The open-weight AI landscape has shifted dramatically in recent weeks. Three major releases — DeepSeek V4, Kimi K2.6, and Qwen 3.6 — have pushed the boundaries of what freely available models can do, each taking a fundamentally different approach to architecture and capability. If you’re tracking the space, here’s what you need to know.
DeepSeek V4: Million-Token Context with Hybrid Attention
DeepSeek’s V4 series comes in two variants: V4-Pro (1.6T total parameters, 49B activated) and V4-Flash (284B total, 13B activated). Both support a one million token context window, which is remarkable on its own. But the real innovation is how they handle it.
V4 introduces a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). The result? At the full 1M-token context length, V4-Pro needs only 27% of the inference FLOPs and 10% of the KV cache compared to V3.2. That’s not an incremental improvement — it’s the kind of architectural leap that makes long-context workloads practical instead of prohibitively expensive.
Other technical highlights include Manifold-Constrained Hyper-Connections (mHC) for better signal propagation across the model’s layers, and training with the Muon optimizer for faster convergence. Both models were pre-trained on 32T tokens with a two-stage post-training pipeline: first cultivating domain-specific experts independently, then consolidating them via on-policy distillation.
The benchmarks are striking. V4-Pro-Max achieves 93.5 on LiveCodeBench and a Codeforces rating of 3206, surpassing GPT-5.4 and Claude Opus 4.6 on coding tasks. It also hits 80.6 on SWE-Bench Verified — firmly competitive with the best closed-source models. The license is MIT, so you can use it for virtually any purpose.
Kimi K2.6: The Agentic Coding Powerhouse
Moonshot AI’s Kimi K2.6 takes a different path. With 1T total parameters and 32B activated per inference, it’s designed as a native multimodal agentic model — built from the ground up for long-horizon coding, autonomous execution, and multi-agent orchestration.
Where K2.6 really shines is in agentic workflows. It can scale to 300 sub-agents executing 4,000 coordinated steps, dynamically decomposing complex tasks into parallel subtasks. It handles frontend workflows, DevOps pipelines, and cross-platform operations autonomously. The model’s “coding-driven design” capability means it can transform visual inputs into production-ready interfaces — not just writing code, but understanding the design intent behind it.
The benchmark numbers tell a compelling story. K2.6 scores 58.6 on SWE-Bench Pro, beating GPT-5.4 (57.7) and Claude Opus 4.6 (53.4). It achieves 96.4 on AIME 2026 and 89.6 on LiveCodeBench. For a model you can download and run yourself, these are competitive numbers with the best proprietary systems. The model uses a modified MIT license and supports both text and image inputs through its MoonViT vision encoder.
Qwen 3.6: Small Model, Big Results
Perhaps the most surprising entry is Qwen 3.6-27B from Alibaba. At just 27 billion parameters — dense, not MoE — it punches well above its weight class. The architecture uses a novel Gated DeltaNet approach for linear attention, combined with traditional gated attention in a 16-block repeating pattern (three DeltaNet + FFN layers followed by one attention + FFN layer).
This hybrid linear-attention design gives Qwen 3.6 a native context length of 262,144 tokens, extendable to over 1M. The model delivers 77.2 on SWE-Bench Verified — outperforming Qwen’s own 397B MoE model (76.2) on the same benchmark. It also scores 83.9 on LiveCodeBench and 94.1 on AIME 2026. For teams that need capable coding assistance without the infrastructure requirements of a 1T-parameter model, this is an extremely practical option.
Qwen 3.6 is released under Apache 2.0, the most permissive license among the three. It’s also the most accessible for local deployment — a 27B dense model runs comfortably on consumer hardware with reasonable quantization.
Running These Models Locally
For developers looking to run these models locally, here’s the practical breakdown:
# DeepSeek V4-Pro via Ollama (quantized)
ollama run deepseek-v4-pro
# Kimi K2.6 via Ollama (quantized)
ollama run kimi-k2.6
# Qwen 3.6-27B - runs on 24GB VRAM at Q4
ollama run qwen3.6:27bFor production serving, vLLM supports all three models. The V4-Pro and K2.6 MoE architectures require more careful memory planning due to their expert parallelism, but the inference speed gains from only activating a fraction of total parameters make them efficient in practice.
What This Means for Developers
The gap between open-weight and closed-source models has effectively closed for coding tasks. DeepSeek V4-Pro leads on LiveCodeBench and Codeforces. Kimi K2.6 leads on SWE-Bench Pro. Qwen 3.6 proves that a 27B dense model can compete with models 15x its size. All three are available under permissive licenses.
For teams building AI-powered developer tools, the implications are clear: you no longer need to depend on proprietary APIs for state-of-the-art coding assistance. The models are there, the licenses are permissive, and the serving infrastructure (vLLM, SGLang, Ollama) is mature. The question is no longer whether open-source can compete — it’s which architecture best fits your specific use case.
Model weights are available on Hugging Face for all three releases: DeepSeek-V4-Pro, Kimi-K2.6, and Qwen3.6-27B.