Open AI Models 2026: DeepSeek V4.1, Qwen 3.8, GLM 5.3, Mistral Large 4

• LargitData
• 2026-10-31

Updated October 31, 2026: Enterprises choosing an open-weight model today have a shortlist that includes DeepSeek V4.1-Flash-1017, Qwen 3.8, GLM 5.3, Kimi K3, Mistral Large 4, Meta Muse Spark Open, the NVIDIA Nemotron 3.5 family, Gemma 4 and Taiwan's TAIDE. This guide assesses those options through five practical questions: can you download the weights, does the license allow commercial use, is the hardware affordable, are Traditional Chinese output and tool calling reliable, and can the data stay inside the organization?

“Open source” can mean three different things

• Hosted API: quick to adopt, though data leaves your environment and the provider controls pricing, versions and retirement schedules. API access can also precede a weight release, as it currently does for Kimi K3.1 Pro.
• Open-weight model: you can download the weights and usually fine-tune the model, although the training data and full production process may remain unavailable. Most models in this guide fall into this category.
• Open-source AI: the Open Source AI Definition 1.0 requires more than downloadable parameters. Users must also have the information and code necessary to study, modify and recreate the system.

These distinctions affect what you can deploy. If inference data must remain inside a private network, a hosted model falls outside the requirements. Even a downloadable model may be unsuitable if its license restricts redistribution, derivative services or particular commercial uses. Each checkpoint therefore needs its own license review.

2026 model landscape: specifications that affect deployment

Specifications and licenses reflect the official model cards and release pages as of the update date. A model's maximum context does not guarantee the same accuracy, speed or cost at that length. Before procurement, check the official model card and license again.

How the leading model families differ

DeepSeek V4.1-Flash: a new architecture with native image input

DeepSeek-V4.1-Flash, released on September 10, introduces a new architecture rather than a post-training refresh of V4-Flash. Its Causal Encoder-Decoder design splits a 40-layer Transformer into a 20-layer causal encoder and a 20-layer decoder. The model has 552B backbone parameters plus 196B of Engram conditional memory, accessed sparsely through token-based lookup. Only about 8B parameters are active per token in prefill and 16B in decode. A DeepSeek-ViT encoder provides native support for image and text input. The model supports 1M context, and its weights are released under the MIT license.

DeepSeek-V4.1-Flash-1017, released on October 17, is a minor update that improves reasoning, coding, agent and visual benchmarks while reducing task time and output length. The deepseek-flash endpoint now serves this version with up to 384K output. The old deepseek-v4-flash name still works, but it temporarily routes requests to the new version and can no longer be used to pin the older model. For self-hosting teams, the key trade-off is memory: per-token compute is low, but all 763B of weights must still fit.

Cost assessment should still focus on agent economics: the turns, retries and total tokens needed to complete a task. DeepSeek moved to peak and off-peak pricing on August 16 (UTC). Peak hours are weekdays 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays; all other hours, including weekends, are off-peak at half price. On the update date the official pricing page listed deepseek-flash cache-miss input at US$0.15 off-peak and US$0.30 peak per million tokens, with output priced at US$0.6 and US$1.2, respectively. Search and Fetch tools cost US$0.75 per 1,000 successful requests. Check the official pricing page for current rates. DeepSeek-V4-Pro-0813 remains the MIT-licensed Pro release for harder reasoning and long-running work.

Qwen 3.8: 27B remains a practical local choice; Flash-Next is a preview

Qwen3.8-27B is a dense vision-language model released under Apache 2.0. It handles text, images and video, offers a 262,144-token native context extendable to 1M, and supports adjustable reasoning. Qwen3.8-Flash-Next, added in late August, is a 125B-backbone MoE with 6B active parameters (plus 51B of n-gram embeddings and 4B of MTP modules) and a Gated DeltaNet plus Qwen Sparse Attention hybrid. Qwen describes it as an experimental preview of the Qwen4 architecture. It is released under the custom qwen-community-1.0 license. Qwen3.8-2.4T-A95B makes Max-class weights available under the qwen3.8-max license. No Qwen 3.9 or Qwen4 weights had been published by the update date.

GLM 5.3: public weights, with licenses that differ by size

Zhipu launched GLM 5.3 as a service in mid-August and published the weights in late August. The roughly 753B text-only model targets complex software engineering, terminal agents and long-running work. Its reasoning effort can be set to low, high or max. It uses the custom GLM-5.3 License rather than MIT. Commercial use, modification and redistribution are allowed, but a licensee that runs a model-as-a-service business and has revenue, including affiliates, exceeding US$10 billion in any 12 consecutive months must pass a Z.ai security review before any commercial use. GLM-5.3-Flash (320B / 18B active) is MIT-licensed and the first natively multimodal GLM-5 model.

Kimi K3 and K3.1 Pro: know which one you can download

Kimi K3 is a native multimodal MoE with 2.8T / 104B-active parameters and 1M context, released under a custom Kimi K3 license. Moonshot launched K3.1 Pro and K3.1 Pro Highspeed (described as 3T-class) on October 23 through its products and API. The company says full weights will follow by November 4. Until the weights and license are published, treat K3.1 Pro as an API model and base local capacity planning on K3 alone.

Mistral Large 4: Europe's trillion-scale open-weight flagship

Mistral launched Large 4 as an API preview on October 6 and released the weights on October 29. According to the official model card, it has about 1T total parameters, with about 49B active per token (about 52B including embedding and output layers). It accepts text and image input, supports 1M context and claims fluency in more than 160 languages. The reference deployment uses vLLM with 8-way tensor parallelism, and an NVFP4 build targets Blackwell GPUs. The Mistral Ecosystem Open-Weight License replaces Large 3's Apache 2.0. This modified MIT license permits commercial use but requires a separate commercial license for a model-as-a-service business.

Meta Muse: Glimmer for local devices, Spark Open for the data center

Muse Glimmer 30B is a roughly 29.6B dense model (including a 1.8B vision encoder), released under Apache 2.0. It supports 131K+ context, tool use and adjustable reasoning. Its official 4-bit releases target 24GB/32GB devices. Muse Spark Open, released on October 15, is an 816B / 42B-active MoE with 1M context and support for text, image, video and audio input. Meta recommends at least 700GB of GPU memory for NVFP4 and about 2TB for BF16. Its Muse Model License requires a visible “Built with Muse Spark Open” notice. Licensees whose revenue, including affiliates, exceeds US$20 million in any 12 consecutive months must notify Meta and sign a separate agreement before offering it as a service. Downloads also require an application form.

NVIDIA Nemotron 3.5: a full range from edge devices to 120B

Nemotron 3.5 Lightning 30B-A3B remains the efficiency-focused option for high-frequency tool calls and low-latency agents. In October, NVIDIA expanded the range with Nemotron 3.5 Super VL (120B / 12B active, 1M context) and Nemotron 3.5 Bolt 74B-A9B, compressed from Super VL. Bolt accepts text, images and video. Its official 24GB build uses 2-bit (IQ2) expert weights trained with quantization-aware distillation and is designed to stay under 24GB at batch size 1 and 128K context. Bolt E4B is an 8B (4B effective) text model post-trained from Gemma-4-E4B-it for the dialogue layer of local voice and device assistants. Lightning, Super VL and Bolt use OpenMDW-1.1. The family also includes Nemotron 3.5 Embed and Rerank, Parse and Transcribe; check their licenses individually. Nemotron 3 Ultra remains the 550B / 55B-active coordinator.

Smaller models are often a better fit for production

• Qwen3.8-27B: balances Chinese, vision, coding and agent capabilities under Apache 2.0. For Taiwan public-sector work, check the source-country policy first.
• Nemotron 3.5 Bolt 74B-A9B (24GB build): multimodal input; the model supports up to 256K context, and the official 24GB build targets under 24GB at batch size 1 and 128K context. Validate 2-bit quality on your own tasks.
• Nemotron 3.5 Lightning: offers strong throughput and integration with NVIDIA's deployment tools.
• Muse Glimmer 30B: suits local image-and-text agents on consumer devices.
• Gemma 4 E2B / E4B / 12B / 26B MoE / 31B: covers edge devices through workstations under Apache 2.0.
• gpt-oss-20b / 120b: Apache 2.0 text reasoning with tool and structured-output strengths, plus gpt-oss-safeguard variants for policy classification.
• Ministral 3 and Mistral Small 4: Apache 2.0 Ministral 3 offers compact 3B, 8B and 14B options for edge and multilingual use; Mistral Small 4 (119B / 6B active per token, about 8B including embeddings, Apache 2.0) is the step up.
• Gemma-3-TAIDE-12B-Chat-2602: focuses on Taiwan text and Traditional Chinese office workflows.

For RAG, summarization, classification, field extraction and customer-service drafts, a mid-sized model paired with a well-maintained knowledge base is often more cost-effective than a flagship used on its own.

Gemma 4 and TAIDE for Taiwan evaluations

Gemma 4 is released under Apache 2.0, with a range spanning E2B, E4B, 12B, 26B MoE and 31B Dense. Gemma 4 12B projects image patches and audio directly into the language model's embedding space without a separate encoder and supports 256K context. It has also become a base for other models: NVIDIA's Nemotron 3.5 Bolt E4B is post-trained from Gemma-4-E4B-it.

Gemma-3-TAIDE-12B-Chat-2602 is based on Gemma 3 rather than Gemma 4 and focuses on Taiwan's language, office tasks and local usage. TAIDE says it was not trained on image data, recommends text-only input and specifies a 128K context. Both this model and the TAIDE embedding model below use the Gemma-version TAIDE model license, which must be accepted before download. TAIDE has also published embeddinggemma-GTAIDE-300m-2605, an EmbeddingGemma-based model fine-tuned to retrieve Taiwan regulations. TAIDE reports that Recall@1 rose from 56.14 to 74.43 on its internal test. Google also launched EmbeddingGemma 2 on October 6 under Apache 2.0, with 740M parameters, 8K context and a shared embedding space for text, images, audio and video. Test agency names, Taiwan laws, Minguo/Gregorian dates, addresses, currency, full-width punctuation and cross-strait vocabulary in the same suite as Gemma 4, Qwen3.8-27B, Muse Glimmer and Nemotron Bolt.

Other August–October releases worth tracking

• MiniMax-M3.1-Flash: about 428B total / 23B active, with text, image and video input and 1M context. API access began in October, followed by weights on Hugging Face. The MiniMax Community License requires a “Built with MiniMax M3.1-Flash” notice for commercial use, prior written permission above US$20 million in related revenue, and a one-time notice below that.
• Xiaomi MiMo-V2.6: the Flash model is 309B / 15B active with text, image, video and audio input and 1M context under MIT; a larger Pro model is also available.
• IBM Granite 4.2: 3B, 8B and 30B models released on August 25 under Apache 2.0. Their 128K native context can be extended to 512K, and Chinese is among the tested languages.
• Liquid AI LFM3-1.7B: 1.73B parameters, 256K context and about 1GB at 4-bit for phones and laptops, under the LFM Open License v1.0, which excludes commercial use by organizations (including affiliates) with annual revenue of US$10 million or more unless separately licensed.

Keep generation, embedding, speech and image models in separate comparisons. MiniMax and MiMo are China-origin models, so the source-country considerations below apply to Taiwan public-sector use.

Hardware planning: active parameters are not memory requirements

Memory for the weights alone is roughly total parameters × bits per parameter ÷ 8. This excludes KV cache, encoders, runtime overhead and concurrency. DeepSeek V4.1-Flash illustrates the distinction: only 8B–16B parameters are active per token, but all 763B must be loaded.

LM Studio, Ollama, vLLM and SGLang serve different stages

LM Studio and Ollama simplify local experimentation. Ollama also offers cloud models, however, so using it does not by itself keep data on the device. For multi-user throughput, long context and production serving, vLLM and SGLang are the relevant tools. The official cards for DeepSeek V4.1, GLM 5.3, Mistral Large 4 and Nemotron 3.5 all include vLLM deployment instructions. New architectures such as DeepSeek V4.1's encoder-decoder and Qwen3.8-Flash-Next's hybrid attention usually need recent engine versions, so validate upgrades in staging. Each engine offers some controls, such as API keys or model revisions, but full enterprise identity, authorization, rate limits, audit logs, content safety and failover still have to be added for production.

Build your own evaluation set

Build a set of 100–300 representative tasks to assess accuracy, citation quality, extraction F1, test pass rate, Traditional Chinese performance, tool/schema reliability, P50/P95 latency, prompt injection resistance and one-year TCO. Repeat each item at least three times, recording the model ID, weight hash, quantization, inference engine and sampling settings. When a provider routes old endpoint names to new versions, as DeepSeek has done, unpinned API results can change without notice.

China, the US and Europe are pursuing different open-model strategies

The idea that “China is open, America is closed” no longer describes the market accurately. Chinese developers compete through very large MoE models, one-million-token contexts, frequent date-stamped releases and aggressive API pricing, and DeepSeek has now changed its underlying architecture with V4.1. Their licenses range from MIT and Apache 2.0 to custom terms.

In the US, Gemma 4, gpt-oss and Nemotron offer models in deployable sizes, while Meta has added an 816B open-weight flagship with Muse Spark Open. In Europe, Mistral Large 4 joins the trillion-scale tier with a multilingual, self-hosting focus. Licensing is increasingly based on “commercial use allowed, with conditions”: GLM 5.3, Muse Spark Open, MiniMax and LFM3 set revenue thresholds, Mistral Large 4 requires a separate license for a model-as-a-service business, and Muse and MiniMax add attribution requirements. Because those conditions differ, legal teams need to review each checkpoint before it is packaged as a service or redistributed.

Jensen Huang's first post on X: open models strengthen safety, innovation and sovereignty

On July 24, 2026, NVIDIA CEO Jensen Huang used his first post on X to share a letter NVIDIA had signed on why open models matter. He argued that AI will transform every industry, power every company and be built by every country, while open models strengthen safety and cybersecurity, accelerate innovation and the spread of technology, and support sovereignty. The post linked to the joint Open Weights and American AI Leadership letter, in which NVIDIA, Meta, Google, OpenAI, Microsoft, Hugging Face, the Linux Foundation and others argued that open weights reduce barriers, increase competition, limit vendor lock-in and preserve customer control over data and deployment. In the following months Meta, one of the signatories, released Muse Spark Open, and NVIDIA continued to expand the Nemotron 3.5 family.

Data sovereignty and procurement policy matter for Taiwan enterprises

On-premises inference does not automatically make a system secure. Verify weight sources and hashes, audit remote code and third-party quantizations, disable unnecessary telemetry, restrict outbound traffic and maintain a software bill of materials.

Taiwan's Ministry of Digital Affairs states that government agencies may not use DeepSeek services, apps or locally downloaded models. No withdrawal of that rule had been announced by the update date. For government procurement, the Public Construction Commission amended the standard contract clauses on October 6, 2026 so that the prohibited-product requirements apply to the performance and deliverables of all procurement types, not only contracts with written deliverables. Private-sector use is not categorically prohibited, but government, critical-infrastructure and government-contracting teams should obtain written security and procurement approval before a PoC and keep alternatives such as Gemma, gpt-oss, Nemotron, Mistral, Muse and TAIDE ready.

A practical adoption sequence

1. First, define the constraints on data location, source country, licensing and hosting.
2. Use one or two hosted flagships to establish the quality ceiling.
3. Establish a local baseline by comparing Qwen3.8-27B, Nemotron Bolt and Lightning, Muse Glimmer, Gemma 4, gpt-oss, Ministral and TAIDE under the same conditions.
4. Review each checkpoint's license, since sizes from the same vendor can have different terms, as GLM 5.3 and GLM-5.3-Flash do. Record the license version and download date.
5. Improve RAG, including the embedding model, before fine-tuning when knowledge changes or citations matter.
6. Load-test real context, output lengths and concurrency before purchasing hardware.
7. Use a model gateway for routing, rollback, cost control, logs and security policy.

Most enterprises ultimately adopt a hybrid architecture. Local models handle sensitive documents and high-frequency routine tasks, while hosted flagships process public information and occasional difficult reasoning tasks. LargitData's RAGi enterprise knowledge platform can be deployed on premises and brings retrieval, model integration and governance into one architecture.

Enterprise selection FAQ for 2026

Conclusion

Model releases accelerated in autumn 2026, outpacing the summer: DeepSeek changed architectures, GLM 5.3 published weights, Mistral and Meta released trillion-scale and near-trillion-scale open-weight flagships, and NVIDIA fit a 74B multimodal model into 24GB. The strongest strategy is still to use hosted flagships as a quality ceiling, deployable open-weight models as the production baseline, and your own evaluation set as the basis for decisions. Qwen3.8-27B, Nemotron Bolt and Lightning, Muse Glimmer, Gemma 4 and TAIDE are practical local candidates; DeepSeek V4.1, GLM 5.3, Kimi K3, Mistral Large 4 and Muse Spark Open require server clusters; Kimi K3.1 Pro remains pending the release of its weights and license. Governance, retrieval, permissions and monitoring matter as much as the model name.

Primary sources: DeepSeek V4.1-Flash, DeepSeek V4.1-Flash-1017, DeepSeek API pricing, Qwen3.8-27B, Qwen3.8-Flash-Next, GLM-5.3, GLM-5.3-Flash, Kimi K3, Kimi K3.1 Pro, Mistral Large 4, Muse Spark Open, Muse Glimmer 30B, Nemotron 3.5 Bolt, Nemotron 3.5 Lightning, Gemma 4 12B, Gemma 4, gpt-oss, MiniMax-M3.1-Flash, MiMo-V2.6-Flash, Granite 4.2, LFM3, EmbeddingGemma 2, TAIDE and TAIDE embedding.