The open-weight AI landscape has shifted dramatically in the first half of 2026. Three major releases — DeepSeek V4 Pro, Kimi K2.6, and Qwen3.5 — have pushed open-source models into territory that was exclusive to closed-source APIs just months ago. Here’s a practical breakdown of what’s new, how the models compare, and how you can run them locally.
DeepSeek V4 Pro: Efficiency Meets Scale
Released on April 24, 2026, DeepSeek V4 Pro is a 1.6-trillion parameter Mixture-of-Experts (MoE) model that activates only 49B parameters per token. The headline feature is its 1 million token context window, achieved through a novel hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). At full context length, V4 Pro requires only 27% of the single-token inference FLOPs and 10% of the KV cache compared to DeepSeek V3.2.
Where V4 Pro really shines is coding. It scores 93.5% on LiveCodeBench (Think Max mode) — the highest score across all models, open or closed — and achieves a Codeforces rating of 3206. On SWE-bench Verified, it hits 80.6%, closely matching Claude Opus 4.6’s 80.8%. The model uses FP4 quantization for expert parameters and FP8 for most other weights, which keeps memory requirements manageable despite the massive parameter count.
V4 Pro offers three reasoning effort modes: Non-think for fast routine tasks, Think High (the default) for complex problem-solving, and Think Max for pushing reasoning boundaries. The Think Max scores are the ones that top the leaderboards, but they come with higher latency and token usage. The model is released under the MIT license and available on Hugging Face.
Kimi K2.6: The Agentic Specialist
Moonshot AI’s Kimi K2.6, released April 20, 2026, takes a different approach. While it’s also a massive MoE model (1T total, 32B active), its real differentiator is agentic capability. K2.6 is designed for long-horizon tasks — think multi-step coding projects, autonomous web browsing, and swarm-based orchestration across up to 300 sub-agents executing 4,000 coordinated steps.
The benchmarks tell the story. On BrowseComp Agent Swarm, K2.6 scores 86.3, well ahead of GPT-5.4’s 78.4 (the standard BrowseComp score is 83.2 vs GPT-5.4’s 82.7). On DeepSearchQA, it hits 92.5 F1, and on SWE-Bench Pro it leads at 58.6. It uses Multi-head Latent Attention (MLA) and includes a dedicated vision encoder (MoonViT, 400M params), making it natively multimodal with a 256K context window.
For coding tasks specifically, K2.6 supports Thinking mode (default, with chain-of-thought reasoning) and Instant mode (faster, no reasoning trace). The recommended temperature is 1.0 for Thinking and 0.6 for Instant. It’s deployable via vLLM or SGLang and supports native INT4 quantization. Available on Hugging Face.
Qwen3.5: Small Models, Surprising Performance
Alibaba’s Qwen3.5 family is arguably the most practical release for developers working with constrained hardware. The family spans eight models from 0.8B to 397B parameters, all open-weight under Apache 2.0. The standout is the efficiency story: Qwen3.5-9B outperforms models over 10x its size on most language benchmarks, and the 4B version beats models 5x larger.
The flagship Qwen3.5-397B-A17B (397B total, 17B active per token) uses MoE and outperformed GPT-5.2, Claude 4.5 Opus, and Gemini-3 Pro on the majority of 44 vision benchmarks. But for most developers, the mid-size models are more relevant: Qwen3.5-27B (dense) and Qwen3.5-122B-A10B (MoE) both exceed GPT-5-mini on most benchmarks while being runnable on consumer-grade multi-GPU setups.
All models support text, image, and video input with a 256K token context window (extensible to 1M). They also support tool use, web search, and chain-of-thought reasoning across 201 languages. The one caveat: smaller Qwen3.5 models still lag behind larger competitors on reasoning and coding tasks. Available on GitHub.
How to Run These Models Locally
The local LLM ecosystem has matured significantly in 2026. Here’s how to get started with each model:
Using Ollama (Quickest Start)
Ollama remains the fastest path to running models locally. The ollama launch command (introduced in v0.15) even sets up coding assistants like Claude Code and OpenCode against local models with a single command:
# Pull and run a model
ollama pull qwen3.5:9b
ollama run qwen3.5:9b
# Launch a coding assistant with a local model
ollama launch claude --model qwen3.5:9bFor Qwen3.5-9B, you’ll need roughly 6-8GB of RAM. Kimi K2.6 and DeepSeek V4 Pro require significantly more — multiple GPUs with 80GB+ VRAM each for full precision, or quantized variants to fit on smaller setups.
Using vLLM (Production Deployment)
For serving models in production, vLLM supports both Kimi K2.6 and DeepSeek V4 Pro with optimized inference:
# Serve DeepSeek V4 Pro with vLLM
python -m vllm.entrypoints.openai.api_server --model deepseek-ai/DeepSeek-V4-Pro --tensor-parallel-size 8 --max-model-len 131072
# Serve Kimi K2.6
python -m vllm.entrypoints.openai.api_server --model moonshotai/Kimi-K2.6 --tensor-parallel-size 4 --max-model-len 65536Using llama.cpp (CPU / Single-GPU)
For running quantized versions on consumer hardware, llama.cpp remains the go-to. The Qwen3.5 models are particularly well-suited here — a Q4_K_M quantized Qwen3.5-9B runs comfortably on a single GPU with 8GB VRAM:
# Convert and quantize (using llama.cpp's convert script)
python convert_hf_to_gguf.py /path/to/Qwen3.5-9B --outfile qwen3.5-9b-f16.gguf
./llama-quantize qwen3.5-9b-f16.gguf qwen3.5-9b-q4_k_m.gguf Q4_K_M
# Run with llama.cpp server
./llama-server -m qwen3.5-9b-q4_k_m.gguf -c 8192 -ngl 99Which Model Should You Use?
The choice depends on your use case:
- Coding and software engineering — DeepSeek V4 Pro leads on coding benchmarks and has the most efficient long-context handling.
- Autonomous agents and multi-step tasks — Kimi K2.6’s agent swarm capabilities and tool-augmented performance make it the clear choice.
- Resource-constrained deployment — Qwen3.5-9B or Qwen3.5-4B deliver impressive results on consumer hardware.
- Multilingual and vision tasks — Qwen3.5’s 201-language support and vision-language architecture cover the widest range of inputs.
- Maximum reasoning depth — DeepSeek V4 Pro in Think Max mode pushes the boundaries on math and reasoning benchmarks.
The gap between open-weight and closed-source models has narrowed to the point where the choice is increasingly about deployment flexibility, cost, and privacy rather than capability. With MIT and Apache 2.0 licensing across these releases, the barrier to running frontier-class AI in your own infrastructure has never been lower.