Olmo Hybrid 7B — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
Ai2's 7B hybrid model: 24 Gated DeltaNet (linear recurrent) layers and 8 full-attention layers, so only a quarter of its layers keep a KV cache and long prompts cost far less memory than in Olmo 3 7B. Ai2 reports about twice Olmo 3's data efficiency in pre-training. This is the Instruct-DPO build, with a 32,768-token context; Apache-2.0.
Olmo Hybrid 7B 在 Q4_K_M 下约需 5 GB 显存——量化权重加框架开销，不含 KV 缓存。在 Apple Silicon 上，这部分来自统一内存。
按量化级别的显存与速度
计算基准：NVIDIA RTX 4090 (24 GB)。包含 8K 上下文的 KV 缓存，因此数值高于上方的主数字。
| 量化 | 显存 | 显存 | 速度（估算） | 适配 |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 4.2 GB | ~163 tok/s | 可运行 | |
| Q3_K_M 3.41 bpw | 5 GB | ~141 tok/s | 可运行 | |
| Q4_K_M 4.83 bpw | 6.3 GB | ~113 tok/s | 可运行 | |
| Q5_K_M 5.67 bpw | 7.1 GB | ~101 tok/s | 可运行 | |
| Q6_K 6.56 bpw | 7.9 GB | ~91 tok/s | 可运行 | |
| Q8_0 8.50 bpw | 9.7 GB | ~75 tok/s | 可运行 | |
| F16 16.00 bpw | 16.7 GB | ~44 tok/s | 可运行 |
黑色标记 = NVIDIA RTX 4090 (24 GB) 上的可用显存。 估算来自内存带宽屋顶线模型，详见 方法说明页. Olmo Hybrid 7B 显存计算器 →
运行 Olmo Hybrid 7B
目录中能运行 Olmo Hybrid 7B 的最便宜 GPU 是 Intel Arc B570 (10 GB).
或在 Vast.ai 比较，低至 $0.35/小时 (typical low · varies)
作为亚马逊联盟成员，我们从符合条件的购买中获得收入。云 GPU 链接为推荐链接——我们可能获得佣金，您无需额外付费。
如何运行 Olmo Hybrid 7B
安装 Ollama，然后运行：
ollama run olmo-3规格
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- 参数量
- 7.4 Billion
- 上下文窗口
- 32,768
- 架构
- Hybrid: 24 Gated DeltaNet + 8 full-attention layers
- 提供商
- Ai2 (Allen Institute for AI)
- 许可证
- Apache-2.0
- Availability
- Weights downloadable
- Practical deployment
- Edge device (phone, SBC)
- 规格量化
- Q4_K_M
- 系统内存
- 16 GB
- 记录更新于
- 2026-11-11
Commercial use permitted. No usage restrictions beyond attribution.
限制与注意事项
- The base model (Olmo-Hybrid-7B) has a 65,536-token context; this instruct build is configured for 32,768.
另请参阅 Olmo 3 7B
质量与使用场景
评分由模型作者或独立评测方发布——衡量质量而非吞吐量，并非我们实测。
Olmo 3 的其他尺寸
Olmo Hybrid 7B — 常见问题
How much VRAM does Olmo Hybrid 7B need?
About 5 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Olmo Hybrid 7B run on an RTX 4090 (24 GB)?
Yes. Olmo Hybrid 7B needs about 5 GB at Q4_K_M, inside a 24 GB card, at an estimated 113 tokens/sec.
How do I run Olmo Hybrid 7B locally?
Install Ollama and run `ollama run olmo-3`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.
What other sizes does Olmo 3 come in?
Olmo 3 7B (5 GB), Olmo 3 32B (20 GB), Olmo 3.1 32B Instruct (20 GB), Olmo 3.1 32B Think (20 GB), Olmo Hybrid 7B (5 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.