Bielik 11B v3.1 Instruct — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
The newest Bielik: 11B, pre-trained on 56 languages with the emphasis on Polish and European languages, then tuned with supervised fine-tuning and reinforcement learning. It adds reasoning (reasoning_effort from none to max) and native tool calling; both need SpeakLeash's bielik-tools parsers on vLLM 0.24 or newer. The chat template prepends a built-in system prompt of about 1,000 tokens, which can be turned off. The repository is gated: accept the conditions on Hugging Face first.
Bielik 11B v3.1 Instruct 在 Q4_K_M 下约需 8 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 | 6.1 GB | ~122 tok/s | 可运行 | |
| Q3_K_M 3.41 bpw | 7.2 GB | ~104 tok/s | 可运行 | |
| Q4_K_M 4.83 bpw | 9.2 GB | ~81 tok/s | 可运行 | |
| Q5_K_M 5.67 bpw | 10.4 GB | ~72 tok/s | 可运行 | |
| Q6_K 6.56 bpw | 11.6 GB | ~65 tok/s | 可运行 | |
| Q8_0 8.50 bpw | 14.3 GB | ~52 tok/s | 可运行 | |
| F16 16.00 bpw | 24.8 GB | ~5 tok/s | 需卸载 |
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. Bielik 11B v3.1 Instruct 显存计算器 →
运行 Bielik 11B v3.1 Instruct
目录中能运行 Bielik 11B v3.1 Instruct 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 Bielik 11B v3.1 Instruct
No Ollama tag. SpeakLeash publishes an official GGUF (Q4_K_M 6.7 GB) for llama.cpp, plus MLX, FP8, NVFP4 and AWQ builds.
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
- 11 Billion
- 上下文窗口
- 32,768
- 架构
- Dense transformer (50 layers, GQA)
- 提供商
- SpeakLeash / ACK Cyfronet AGH
- 许可证
- Apache-2.0
- Availability
- Weights downloadable
- Access
- Gated: request access on Hugging Face and accept the licence first
- Practical deployment
- Laptop
- 规格量化
- Q4_K_M
- 系统内存
- 16 GB
- 记录更新于
- 2026-11-11
- Runs with
- vLLM: supported source llama.cpp: supported source
Commercial use permitted. No usage restrictions beyond attribution.
限制与注意事项
- Reasoning and tool calling need the speakleash/bielik-tools vLLM parsers (vLLM 0.24 or newer); plain llama.cpp chat works without them.
- The built-in system prompt takes about 1,000 tokens of the context unless you disable it.
- SpeakLeash publishes DFlash speculative-decoding drafters for this model.
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。
Bielik 的其他尺寸
Bielik 11B v3.1 Instruct — 常见问题
How much VRAM does Bielik 11B v3.1 Instruct need?
About 8 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 Bielik 11B v3.1 Instruct run on an RTX 4090 (24 GB)?
Yes. Bielik 11B v3.1 Instruct needs about 8 GB at Q4_K_M, inside a 24 GB card, at an estimated 81 tokens/sec.
How do I run Bielik 11B v3.1 Instruct locally?
No Ollama tag. SpeakLeash publishes an official GGUF (Q4_K_M 6.7 GB) for llama.cpp, plus MLX, FP8, NVFP4 and AWQ builds. Running the published tag would send your prompts to a hosted GPU rather than your own machine.
What other sizes does Bielik come in?
Bielik PL 11B v3.0 Instruct (7 GB), Bielik 11B v3.1 Instruct (8 GB), Bielik 11B v3.0 (multilingual) (8 GB), Bielik Minitron 7B v3.0 Instruct (5 GB), Bielik PL Minitron 7B v3.0 Instruct (5 GB), Bielik 4.5B v3.0 Instruct (4 GB), Bielik 1.5B v3.0 Instruct (2 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.