Chasing the Frontier Open-Weight AI Models in 2026
Fifteen cents per million input tokens. That is DeepSeek’s off-peak API price for V4.1-Flash. Surely something that cheap must be second-rate? Not in 2026. DeepSeek V4, Qwen 3.8, GLM-5.2, and Kimi K3 match or beat proprietary models on some developer-reported benchmarks. The catch: once you download the weights, the hardware bill and the license terms are yours.
Public-source analysis · Edited by Dandan Dad · Editorial policy
- DeepSeek
- Qwen
- Mistral
- GLM
- Llama
- Open-weight models now match or beat proprietary models on some developer-reported benchmarks. Independent evaluations can tell a different story.
- Through an API they usually cost less per token than leading proprietary models. Run them yourself and the token bill disappears, but GPU and staffing costs take its place.
- Licenses differ model by model. Even within Qwen 3.8, the 27B model is Apache 2.0, while the Max-class open model carries a custom license whose conditions kick in as a business grows.
Major Open-Weight Model Comparison
Open-weight means the trained weights are published, so you can download them and run the model on your own hardware. Whether you may use it commercially or redistribute it depends on each model’s license. The table lists major models with public weights; licenses follow each released model’s model card (checked October 6, 2026).
Another one is on the way. On October 6, Mistral opened a preview API for Mistral Large 4, a model of about 1 trillion parameters, and said the weights will follow by the end of the month. Until then, Large 3 in the table is Mistral’s newest large model you can download.
| Model | Developer | Parameters (total/active) | Context | License |
|---|---|---|---|---|
| DeepSeek V4-Pro | DeepSeek | 1.6T / 49B | 1M | MIT |
| DeepSeek V4-Flash | DeepSeek | 284B / 13B | 1M | MIT |
| Qwen3.8-27B | Alibaba | 27B (dense) | 262K | Apache 2.0 |
| GLM-5.2 | Z.ai | 753B (MoE) | 1M | MIT |
| Kimi K3 | Moonshot AI | 2.8T / 104B | 1M | Kimi K3 License (custom) |
| MiniMax M3 | MiniMax | ~428B / 23B | 1M | MiniMax Community License (custom) |
| Mistral Large 3 | Mistral AI | 675B / 41B | 256K | Apache 2.0 |
| Muse Glimmer | Meta | ~29.6B (dense) | 131K | Apache 2.0 |
Similar names, different models
- Qwen 3.8-Max (Alibaba): the official hosted version on Alibaba’s Qwen Cloud API. The weights of its base model, Qwen3.8-2.4T-A95B (2.4T total / 95B active), have been public since August 12, 2026, but that open release is text-only and comes under a custom license (the Qwen3.8-Max License). The API version adds image and video input and a 1M-token default context.
- Muse Spark (Meta): the proprietary model behind the Muse app. Its weights are not public, and it is a different model from the open Muse Glimmer.
Cost: Using an API vs. Running It Yourself
The same model can come with two very different bills. Call the developer’s API and you pay per token. Download the weights and host them yourself, and you pay for hardware and people instead.
| Via API | Input / output (per 1M tokens) |
|---|---|
| DeepSeek V4.1-Flash (replaces the retired V4-Flash API) | $0.15 / $0.60 2x at peak hours |
| DeepSeek V4-Pro | $0.66 / $1.98 2x at peak hours |
| Mistral Large 3 | $0.50 / $1.50 |
| GLM-5.2 | $1.40 / $4.40 |
| Kimi K3 | $3.00 / $15.00 |
| Qwen 3.8-Max (official API version) | $2.00 / $6.00 |
Official API list prices as of October 6, 2026; cache discounts not included.
Costs of self-hosting
- No per-token fee, but hardware isn’t free. You pay for buying or renting GPUs, power, and the time of whoever runs the servers.
- Model size changes the hardware dramatically. The ~30B Muse Glimmer targets 24–32 GB machines at 4-bit, while the 1.6T DeepSeek V4-Pro needs datacenter-class GPU servers.
- Low volume usually favors an API; high volume or strict data rules can favor self-hosting. Work out the real cost from your throughput and hardware utilization.
DeepSeek V4: The Price-Performance Standard
DeepSeek released the V4 series weights under the MIT license, and its API prices are well below those of leading proprietary models. By DeepSeek’s own measurements, V4-Pro scored 80.6% on SWE-bench Verified and V4.1-Flash 90.6% on Terminal-Bench 2.1. Both are DeepSeek’s numbers, and independent evaluations can differ.
Meta: Big Models Closed, Small Models Open
Meta released Llama 4 Scout and Maverick in April 2025, but the largest model, Llama 4 Behemoth, never shipped. In April 2026 it launched Muse Spark, a proprietary model whose weights stay private. That did not end Meta’s open releases. On August 10, 2026, it published the weights of Muse Glimmer, a model of about 30 billion parameters, under the Apache 2.0 license. Meta now plays both sides: its big models stay closed, while a model sized for personal hardware is open.
Muse Glimmer: Meta’s Open Model for Your Own Hardware
Muse Glimmer, from Meta Superintelligence Labs, is a dense model of about 29.6 billion parameters, including a ~1.8B-parameter vision encoder. Dense means the whole model works on every token, unlike the mixture-of-experts (MoE) design behind most of the big models in the table, which sends each token to a few “experts.” Glimmer reads text and images, answers in text, and handles about 131K tokens of context by default. Meta says it trained Glimmer by distilling Muse Spark’s outputs. Commercial use is allowed under Apache 2.0, subject to the license’s conditions.
When sizing hardware, separate the size of the weight file you download from the memory needed to run it. At runtime the KV cache (which grows with conversation length), the vision encoder, and the speed-up drafter model all take memory on top of the weights.
| Format | Weight file | Runtime memory guidance |
|---|---|---|
| BF16 original | ~58 GB (Meta: over 55 GB) | 64 GB VRAM class (Meta’s target hardware) |
| 4-bit K-Quant-Dynamic (Meta) | Under 20 GB | 32 GB VRAM class · 0.2% average degradation |
| 4-bit K-Quant-17GB (Meta) | ~17 GB | 24 GB VRAM class · 1.0% average degradation |
| Unsloth GGUF, 2–8 bit | Varies by bit width | From 12–14 GB (2-bit) to 34 GB+ (8-bit); more for long contexts |
Meta measured degradation as the average accuracy change across 15 benchmarks. Unsloth counts total memory (RAM plus VRAM, or unified memory) and says available memory should exceed the downloaded file, with more needed for long contexts. So “about 17 GB” is the weight file, not the total memory needed to run the model.
- Speed: Glimmer ships with a DFlash drafter that proposes 16-token blocks for the main model to verify. In Meta’s measurements, an RTX 5090 went from 74.9 to 233.4 tokens per second (about 3.1x), an Apple M4 Max 1.5x, and an M5 Max 1.8x. AMD reports up to 24 tokens per second on Ryzen AI Max+ 395 and up to 53 on Radeon AI PRO R9700. All are vendor measurements.
- Performance: In Meta’s comparison with the similarly sized Gemma4-31B and Qwen3.6-27B, Glimmer led on some agent tasks such as MCP Atlas (75.5), while Qwen3.6-27B led on SWE-bench Verified and OSWorld-Verified (Meta’s figures). Four days after Glimmer’s release, Alibaba’s model card for Qwen3.8-27B showed it ahead of Glimmer on Terminal-Bench 2.1 (73.0 vs 51.7) and OSWorld-Verified (84.3 vs 65.9), by Alibaba’s own measurements. Among small models, the lead can change within days.
- Safeguards: Meta’s model card recommends deploying Glimmer inside a system with extra guardrails, such as human confirmation for irreversible actions, rather than on its own.
Licenses by released model (model cards, checked October 6, 2026)
- MIT: DeepSeek V4-Pro, V4-Flash and V4.1-Flash, GLM-5.2. Commercial use allowed; keep the copyright and permission notices.
- Apache 2.0: Qwen3.8-27B, Mistral Large 3, Muse Glimmer. Commercial use allowed; include the license and note your changes, among other conditions.
- Custom licenses: These generally allow commercial use, but the strings tighten as a business grows. The Kimi K3 License requires products above 100 million monthly users or 20 million US dollars in monthly revenue to display “Kimi K3.” Model-as-a-service (MaaS) businesses, meaning companies that let others call the model through an API, need a separate agreement once their revenue tops 20 million US dollars over 12 months. The Qwen3.8-Max License on Qwen3.8-2.4T-A95B has the same name-display rule and requires a separate license for MaaS or AI coding and office-assistant businesses above $50 million in 12-month revenue. The GLM-5.3 License requires MaaS businesses above $10 billion in 12-month revenue to pass a Z.ai security review. The MiniMax Community License asks more: commercial users must display “Built with MiniMax M3,” and products above $20 million in yearly revenue need prior written authorization.
Licenses can split even within one company’s lineup. GLM-5.2 is MIT, but GLM-5.3, a post-trained update of the same base model, comes under its own GLM-5.3 License. Check the license of the exact model you download, not the family name.
- Most open-weight benchmark scores are self-reported. Test setups vary, so recheck them against independent evaluations and your own workload.
- Running a model on your own server does not by itself keep data in-house. If web search, external tools, logging, or telemetry are enabled, data can still leave. Check local execution and outbound data settings separately.
Sources
- Progressive Robot · Open-Weight Guide 2026
- GitHub · Qwen 3.8
- AI Release Tracker · DeepSeek
- Meta Research · Introducing Muse Glimmer (Aug 10, 2026)
- Hugging Face · Muse Glimmer 30B model card
- Unsloth · Muse Glimmer: how to run locally
- AMD · Muse Glimmer 30B on Ryzen AI Max and Radeon
- Hugging Face · DeepSeek-V4-Pro model card (MIT)
- DeepSeek API Docs · Models & Pricing (checked Oct 6, 2026)
- Hugging Face · GLM-5.2 model card (MIT)
- Hugging Face · Qwen3.8-27B model card (Apache 2.0)
- Hugging Face · Kimi K3 License
- Hugging Face · MiniMax-M3 model card
- Hugging Face · MiniMax Community License (M3)
- MiniMax API Docs · Pay as You Go (checked Oct 6, 2026)
- Hugging Face · Qwen3.8-2.4T-A95B model card (Qwen3.8-Max License)
- Qwen Cloud · Qwen3.8-Max pricing (checked Oct 6, 2026)
- Hugging Face · GLM-5.3 License
- Z.ai API Docs · Pricing (checked Oct 6, 2026)
- Kimi API Platform · K3 pricing (checked Oct 6, 2026)
- Mistral AI · Pricing (checked Oct 6, 2026)
- Mistral AI · Introducing Mistral Large 4 (Oct 6, 2026)
