documents, not evidence that every local runtime supports that window.

Both activate 17B parameters, but their total sizes differ sharply. Budget for the full checkpoint and cache. Llama 4 card and license.

OpenAI: gpt-oss

For a text-only local reasoning assistant, gpt-oss provides clear deployment starting points: the publisher documents 16 GB for 20b and an 80 GB GPU for 120b with MXFP4. Publisher README.

Use its Harmony formatting and a compatible runtime. The model names round their sizes; the cards' actual counts are shown in the table. Context is 131,072 in both configurations. 20b config, 120b config.

Google: Gemma 4

Gemma 4 offers E2B, E4B, 12B, 26B A4B and 31B releases under Apache-2.0. The smaller E models have 128K context; the other sizes have 256K. Family card.

The E labels describe effective parameters, not complete weight storage. Include embeddings and any modality encoders. Gemma is my starting family when device constraints matter more than testing the largest available checkpoint.

How to choose

1. Set the deployment boundary. Decide whether weights must stay on your workstation, an internal server or a managed host before comparing models.
2. Select the exact checkpoint. Record its publisher, revision, license and quantization. A family name is too broad for approval or reproducibility.
3. Test completed work. Score correct patches, grounded answers and valid tool calls. Track retries, elapsed time and peak memory alongside answer quality.
4. Buy capacity after testing. Extra memory should solve an observed constraint, such as context or concurrency, rather than an assumed quality gain.

Use the model picker to narrow the workload and the release timeline to find candidates. Keep a small set of representative tasks so a newer release has to earn its place.

For document work, include questions with known answers and questions the documents cannot answer. For coding, require a patch that passes tests. Evaluate the artifact you intend to deploy, including its quantization and tool interface.

Choose acceptance criteria before comparing outputs. Otherwise, an impressive demonstration can distract from a recurring failure in the ordinary work that motivated the deployment.

Running them locally

Start with a quantized build, short context and one conversation. Confirm runtime support for the architecture and modalities, then increase context while watching memory. Downloadability and easy deployment are separate questions.

Follow the local LLMs guide for setup and use the VRAM calculator for planning. Validate estimates against a real load.

FAQ

What is the best open-source LLM for a home computer?

My first workstation pick is Qwen3.8-27B at Q4_K_M when it fits with headroom. For smaller machines, evaluate Gemma 4's smaller releases. Choose from your available memory before choosing from benchmark tables.

Is Kimi K3 open source or API-only?

Kimi K3 has downloadable weights and a custom license. It belongs in an open-weight comparison, but should not be presented as having an unrestricted MIT or Apache license.

Are all Qwen and Mistral models Apache-2.0?

No. Qwen3.8-27B and Mistral Small 4 use Apache-2.0; Qwen's largest 3.8 release and Mistral Medium 3.5 use different terms. Check the precise repository's license.

Does a low active parameter count mean low VRAM use?

No. Active parameters describe computation per token. Resident experts, embeddings, encoders, caches and runtime allocations determine memory needs. Use the full artifact and serving configuration when sizing hardware.

Sources

• Qwen3.8-27B card
• Qwen3.8-27B Ollama package
• Qwen3.8-2.4T-A95B card
• Qwen3.8-Flash-Next card
• Qwen3.7 Max, Plus, Flash
• DeepSeek V4 card and download table
• DeepSeek V4.1 Flash card
• GLM-5.2 card
• GLM-5.3 card, configuration, license
• GLM-5.3-Flash card, configuration
• Kimi K2.6 card
• Kimi K3 card, license
• Mistral Small 4 card
• Mistral Medium 3.5 card, license
• Llama 4 card and license
• gpt-oss-20b card, configuration
• gpt-oss-120b card, configuration, publisher README
• Gemma 4 26B A4B card
• Gemma 4 31B and family card, memory requirements

Last reviewed and updated: September 2026.