本文へ

State of LLMs, July 2026: frontier, open-weight, and what actually fits on your own hardware

For readers deciding between an API call and running inference themselves -- on murakumo.cloud fleet hardware or their own machine

Two things happened at once in the ~10 weeks before this post: US frontier labs mostly iterated on cost-efficiency rather than raw capability, while Chinese open-weight labs kept scaling and shipping under permissive licenses. Below is where that landed as of 2026-07-25, split into three tables -- frontier (API-only), open-weight (self-hostable, if you have the hardware), and what's realistic on a 16GB local device, which is the question that actually matters if you're deciding whether to run something yourself on murakumo.cloud fleet hardware instead of calling an API.

Frontier (API-only)

Model Vendor Released Notable
Claude Fable 5 Anthropic Top widely-available Anthropic model as of this writing
Claude Opus 5 Anthropic 2026-07-24 Cheaper, near-Fable-5 capability; new Max/Pro default
Claude Sonnet 5 Anthropic 2026-06-30 API launch
GPT-5.6 (Sol/Terra/Luna) OpenAI 2026-07-09 Sol: 88.8% Terminal-Bench 2.1, $5/$30 per 1M tok; now ChatGPT default
Grok 4.5 xAI 2026-07-08 1.5T-param MoE, Cursor-trained coding focus; 83.3% Terminal-Bench 2.1 at ~25% Opus 4.8's output tokens; $2/$6 per 1M tok
Gemini 3.6 Flash / 3.5 Flash Lite Google DeepMind 2026-07-21 Both shipped same day

Open-weight (self-hostable, if you have the hardware)

Model Vendor Size License Notable
GLM-5.2 Zhipu / Z.ai 753B total / ~40B active MoE MIT NIST/CAISI: roughly on par with GPT-5.2 overall; tops open-weight SWE-bench Pro at 62.1
Kimi K2.6 Moonshot AI 1T total / 32B active MoE Modified MIT Multimodal, 262K context; ties GPT-5.5 on SWE-Bench Pro (58.6%)
Kimi K3 Moonshot AI 2.8T total Modified MIT Announced 2026-07-16, largest open-weight model to date; full weights promised 2026-07-27; #1 on Arena's blind Frontend Code eval, ahead of Fable 5
Ornith-1.0 (flagship) DeepReinforce 397B open-source Self-scaffolding RL training; 82.4 SWE-Bench Verified, ~matches Claude Opus 4.7
KAT-Coder-Pro V2 Kwaipilot 256K context, coding/agentic specialist, $0.30/$1.20 per 1M tok

None of the above run on a single consumer machine -- even the smallest (Ornith-1.0's 397B) needs multi-GPU server hardware. They're the models a murakumo.cloud fleet node with enough GPU memory could serve; they are not laptop-runnable.

What actually fits a 16GB local device

If the question is "what's the best model I can run on my own machine" -- a 16GB unified-memory Apple Silicon box, or comparable -- the frontier and open-weight tables above are irrelevant. Realistic budget after OS overhead is roughly 9-10GB for weights + KV cache, which at Q4 quantization (~0.6GB/B params) caps you at dense models around 14B.

Model Params Q4 footprint Fit
Qwen3.6-14B 14B dense ~9.5GB Largest dense model that fits comfortably
Phi-4 14B dense, MIT ~9.5GB Strong reasoning for its size; good dev-assistant pick
Gemma 4 12B 12B dense ~6-8GB Best MLX tooling support, safer headroom
Qwen3.6 8-9B 8-9B dense ~7GB Pick this if you need RAM headroom for other apps running at the same time
Gemma 4 26B-A4B 25.2B total / 3.8B active MoE ~15GB Technically fits, near-zero headroom -- only worth it on a dedicated inference box

murakumo.cloud's read

For a genuinely dedicated local device (16GB, nothing else competing for memory), Qwen3.6-14B at Q4_K_M is the highest dense-model capability that still leaves real headroom -- Phi-4 is the closest alternative if you specifically want a smaller, MIT-licensed reasoning-focused model. If you want a faster daily driver instead of the ceiling, Gemma 4 12B is the community-favored default with the best MLX tooling.

The wider point this table makes: the gap between "what fits on a laptop" (14B-class) and "what's actually competitive with frontier APIs" (GLM-5.2, Kimi K2.6/K3, hundreds of billions of active-or-total parameters) has not closed -- if anything the open-weight scaling race of the last few months widened it. That's the concrete case for self-hosted inference on real GPU hardware rather than a laptop: a murakumo.cloud fleet node can serve the open-weight column above; your own 16GB machine can't, no matter how good next quarter's small models get.

Figures below are compiled from public search results and vendor/press announcements as of 2026-07-25, not independently re-benchmarked by us. Some entries (small open-weight releases in particular) could not be cross-checked against a primary source and are marked accordingly. Treat benchmark numbers as directional, not as a guarantee of reproduced performance on murakumo.cloud or any other fleet.