eaddario/K2-Horizon-7B-GGUF

Experimental global target bits‑per‑weight quantization of IFM/K2-Horizon-7B

Using non-standard (forked) LLaMA C++ release b11859 for quantization.

Original model: IFM/K2-Horizon-7B

From the original model creators:

K2-Horizon-7B is the medium dense member of the K2-Horizon family: a 7B-core decoder-only model with a 512K context window.

K2-Horizon-7B Highlights

• Strong dense baseline. A 7B-class dense model evaluated across agentic, coding, long-context, and reasoning benchmarks.
• 512K context. Native 524,288-token context from the midtraining stages onward.
• Diffusion Adapters. For faster inference (HF).
• Intermediate checkpoints. Intermediate checkpoints are released so capability changes can be studied across training rather than at a single checkpoint.
• Fully open. Training data and recipe, training code, and evaluation resources are public.

---

⚠️ PLEASE READ THIS BEFORE USING THESE EXPERIMENTAL VERSIONS! ⚠️

An area of personal interest is finding ways to optimize the inference performance of LLMs when deployed in resource-constrained environments like commodity hardware, desktops, laptops, mobiles, edge devices, etc. There are many approaches to accomplish this, including architecture simplification and knowledge distillation, but my focus has been primarily on quantization and pruning.

The method to produce these experimental versions involves using a custom version of llama-quantize, which computes a per-tensor quantization error, to automatically select the lowest error quantization recipe that achieves a global target bits‑per‑weight (bpw). More details on the implementation and test results here

There's a pull request (#15550) to merge the changes back into the core llama.cpp project. This may or may not ever happen so, until then, the modified versions will be available on GitHub.

For testing and comparison, I use models produced by Bartowski (see credits below) and Unsloth (Daniel and Michael Han do some really interesting stuff!) but when they don't provide versions of the required model, tests and comparisons are against standard quantization obtained by simply running llama-quantize with no further optimizations.

All experimental versions were generated using an appropriate imatrix created from datasets available at eaddario/imatrix-calibration. In llama.cpp, an imatrix is a calibration file derived from running representative text through the model and collecting activation statistics. It is used to weight quantization error so that error in more “important” directions (as estimated from activations) is penalized more heavily.

The process to generate these models is roughly as follows:

1. Convert the original model's safetensors to GGUF F16
2. Estimate the Perplexity score for the F16 model (baseline) using the wikitext-2-raw-v1 dataset, and save the logits
3. Generate an imatrix from the most appropriate calibration dataset
4. Quantize the baseline model targeting a bpw average (e.g. llama-quantize --target-bpw 4.5678 --state-file --imatrix imatrix.gguf baseline-model-F16.gguf 12)
5. Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), GPQA-Diamond, HellaSwag, MMLU-Redux, Truthful QA and WinoGrande scores for each quantized model
6. Keep version with the best 𝜌PPL and μKLD scores
7. Repeat until all desired quants are created

Misconceptions about BF16 to F16 Conversion

A common concern when converting BFloat16 (BF16) models to Float16 (F16) is the potential for accuracy loss. Specifically:

• Weight Clipping (Overflow): Clipping, or overflow, is often feared but only occurs if a model's weights exceed the range of ±65,503. This is a relatively rare issue in practice.
• Subnormal Zeroing (Underflow): A more frequent occurrence is underflow, where weights smaller than approximately 5.96x10⁻⁸ are converted to zero.

Crucially, when the F16 model is subsequently used for quantization, the resulting degradation in metrics like Perplexity (PPL) or Kullback–Leibler Divergence (KLD) is minimal. Any variations are typically restricted to the hundreds or thousandths decimal places compared to the BF16 model.

However, considering that weight clipping presents a more substantial risk to model integrity, every BF16 base model undergoes validation prior to the conversion process. Consequently, no models hosted in this repository exhibit performance degradation due to overflow clipping.

While BF16 offers precision benefits, performance remains a key factor.

• Conversion Speed: Tests, such as timing convert_hf_to_gguf.py, show a notable performance difference, with conversion to BF16 being 15–30% slower than to F16.
• Inference Speed: A less pronounced but still present difference (3–6%) is observed during inference. Although native BF support has been introduced by many chip manufacturers, the slower performance may stem from the entire software and hardware stack (firmware, libraries, etc.) not being fully optimized yet.

The choice to prioritize F16 over BF16 is driven by a focus on maximizing performance in specific deployment environments. My primary objective is not large-scale quantization production, a domain where others like Bartowski and Unsloth excel at, but rather optimizing inference performance for resource-constrained environments. Since BF16 support is not yet widespread in areas like mobile, edge, and embedded devices, using F16 ensures broader compatibility and easier optimization for these use cases.

Advantages and disadvantages of the global target bits‑per‑weight quantization process

Advantages

1. Target arbitrary size models

   • When specifying --target-bpw 4.5678 for instance, the algorithm will produce a model (nearly) exactly of that size, which is very useful for maximizing VRAM usage. In a system with 24GB VRAM and a 70B model, standard quants might produce a 16.8GB file (too small, quality left on table) or a 24.1GB file (won't fit). This approach can generate a 23.85GB file to utilize the hardware fully.

2. Data-driven mixed precision often can improve quality at fixed size

   • Instead of using hardcoded heuristics (e.g. make attn_v Q5_K for a 70B model), that may be sub‑optimal for a given architecture or size, the quantization mix is determined by the actual error sensitivity of the specific model's weights. This, in practice, often yields a better quality/size trade-off, especially in aggressive quantization scenarios (1.5 to 3.5 bpw), or for unusual architectures.

   • Please note: llama.cpp’s heuristics have been tuned across many models and are highly optimized; although the target bpw method produces better quality often (>75% based on tests with 130 models from 11 different families), it can also lose in surprising cases.

3. Allows better like-for-like comparisons between models and families

   • Standard llama.cpp quantization uses hardcoded rules like: "use Q4_K_M, except bump some tensors up/down, except fall back if incompatible, except keep some tensors unquantized..." and for that reason, two different models quantized with the same Q4_K_M type can end up with very different bpw (e.g. 4.75 and 4.30).

   • All things being equal, the performance of a model is usually proportional to its overall bpw size; models with a higher bpw tend to perform better than lower bpw models. Since model A has simply been given more bits, it will typically perform better (lower perplexity, better eval scores, etc.) even if the underlying quantization method is identical. That makes comparing the performance not a controlled experiment, because the comparison is between models with different effective compression ratios.

   • --target-bpw tries to address that by making the experiment more controlled: each model gets quantized to land on (approximately) the same global byte budget, so that the models' performance differences are more attributable to architecture/training differences, quantization error behaviour at the same compression ratio, optimizer’s allocation decisions, etc.

Disadvantages

1. Quantization process is significantly slower than standard

   • This approach can take 5x-10x longer as it quantizes a sample of most tensors into 15 different formats, dequantizes them back to floats, computes error diffs, and selects the best size/error option that fits the global bpw budget.

   • However, the --state-file option will save/use the above-mentioned computations so that future quantizations, for the same model, can be generated at normal speed. It also allows to interrupt the computation process and resume it at a later time.

2. The optimization target is only a proxy for the model's performance quality

   • The process minimizes a per-tensor estimated error computed from sampled rows, not actual perplexity or divergence of output distributions (a future version may address this). Since errors interact nonlinearly across layers, there are no guarantees it will select the best possible quantization recipe subject to the bpw size constraint.

3. An imatrix with activations data is required for best results

   • Activation data is required to compute the bias factor (i.e. the systematic error projected onto activation directions). If the imatrix file does not contain activation data, the --target-bpw option will refuse to run.

---

Models

Bits per weight, size, perplexity and KL Divergence scores

Model	BPW	Size (GB)	μPPL	𝜌PPL	μKLD	Same Top-P
K2-Horizon-7B-F16	16.0005	16.77	8.948840 ±0.060298	100%	N/A	N/A
K2-Horizon-7B-Q2_K	2.5000	2.63	15.651262 ±0.115594	88.86%	0.601087 ±0.001969	64.072 ±0.128
K2-Horizon-7B-Q3_K	3.4999	3.68	9.917631 ±0.067642	97.78%	0.122294 ±0.000513	82.513 ±0.101
K2-Horizon-7B-Q4_K	4.5000	4.73	9.191470 ±0.062134	99.32%	0.035391 ±0.000207	90.659 ±0.078
K2-Horizon-7B-Q5_K	5.5000	5.77	9.007803 ±0.060800	99.72%	0.013705 ±0.000156	94.570 ±0.060
K2-Horizon-7B-Q6_K	6.5000	6.82	9.002577 ±0.060768	99.72%	0.012925 ±0.000210	95.475 ±0.055
K2-Horizon-7B-Q7_K	7.4996	7.87	8.989154 ±0.060652	99.78%	0.009608 ±0.000196	96.707 ±0.048
K2-Horizon-7B-Q8_0	8.4969	8.91	8.952100 ±0.060364	99.98%	0.000774 ±0.000029	98.669 ±0.031

ARC, GPQA-Diamond, HellaSwag, MMLU-Redux, Truthful QA, and WinoGrande scores

Scores generated using llama-perplexity with 750 tasks per test, and a context size of 1024 tokens.

For the test data used in the generation of these scores, follow the appropriate links: ARC Challenge, Truthful QA, GPQA-Diamond, HellaSwag, MMLU-Redux, WinoGrande

Model	ARC Challenge	GPQA-Diamond	HellaSwag	MMLU-Redux	Truthful QA	WinoGrande	Avg Score
K2-Horizon-7B-Q2_K	63.7333 ±1.7567	28.7879 ±3.2259	64.93	46.0000 ±1.8211	30.4000 ±1.6807	64.2667 ±1.7510	49.69
K2-Horizon-7B-Q3_K	67.4667 ±1.7119	19.1919 ±2.8058	74.27	66.6667 ±1.7225	30.8000 ±1.6869	67.3333 ±1.7137	54.29
K2-Horizon-7B-Q4_K	68.8000 ±1.6929	24.7475 ±3.0746	76.80	68.0000 ±1.7045	32.5333 ±1.7119	66.5333 ±1.7242	56.24
K2-Horizon-7B-Q5_K	68.6667 ±1.6949	24.7475 ±3.0746	76.40	69.2000 ±1.6869	32.2667 ±1.7082	66.6667 ±1.7225	56.32
K2-Horizon-7B-Q6_K	69.7333 ±1.6787	22.2222 ±2.9620	76.27	68.8000 ±1.6929	32.8000 ±1.7155	67.7333 ±1.7082	56.26
K2-Horizon-7B-Q7_K	69.4667 ±1.6828	20.2020 ±2.8606	76.27	68.8000 ±1.6929	32.8000 ±1.7155	66.5333 ±1.7242	55.68
K2-Horizon-7B-Q8_0	68.6667 ±1.6949	22.7273 ±2.9858	76.13	69.6000 ±1.6807	32.1333 ±1.7063	66.4000 ±1.7259	55.94

Tokens per second benchmarks

Scores generated using llama-bench. Standard (llama-quantize with no optimization) Q4_K_M quantization included for comparison.

model	size	params	backend	threads	test	t/s
K2-Horizon-7B-Q2_K	2.62 GiB	9.00 B	MTL,BLAS	12	pp512	912.43 ± 0.83
K2-Horizon-7B-Q2_K	2.62 GiB	9.00 B	MTL,BLAS	12	tg128	85.93 ± 0.14
K2-Horizon-7B-Q2_K	2.62 GiB	9.00 B	MTL,BLAS	12	pp1024+tg1024	133.12 ± 8.74
K2-Horizon-7B-Q3_K	3.67 GiB	9.00 B	MTL,BLAS	12	pp512	721.22 ± 4.98
K2-Horizon-7B-Q3_K	3.67 GiB	9.00 B	MTL,BLAS	12	tg128	76.83 ± 1.66
K2-Horizon-7B-Q3_K	3.67 GiB	9.00 B	MTL,BLAS	12	pp1024+tg1024	121.54 ± 0.30
K2-Horizon-7B-Q4_K	4.71 GiB	9.00 B	MTL,BLAS	12	pp512	823.58 ± 8.76
K2-Horizon-7B-Q4_K	4.71 GiB	9.00 B	MTL,BLAS	12	tg128	78.19 ± 2.53
K2-Horizon-7B-Q4_K	4.71 GiB	9.00 B	MTL,BLAS	12	pp1024+tg1024	120.43 ± 0.45
K2-Horizon-7B-Q5_K	5.76 GiB	9.00 B	MTL,BLAS	12	pp512	830.86 ± 3.91
K2-Horizon-7B-Q5_K	5.76 GiB	9.00 B	MTL,BLAS	12	tg128	58.78 ± 3.61
K2-Horizon-7B-Q5_K	5.76 GiB	9.00 B	MTL,BLAS	12	pp1024+tg1024	101.87 ± 1.96
K2-Horizon-7B-Q6_K	6.81 GiB	9.00 B	MTL,BLAS	12	pp512	864.93 ± 7.36
K2-Horizon-7B-Q6_K	6.81 GiB	9.00 B	MTL,BLAS	12	tg128	62.07 ± 0.98
K2-Horizon-7B-Q6_K	6.81 GiB	9.00 B	MTL,BLAS	12	pp1024+tg1024	106.96 ± 2.80
K2-Horizon-7B-Q7_K	7.86 GiB	9.00 B	MTL,BLAS	12	pp512	895.57 ± 1.47
K2-Horizon-7B-Q7_K	7.86 GiB	9.00 B	MTL,BLAS	12	tg128	57.06 ± 0.00
K2-Horizon-7B-Q7_K	7.86 GiB	9.00 B	MTL,BLAS	12	pp1024+tg1024	102.54 ± 0.75
K2-Horizon-7B-Q8_0	8.90 GiB	9.00 B	MTL,BLAS	12	pp512	940.40 ± 1.14
K2-Horizon-7B-Q8_0	8.90 GiB	9.00 B	MTL,BLAS	12	tg128	53.12 ± 0.03
K2-Horizon-7B-Q8_0	8.90 GiB	9.00 B	MTL,BLAS	12	pp1024+tg1024	96.03 ± 0.65

Metrics used

Perplexity: one of the key metrics used in NLP evaluation. It measures the quality of a language model by evaluating how well it predicts the next token given a particular sequence of words. A PPL of 1 indicates an exact match between predicted and actual, whereas values greater than one indicate a degree of "surprise" the generated token differs from the expected.

Kullback–Leibler (KL) Divergence: a statistical measure of how much a probability distribution differs from another. When quantizing models (or altering the original tensors in any way for that matter), the closest we can preserve the weights' probability distribution to the original model the better, thus the closest to 0 the better.

AI2 Reasoning Challenge (ARC): a benchmark to evaluate the ability of AI models to answer complex science questions that require logical reasoning beyond pattern matching.

GPQA-Diamond: a challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry.

HellaSwag: the Harder Endings, Longer contexts, and Low-shot Activities for Situations With Adversarial Generations (bit of a mouthful!) is a benchmark designed to test commonsense natural language inference. It requires the model to predict the most likely ending of a sentence.

MMLU: the Massive Multitask Language Understanding evaluates LLMs’ general knowledge and problem-solving abilities across 57 subjects, including elementary mathematics, US history, computer science, and law.

Truthful QA: evaluates how well LLMs generate truthful responses to questions. It identifies whether AI models can avoid generating false or misleading information, particularly in areas where human knowledge is prone to misconceptions.

Winogrande: based on the Winograd Schema Challenge, is a natural language understanding task requiring models to resolve ambiguities in sentences involving pronoun references.

Credits

LLaMa C++ has a large and vibrant community of contributors (~1,600 last time I checked) that actively maintain and extend its functionality, adding new models and architectures almost as fast as they appear. Considering the breakneck speed at which the AI/ML field is advancing, this alone is a remarkable feat!

While I'm grateful to all contributors, I want to recognise three in particular:

• Colin Kealty (Bartowski), for the many contributions and for being one of the best sources of high quality quantized models available on Hugging Face
• Georgi Gerganov for his amazing work with llama.cpp and the ggml/gguf libraries
• Iwan Kawrakow for being one of the key authors behind the many quantization algorithms and the imatrix functionality.