ai21labs_AI21-Jamba-Reasoning-3B GGUF size and VRAM requirements

⬇ 13,995 ❤ 6
Parameters3.2B
Context262,144

bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF is a mid-size reasoning-focused model with 3.2 billion parameters, built on the jamba architecture. It has been downloaded 13,995 times.

ai21labs_AI21-Jamba-Reasoning-3B-GGUF is a Mixture-of-Experts model with 1 experts, of which 1 are active on each token. Routing only 1 of 1 experts makes it noticeably faster than a dense model of the same size — but every expert still has to be held in memory, so the numbers below are set by the full parameter count, not the active one. It is built from 28 transformer layers, a hidden size of 2,560, 20 attention heads. It uses grouped-query attention (20 query heads sharing 1 key/value heads), which already trims KV-cache memory compared with full multi-head attention.

To run bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF locally at a 4,096-token context, its quantized versions need between 1.91 GB (IQ2_M, lowest quality) and 6.82 GB (BF16, highest quality) of memory, weights plus KV cache and a system margin included. Context length drives that memory directly: at 4,096 tokens the KV cache for ai21labs_AI21-Jamba-Reasoning-3B-GGUF is about 0.05 GB, rising to roughly 3.5 GB at its full 262,144-token context. Shorter prompts free up memory for a higher-quality quantization.

For most users the best balance is BF16, needing about 6.82 GB. That means bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF fits entirely in the VRAM of a 6 GB GPU or larger, running fully on the GPU.

Available GGUF quantizations for bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF include IQ2_M, Q2_K, IQ3_XXS, Q2_K_L, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, Q3_K_XL, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q4_K_L, Q5_K_S, Q5_K_M, Q5_K_L, Q6_K, Q6_K_L, Q8_0, BF16. The model supports a native context length of up to 262,144 tokens; a longer context grows the KV cache and the memory needed.

→ Guide: How much VRAM do you need?

GGUF file size and memory by quantization

Compare real GGUF weight sizes, estimated KV cache and total memory for Q4, Q5, Q8 and every quantization published in this repository.

Quant.Bits QualityWeights KVTotal Speed~Verdict
IQ2_M 2.83 Low 1.05 GB 0.05 GB 1.91 GB 379.8 t/s Fits in VRAM
Q2_K 3.03 Low 1.13 GB 0.05 GB 1.98 GB 354.7 t/s Fits in VRAM
IQ3_XXS 3.25 Low 1.21 GB 0.05 GB 2.07 GB 330.5 t/s Fits in VRAM
Q2_K_L 3.44 Fair 1.28 GB 0.05 GB 2.14 GB 312.4 t/s Fits in VRAM
IQ3_XS 3.54 Fair 1.32 GB 0.05 GB 2.17 GB 303.9 t/s Fits in VRAM
Q3_K_S 3.64 Fair 1.36 GB 0.05 GB 2.21 GB 295.2 t/s Fits in VRAM
IQ3_M 3.67 Fair 1.36 GB 0.05 GB 2.22 GB 293.1 t/s Fits in VRAM
Q3_K_M 3.85 Fair 1.43 GB 0.05 GB 2.29 GB 279.3 t/s Fits in VRAM
Q3_K_L 4.03 Fair 1.5 GB 0.05 GB 2.35 GB 266.7 t/s Fits in VRAM
Q3_K_XL 4.4 Good 1.64 GB 0.05 GB 2.49 GB 244.5 t/s Fits in VRAM
IQ4_XS 4.4 Good 1.64 GB 0.05 GB 2.49 GB 244.0 t/s Fits in VRAM
IQ4_NL 4.64 Good 1.73 GB 0.05 GB 2.58 GB 231.6 t/s Fits in VRAM
Q4_0 4.65 Good 1.73 GB 0.05 GB 2.59 GB 231.1 t/s Fits in VRAM
Q4_K_S 4.67 Good 1.74 GB 0.05 GB 2.59 GB 230.3 t/s Fits in VRAM
Q4_K_M 4.84 Good 1.8 GB 0.05 GB 2.65 GB 222.2 t/s Fits in VRAM
Q4_1 5.11 Very good 1.9 GB 0.05 GB 2.76 GB 210.2 t/s Fits in VRAM
Q4_K_L 5.15 Very good 1.92 GB 0.05 GB 2.77 GB 208.8 t/s Fits in VRAM
Q5_K_S 5.59 Very good 2.08 GB 0.05 GB 2.94 GB 192.3 t/s Fits in VRAM
Q5_K_M 5.69 Very good 2.12 GB 0.05 GB 2.97 GB 189.0 t/s Fits in VRAM
Q5_K_L 5.95 Very good 2.21 GB 0.05 GB 3.07 GB 180.7 t/s Fits in VRAM
Q6_K 6.6 Excellent 2.46 GB 0.05 GB 3.31 GB 162.7 t/s Fits in VRAM
Q6_K_L 6.81 Excellent 2.53 GB 0.05 GB 3.39 GB 157.9 t/s Fits in VRAM
Q8_0 8.53 Excellent 3.17 GB 0.05 GB 4.03 GB 126.0 t/s Fits in VRAM
BF16 16.02 Excellent 5.96 GB 0.05 GB 6.82 GB 67.1 t/s Fits in VRAM

KV cache computed from the model's exact architecture. Speed is a rough estimate bounded by memory bandwidth.

Frequently asked questions

What kind of model is bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF?

bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF is a reasoning-focused model with 3.2 billion parameters, based on the jamba architecture. It is distributed as GGUF files for local inference.

Is bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF a Mixture-of-Experts (MoE) model?

Yes. bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF is a Mixture-of-Experts model with 1 experts, of which 1 are activated per token. That makes it faster than a dense model of the same size, but all 1 experts must be loaded into memory, so the VRAM/RAM it needs is driven by the total parameter count, not the active one.

How much VRAM do you need to run bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF?

You need about 4.03 GB of VRAM to run bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF entirely on the GPU using the Q8_0 quantization (at a 4,096-token context). Smaller quantizations lower the requirement at the cost of quality.

Can I run bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF on an 8 GB GPU?

Yes. With 8 GB of VRAM you can run bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF fully on the GPU using BF16 (about 6.82 GB).

Can I run bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF on a 16 GB GPU?

Yes. With 16 GB of VRAM you can run bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF fully on the GPU using BF16 (about 6.82 GB).

Can I run bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF on a 24 GB GPU?

Yes. With 24 GB of VRAM you can run bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF fully on the GPU using BF16 (about 6.82 GB).

What context length does bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF support?

bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF supports a native context length of up to 262,144 tokens. A longer context grows the KV cache, so it increases the memory needed to run the model.

What is the best quantization for bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF?

For bartowski/ai21labs_AI21-Jamba-Reasoning-3B-GGUF, a strong default is Q4_K_M, which needs about 2.65 GB and keeps most of the quality while roughly halving the memory versus 8-bit. With VRAM to spare, Q5_K_M or Q6_K add a little more quality; if you are tight on memory, a smaller quantization still runs. Pick the highest quantization that fits your VRAM.