New open AI models, and what each one needs to run
Some links on this page are referral links: we may earn a commission, at no cost to you.
The newest open models we track are Inkling 2 Mini (Nov 20, 2026), Tencent Hy4 Flash (Nov 17, 2026) and Inkling 2 Small (Nov 4, 2026). Inkling 2 Mini needs about 16 GB at Q4_K_M, the smallest consumer card that runs it is a 20 GB AMD Radeon RX 7900 XT, a Mac with 32 GB of unified memory also works. Tencent Hy4 Flash needs about 125 GB at Q4_K_M, no single consumer graphics card holds it, it fits a Mac with 192 GB of unified memory. Inkling 2 Small needs about 368 GB at NVFP4, no single consumer graphics card holds it, it fits a Mac with 512 GB of unified memory. Every row below is computed with the same engine as our cost-to-run pages, so the numbers match.
63 models, newest release first (latest Nov 20, 2026). Rental prices captured , refreshed hourly. Memory at a 16K context.
Every open model, newest first
Memory = weights + context cache + about 0.8 GB for the runtime. Smallest GPU and Mac are the smallest that run the 4-bit size with 10% to spare (Macs: 8 GB left for macOS first). Rent is the card with the lowest typical price that fits: typical is the median of live offers (from a provider with 3+ offers or a list price), the cheapest listing is one offer.
We earn a commission if you buy through this link.
~2-bit is Q2_K, which we size on the safe side (real files of big models come out a little smaller). Models released at 4-bit (MXFP4, NVFP4, FP4) show their own download in the 4-bit column and nothing at 2-bit or 8-bit: a bigger copy stores the same numbers in more bytes, and how far sub-4-bit builds shrink depends on who made them. FP8 releases show their own weights at 8-bit. Hybrid-attention models (Qwen3.5 and newer, Gemma) count only their full-attention layers for the context cache. How we size memory: methodology.
What changed in the newest releases
From each official model card, and runtime support from the Ollama library, LM Studio, Hugging Face (GGUF and mlx-community builds) and vLLM, checked by hand on the date shown. "Cloud only" means the runtime lists it as hosted inference, not a download you can run at home.
A looped MoE transformer. The 24 stored layers include a middle block of 12 that runs three times with shared weights, so each token gets 48 layer passes (3.4B active, about 5.4B effective). 26.4B total, 8 of 128 experts per token, hybrid local and global attention, text and image input, up to 1M context, shipped as one mixed NVFP4 and MXFP8 checkpoint of about 20 GB. The card makes no comparison with an earlier Inkling model. model card
Compared with Hy4 the card lists three architecture changes. Hybrid attention (sparse IndexPool and 512-token sliding-window layers at 1 to 3), LatentMoE (16 of 512 routed experts work in a smaller latent space, plus one shared expert) and Engram n-gram embeddings (learned bigram and trigram tables read by hash lookup). 220B total including 64B of Engram tables, 12B active, text and image input, 1M context and a built-in DFly2 drafter for speculative decoding. On Tencent's WorkIndex it scores 42.1 against 39.7 for Hy4 preview at about a thirteenth of the cost per task. model card
A 48-layer MoE that routes each token to 8 of 384 experts, 594B total and 17B active, with hybrid local and global attention, text and image input and up to 1M context. Pretrained, midtrained and post-trained in NVFP4, which the card calls its native precision; the MXFP8 copy scores only 1 to 2 points higher on evals. The card makes no comparison with an earlier Inkling model. model card
Builds on MiniMax-M3 with sparse attention in every layer, quantization-aware training and DSpark speculative decoding, which the card credits for better task scores and faster inference. About 428B total and 23B active parameters, text, image and video input, up to 1M context. model card
The language backbone starts from Nemotron 3 Super Base. NVIDIA added a C-RADIO vision encoder for image and video input, a shared MTP head for speculative decoding, distillation from Nemotron 3 Ultra Base and a context extension to 1M tokens. Still 120B total and 12B active, a Mamba-2 and Transformer hybrid MoE. model card
A new causal encoder-decoder design with 8B active parameters per token in prefill and 16B in decode. CSA2 sparse attention and an FP4 KV cache cut the global KV cache to about a quarter of DeepSeek V4 Flash. Natively multimodal (image and text) with a 1M context; V4 Flash had 284B parameters and 13B active. model card
Same base model as GLM-5.2; every change is post-training. Z.ai reports stronger complex coding and long agent tasks (50 percent better than GLM-5.2 on its in-house Z.ai Code Bench). Text only. Not the same model as GLM-5.3-Flash. model card
An experimental preview of the architecture planned for Qwen4. Gated DeltaNet is now paired with Qwen Sparse Attention (block-level sparse attention), with gated residual streams and an n-gram embedding added. 125B total and 6B active, plus 51B of n-gram embedding and a 4B MTP head. The card names no direct predecessor. model card
Ornith-1.0 was built on Qwen3.5 and Gemma 4 with continued pretraining and post-training. Ornith-1.5 widens the self-improvement loop to jointly optimise task generation, scaffold construction and solution rollouts, and trains with reinforcement learning on tasks it generates itself. model card
30B total and 3B active, a Mamba-2, MoE and attention hybrid, released with DSpark, MTP and DFlash speculative decoding options and up to 1M context. In the card's own table it beats Nemotron 3 Nano on SWE-bench Verified (51.6 vs 34.1), Terminal-Bench 2.1 (24.6 vs 8.3) and AA-LCR long context (52.0 vs 32.8), and is about even on IFBench (71.9 vs 72.2). The card lists no architecture changes against Nano. model card
Built on the Qwen3.5 architecture (dense, 3 Gated DeltaNet layers per Gated Attention layer). The card reports gains in coding, research and long agent tasks and broader agent-harness support. Thinking is on by default and can be turned off per request. The card gives no point-by-point comparison with Qwen3.6 27B. model card
A new architecture built on Kimi Delta Attention and Attention Residuals, with a mixture of experts that activates 16 of 896 experts. 2.8T total and 104B active, native vision, a 1M-token context, and weights trained to MXFP4. The card claims about 2.5 times the scaling efficiency of Kimi K2. model card
Of the open-weight models we track, the newest is Inkling 2 Mini, released Nov 20, 2026: 26.4B parameters with 3.4B active per token. It needs about 16 GB at Q4_K_M, the smallest consumer card that runs it is a 20 GB AMD Radeon RX 7900 XT, a Mac with 32 GB of unified memory also works.
Mini, Small or the big one? How do I pick a model size?
Pick the biggest model whose 4-bit size fits in about 90% of your graphics memory; the rest is left for the context and your desktop. In the newest line-up with several sizes, Inkling 2 Mini needs about 16 GB (a 20 GB card); Inkling 2 Small needs about 368 GB (a 512 GB Mac, no single consumer card). Of all the models released in the last six months, on a 24 or 32 GB card the biggest is Ornith-1.5 35B-A3B (about 21 GB), and on a Mac with up to 64 GB it is Ornith-1.5 35B-A3B (needs a 32 GB Mac). A smaller model at 4-bit usually beats a bigger one squeezed to 2-bit.
Which new models run on a 16 GB or 24 GB graphics card?
At 4-bit with a 16K context, of the models released in the last six months: none fits a 16 GB card; 24 GB runs Inkling 2 Mini, Ornith-1.5 35B-A3B, NVIDIA Nemotron 3.5 Lightning 30B-A3B, Qwen3.8 27B. Everything else needs a bigger card, a Mac with more unified memory, MoE offload to system RAM, or a rented GPU.
Why does a mixture-of-experts model need so much memory when only a few billion parameters are active?
Every expert has to sit in memory, even though each token only uses a few of them. Inkling 2 Mini has 26.4B parameters with 3.4B active, so it needs about 16 GB at Q4_K_M. The small active count is why it runs fast once it fits, not a smaller download. With llama.cpp you can keep the experts in system RAM (MoE offload), which is slower but works.