New AI models · updated 24 October 2026, 18:09 UTC
New AI models,and the GPU each one needs.
Open models released in the last 60 days that people are downloading, with how much card memory each needs and the cheapest card that runs it. Then new pay-as-you-go AI services, and every price change we've seen. Read every 6 hours.
Open models
Download them, run them on a GPU.
From Hugging Face, most liked first. Chat models are counted at 4-bit, the usual way to run them, plus room for a short conversation; others as published. How we work it out
| Model | Made by | Size | Does | Memory needed | Smallest card that fits | From | Added |
|---|---|---|---|---|---|---|---|
| laya | convaiinnovations | 421M | Sorts text | about 1 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 18 Sep |
| DeepSeek-V4.1-Flash | deepseek-ai | 763B | Chat, reads images | about 459 GB (4-bit) | Needs several cards | — | 10 Sep |
| Qwen-Image-2.1 | Qwen | 7.1B | Makes images | about 14 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 14 Sep |
| ZDTaichu5.0-9B | TaichuAI | 9.8B | Chat, reads images | about 7 GB (4-bit) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 4 Sep |
| embeddinggemma-2 | 744M | Search and ranking | about 1 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 14 Sep | |
| DeepSeek-V4.1-Flash-1017 | deepseek-ai | 763B | Chat, reads images | about 459 GB (4-bit) | Needs several cards | — | 17 Oct |
| MiniCPM5-2B | openbmb | 2.5B | Chat and text | about 3 GB (4-bit) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 6 Sep |
| YuE2-3B | m-a-p | 3.6B | Speech from text | about 7 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 9 Sep |
| MiniCPM-V-4.7-35B-A3B | openbmb | 35B | Chat, reads images | about 23 GB (4-bit) | RTX 4090 (24 GB) | $0.27/hr on an RTX 3090 | 7 Oct |
| drama-3 | fishaudio | 36B | Speech from text | about 40 GB (as published) | RTX A6000 (48 GB) | $0.40/hr on an RTX A6000 | 1 Oct |
| MiMo-V2.6-Pro-RL | XiaomiMiMo | 1.0T | Chat and text | about 616 GB (4-bit) | Needs several cards | — | 21 Sep |
| flint | RecursiveSI | 40B | Chat and text | about 25 GB (4-bit) | RTX 5090 (32 GB) | $0.40/hr on an RTX A6000 | 21 Oct |
| Muse-Spark-Open-816B-A42B | meta-models | 822B | Chat, reads images | about 495 GB (4-bit) | Needs several cards | — | 15 Sep |
| flux-3-image-klein | black-forest-labs | 5.5B | Makes images | about 11 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 13 Oct |
| MOSS-Image-Base | OpenMOSS-Team | 5.1B | Makes images | about 10 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 11 Oct |
| whistle | Cactus-Compute | — | Text from speech | Not published | — | — | 30 Sep |
| ema-lightning | canberkkkkkk | — | Speech from text | Not published | — | — | 1 Oct |
| Ming-Image-0.1-Design | inclusionAI | 6.2B | Makes images | about 12 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 17 Sep |
| indra-ocr | datalab-to | — | Works with images | Not published | — | — | 10 Oct |
| MiniCPM-V-4.7-1B | openbmb | 1.6B | Chat, reads images | about 2 GB (4-bit) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 6 Oct |
| Timestral-209M | mistralai | 209M | Other | about 1 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 14 Oct |
| needle3 | Cactus-Compute | — | Chat and text | Not published | — | — | 16 Sep |
| iris-3b | speridlabs | 3.0B | Makes images | about 12 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 5 Oct |
| Hy-Memory-30B-A3B | tencent | 30B | Chat and text | about 20 GB (4-bit) | RTX 4090 (24 GB) | $0.27/hr on an RTX 3090 | 15 Oct |
| MiniMax-M3.1-Flash | MiniMaxAI | 220B | Chat, reads images | about 134 GB (4-bit) | H200 (141 GB) | $4.75/hr on an H200 | 17 Oct |
| parakeet-redux | moondream | 149M | Text from speech | about 0 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 18 Sep |
| Naive-N0.5-Flash | NaiveAI | — | Chat and text | Not published | — | — | 27 Sep |
| Sage-A1.5 | InternScience | 36B | Chat and text | about 23 GB (4-bit) | RTX 4090 (24 GB) | $0.27/hr on an RTX 3090 | 15 Oct |
| Beam-1.0-501B-A23B | reflection | 501B | Chat and text | about 302 GB (4-bit) | Needs several cards | — | 7 Oct |
| Phocinae-Largha-150M-v1 | Phocinae | 144M | Sorts text | about 0 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 8 Oct |
| FaceSwap_MiniMaxH3_REF2VA | UntMods | — | Other | Not published | — | — | 15 Sep |
| NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 | nvidia | 4.0B | Chat and text | about 4 GB (4-bit) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 16 Oct |
| Prism | FrancisRing | — | Makes video | Not published | — | — | 24 Sep |
| flux-3-image-klein-base | black-forest-labs | 5.5B | Makes images | about 11 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 13 Oct |
| Raon-1-mini-preview | KRAFTON | 29B | Chat and text | about 19 GB (4-bit) | RTX 4090 (24 GB) | $0.27/hr on an RTX 3090 | 21 Oct |
| FrogNano-4B-2609 | microsoft | 4.7B | Other | about 9 GB (as published) | RTX 4080 (16 GB) | $0.18/hr on an RTX 4060 Ti 16GB | 17 Sep |
| Ornith-1.5-122B | ornith-ai | 125B | Chat and text | about 77 GB (4-bit) | A100 80GB (80 GB) | $1.34/hr on an A100 80GB | 16 Oct |
| Ling-3.1-flash | inclusionAI | 572B | Chat and text | about 345 GB (4-bit) | Needs several cards | — | 13 Oct |
| taiga-s1 | shhivv | — | Other | Not published | — | — | 27 Sep |
| AutoTrust-JEV-27B-v1.1 | autotrust | 27B | Sorts text | about 55 GB (as published) | A100 80GB (80 GB) | $1.34/hr on an A100 80GB | 17 Oct |
"From" is the cheapest typical hourly price on Vast.ai among cards with enough memory. We leave out copies, compressed versions and add-ons of other models, and "uncensored" remakes. Fits in 32 GB? Rent a GPU on KernelMachine
AI services
New in the last 30 days.
From OpenRouter, a site that sells access to hundreds of AI models through one account. Prices are per million tokens, about 750,000 words: input is what you send, output is what the model writes.
| Service | Made by | Added | Input | Output | Reads up to |
|---|---|---|---|---|---|
| Kimi K3.1 Pro | MoonshotAI | 23 Oct | $2 | $10 | 1M tokens |
| MiniMax M3.1 Flash | MiniMax | 19 Oct | $0.15 | $0.60 | 1M tokens |
| DeepSeek V4.1 Flash 1017 | DeepSeek | 17 Oct | $0.15 | $0.60 | 1M tokens |
| Glimpse | Elorian | 17 Oct | $0.60 | $3 | 262K tokens |
| Beam 1.0 501B A23B | Reflection | 16 Oct | $0.25 | $1 | 1M tokens |
| Mercury 3 | Inception | 16 Oct | $0.10 | $0.35 | 1M tokens |
| Moondream 3 Preview | Moondream | 15 Oct | $0.30 | $2.50 | 32K tokens |
| Moondream 3.1 | Moondream | 15 Oct | $0.30 | $2.50 | 32K tokens |
| Muse Spark Open | Meta | 15 Oct | $1.25 | $4.25 | 1M tokens |
| Nemotron 3.5 Super VL | NVIDIA | 15 Oct | $0.15 | $0.50 | 1M tokens |
| Pareto 26.10 | unbiased | 15 Oct | $0.80 | $3.20 | 1M tokens |
| Pareto Latest | ~unbiased | 13 Oct | $0.80 | $3.20 | 1M tokens |
| Pareto 26.9 | unbiased | 13 Oct | $2.50 | $7.50 | 262K tokens |
| LFM2.5-24B-A2B | LiquidAI | 13 Oct | $0.050 | $0.70 | 128K tokens |
| Gemini 4 Argon | 13 Oct | $2 | $10 | 1M tokens | |
| Summarize | ScaleDown | 13 Oct | $0.040 | free | 1M tokens |
| Extract | ScaleDown | 13 Oct | $0.040 | free | 1M tokens |
| Muse Wonder | Meta | 12 Oct | $4 | $20 | 1M tokens |
| Step 5 Preview | StepFun | 8 Oct | $1 | $2.70 | 1M tokens |
| Claude Haiku 5.5 | Anthropic | 7 Oct | $0.10 | $0.50 | 1M tokens |
| Nano Banana 2.1 | 6 Oct | $1.50 | $7.50 | 66K tokens | |
| Mistral Large 4 | Mistral | 6 Oct | $1.36 | $4.18 | 1M tokens |
| Ling 3.1 Flash | inclusionAI | 2 Oct | $0.075 | $0.23 | 262K tokens |
| GPT-6.1 Sol Pro | OpenAI | 29 Sep | $2 | $10 | 1.1M tokens |
| GPT-6.1 Sol | OpenAI | 29 Sep | $2 | $10 | 1.1M tokens |
| Claude Sonnet 5.5 | Anthropic | 28 Sep | $2 | $10 | 1M tokens |
| Jev Router | TypeSafe | 25 Sep | varies | varies | 1M tokens |
| Perceptron Mk1.5 | Perceptron | 25 Sep | $0.15 | $1.50 | 37K tokens |
Price changes
Who changed their prices.
Each time a service's price moved between two of our readings, in the last 30 days. Where the output price changed, that's the one shown; it's usually the bigger cost.
| Service | Made by | Changed | Price per million tokens | Change |
|---|---|---|---|---|
| DeepSeek V4 Pro 0423 | DeepSeek | 24 Oct | $0.57 → $0.42 | −27% |
| Kimi Latest | MoonshotAI | 24 Oct | $16 → $8.94 | −44% |
| Kimi K2.7 Code | MoonshotAI | 24 Oct | $3.35 → $3.10 | −7% |
| Hy3 | Tencent | 24 Oct | $0.53 → $0.33 | −38% |
| DeepSeek V4 Flash Latest | DeepSeek | 24 Oct | $0.17 → $0.18 | +4% |
| GLM 5.3 | Z.ai | 24 Oct | $1.40 → $1.61 | +15% |
| GLM 5.3 Flash | Z.ai | 24 Oct | $0.079 → $0.50 | +535% |
| GLM Flash Latest | Z.ai | 24 Oct | $0.079 → $0.070 | −11% |
| Hy4 preview | Tencent | 24 Oct | $2.50 → $2.25 | −10% |
| MiniMax M2.5 | MiniMax | 24 Oct | $0.95 → $1.20 | +26% |
| DeepSeek V4 Pro 0423 | DeepSeek | 24 Oct | $1.02 → $0.57 | −44% |
| Kimi Latest | MoonshotAI | 24 Oct | $24 → $16 | −33% |
| GLM Latest | Z.ai | 24 Oct | $1.40 → $4.29 | +207% |
| GLM 5.3 Flash | Z.ai | 24 Oct | $0.50 → $0.079 | −84% |
| GLM Flash Latest | Z.ai | 24 Oct | $0.018 → $0.049 | +181% |
| DeepSeek V4 Pro 0423 | DeepSeek | 24 Oct | $1.32 → $1.02 | −23% |
| Kimi Latest | MoonshotAI | 24 Oct | $16 → $24 | +50% |
| DeepSeek V4 Flash Latest | DeepSeek | 24 Oct | $0.17 → $0.17 | −2% |
| GLM Flash Latest | Z.ai | 24 Oct | $0.085 → $0.079 | −8% |
| Gemma 4 26B A4B | 24 Oct | $0.30 → $0.23 | −25% | |
| DeepSeek V4 Pro 0423 | DeepSeek | 24 Oct | $1.67 → $1.32 | −21% |
| Kimi Latest | MoonshotAI | 24 Oct | $24 → $16 | −33% |
| Hy3 | Tencent | 24 Oct | $0.33 → $0.53 | +60% |
| GLM Flash Latest | Z.ai | 24 Oct | $0.20 → $0.085 | −57% |
| Hy4 preview | Tencent | 24 Oct | $2.25 → $2.50 | +11% |
| DeepSeek V4.1 Flash | DeepSeek | 24 Oct | $1.20 → $1 | −17% |
| DeepSeek Flash Latest | DeepSeek | 24 Oct | $0.50 → $1 | +100% |
| DeepSeek V4 Pro 0423 | DeepSeek | 23 Oct | $1.91 → $1.67 | −13% |
| Kimi Latest | MoonshotAI | 23 Oct | $16 → $24 | +50% |
| Hy3 | Tencent | 23 Oct | $0.53 → $0.33 | −38% |
Common questions
Questions, answered plainly.
What new AI models came out recently?
Of open models released in the last 60 days, the most liked on Hugging Face are laya (convaiinnovations), DeepSeek-V4.1-Flash (deepseek-ai), Qwen-Image-2.1 (Qwen), ZDTaichu5.0-9B (TaichuAI), embeddinggemma-2 (google). The table above has more, with the card each one needs.
What GPU do I need to run laya?
laya needs about 1 GB (as published). The smallest card that fits is an RTX 4080 (16 GB), and the cheapest card that fits costs $0.18/hr on an RTX 4060 Ti 16GB.
How do you work out how much memory a model needs?
For chat models: the number of parameters in billions times 0.6 GB at 4-bit, the usual way to run them, plus about 1.5 GB for the program and a short conversation. For other models: the size of their published files. Long conversations, big images and batches need more.
Which AI services changed their prices?
In the last 30 days, per million tokens: DeepSeek V4 Pro 0423 (DeepSeek, 24 Oct): $0.57 → $0.42; Kimi Latest (MoonshotAI, 24 Oct): $16 → $8.94; Kimi K2.7 Code (MoonshotAI, 24 Oct): $3.35 → $3.10. The table above has every change we've seen.
Preparing an invite-only beta
Get an email when access opens.
No launch date is set. We'll email you once when KernelMachine opens. Nothing else.