Kolibri 1 BF16 — Hardware Requirements & GPU Compatibility
ChatReasoningFunctions
Kolibri 1 is Aleph Alpha's mixture-of-experts reasoning model, with 78 billion total and 3.46 billion active parameters per token, focused on German and English. It has an explicit reasoning mode and tool calling, and the card lists multi-step reasoning, retrieval-augmented generation, agentic tool calling and coding as its main uses. It is a 50-layer transformer with 384 experts per layer (one shared and six routed) and a 4:1 mix of sliding-window and grouped-query attention. This is the BF16 original; Aleph-Alpha/Kolibri-1 is the official FP8 version of the same model. The BF16 weights take about 156 GB, and the card lists 4x A100 80 GB or 2x H200 as the minimum, so it is a server-class model rather than a local one. The card gives a context length of 1,048,576 tokens, with 262,144 as its native length and the recommended limit for serving. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use. Published in October 2026, it is the BF16 release of Kolibri 1, which also ships as an FP8 repository.
1.4K downloads
56 likesSpecifications
- Publisher
- Aleph-Alpha
- Parameters
- 78.1B
- Architecture
- Kolibri1ForCausalLM
- Context Length
- 262,144 tokens
- Vocabulary Size
- 128,000
- Release Date
- 2026-10-02
- License
- Apache 2.0
Get Started
HuggingFace
Run in cloud
Fits on A100 80GB (32 GB headroom) · Q4_K_M
- Generation speed
- ~231 tok/s
- generation speed
- Cost per 1M output tokens
- $1.30
- per 1M output tokens
How Much VRAM Does Kolibri 1 BF16 Need?
Select a quantization to see compatible GPUs below.
| Quantization | Bits | VRAM | + Context | File Size | Quality |
|---|---|---|---|---|---|
| Q2_Kest. | 3.40 | 33.6 GB | 44.7 GB | 33.19 GB | 2-bit quantization with K-quant improvements |
| Q3_K_Mest. | 3.90 | 38.5 GB | 49.6 GB | 38.08 GB | 3-bit medium quantization |
| Q4_K_Mest. | 4.80 | 47.3 GB | 58.4 GB | 46.86 GB | 4-bit medium quantization — most popular sweet spot |
| Q5_K_Mest. | 5.70 | 56.0 GB | 67.1 GB | 55.65 GB | 5-bit medium quantization — good quality/size tradeoff |
| Q6_Kest. | 6.60 | 64.8 GB | 75.9 GB | 64.44 GB | 6-bit quantization, very good quality |
| Q8_0est. | 8.00 | 78.5 GB | 89.6 GB | 78.10 GB | 8-bit quantization, near-lossless |
| BF16est. | 16.00 | 156.6 GB | 167.7 GB | 156.21 GB | Brain floating point 16 — preferred for training |
est.= calculated VRAM estimate; no published GGUF file found for that quantization yet. Other rows are verified against real community uploads.
Which GPUs Can Run Kolibri 1 BF16?
Q4_K_M · 47.3 GBKolibri 1 BF16 (Q4_K_M) requires 47.3 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 62+ GB is recommended. Using the full 262K context window can add up to 11.1 GB, bringing total usage to 58.4 GB. No consumer GPU has enough memory.
Which Devices Can Run Kolibri 1 BF16?
Q4_K_M · 47.3 GB23 devices with unified memory can run Kolibri 1 BF16, including NVIDIA DGX H100, NVIDIA DGX A100 640GB, Mac Studio M4 Max (64 GB).
Runs great
— Plenty of headroom
Related Models
Frequently Asked Questions
- How much VRAM does Kolibri 1 BF16 need?
Kolibri 1 BF16 requires 47.3 GB of VRAM at Q4_K_M, or 156.6 GB at BF16. Full 262K context adds up to 11.1 GB (58.4 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 78.1B × 4.8 bits ÷ 8 = 46.9 GB
KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 11.5 GB (at full 262K context)
VRAM usage by quantization
Q4_K_M47.3 GBQ4_K_M + full context58.4 GB- Can NVIDIA GeForce RTX 5090 run Kolibri 1 BF16?
No — Kolibri 1 BF16 requires at least 33.6 GB at Q2_K, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
- What's the best quantization for Kolibri 1 BF16?
For Kolibri 1 BF16, Q4_K_M (47.3 GB) offers the best balance of quality and VRAM usage. Q5_K_M (56.0 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 33.6 GB.
VRAM requirement by quantization
Q2_K33.6 GBQ4_K_M ★47.3 GBQ5_K_M56.0 GBQ6_K64.8 GBQ8_078.5 GBBF16156.6 GB★ Recommended — best balance of quality and VRAM usage.
- Can I run Kolibri 1 BF16 on a Mac?
Yes, but only at lower quantizations. The smallest Mac that can run Kolibri 1 BF16 is Mac Mini M4 Pro (48 GB) at Q2_K; 13 of the 39 Macs we list can run it at some quantization. For Q4_K_M (47.3 GB) you need a Mac with more unified memory.
- Can I run Kolibri 1 BF16 locally?
Yes — Kolibri 1 BF16 can run locally on consumer hardware. At Q4_K_M quantization it needs 47.3 GB of VRAM. Popular tools include Ollama, LM Studio, and llama.cpp.
- How fast is Kolibri 1 BF16?
At Q4_K_M, Kolibri 1 BF16 can reach ~96 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: only the active experts are read per token, plus a fixed per-layer routing cost.
Example: NVIDIA B200 → 2.1 GB active ÷ (8000 × 0.65) = 0.40 ms, plus 50 layers × 0.055 ms = 2.75 ms, so 1000 ÷ 3.15 ms = ~317 tok/s
Estimated speed at Q4_K_M (47.3 GB)
~317 tok/s~317 tok/s~294 tok/sReal-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
- What's the download size of Kolibri 1 BF16?
At Q4_K_M, the download is about 46.86 GB. The full-precision BF16 version is 156.21 GB. The smallest option (Q2_K) is 33.19 GB.
- Which GPUs can run Kolibri 1 BF16?
No single consumer GPU has enough VRAM to run Kolibri 1 BF16 at Q4_K_M (47.3 GB). Multi-GPU or professional hardware is required.
- Which devices can run Kolibri 1 BF16?
23 devices with unified memory can run Kolibri 1 BF16 at Q4_K_M (47.3 GB), including ASUS Ascent GX10, Asus ROG Flow Z13 (2025, Ryzen AI Max+ 395, 128 GB), Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.