eaddario/Apriel-1.6-15b-Thinker-GGUF

Experimental global target bits‑per‑weight quantization of ServiceNow-AI/Apriel-1.6-15b-Thinker

Using non-standard (forked) LLaMA C++ release b7520 for quantization.

Original model: ServiceNow-AI/Apriel-1.6-15b-Thinker

From the original model creators:

Summary

Apriel-1.6-15B-Thinker is an updated multimodal reasoning model in ServiceNow’s Apriel SLM series, building on Apriel-1.5-15B-Thinker. With significantly improved text and image reasoning capabilities, Apriel-1.6 achieves competitive performance against models up to 10x its size. Like its predecessor, it benefits from extensive continual pre-training across both text and image domains. We additionally perform post-training that focuses on Supervised Finetuning (SFT) and Reinforcement Learning (RL). Apriel-1.6 obtains frontier performance without sacrificing reasoning token efficiency. The model improves or maintains task performance when compared with Apriel-1.5-15B-Thinker, while reducing reasoning token usage by more than 30%.

Highlights

• Achieves a score of 57 on the Artificial Analysis index outperforming models like Gemini 2.5 Flash, Claude Haiku 4.5 and GPT OSS 20b. It obtains a score on par with Qwen3 235B A22B, while being significantly more efficient.
• Reduces reasoning token usage by more than 30%, delivering significantly better efficiency than Apriel-1.5-15B-Thinker.
• Scores 69 on Tau2 Bench Telecom and 69 on IFBench, which are key benchmarks for the enterprise domain.
• At 15B parameters, the model fits on a single GPU, making it highly memory-efficient.
• Based on community feedback on Apriel-1.5-15B-Thinker, we simplified the chat template by removing redundant tags and introduced four special tokens to the tokenizer (<tool_calls>, </tool_calls>, [BEGIN FINAL RESPONSE], <|end|>) for easier output parsing.

Please see our blog post for more details

⚠️ 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-imatrix to generate an imatrix including the mean activations, and a custom version of llama-quantize, which computes a per-tensor weighted mean squared quantization error and a bias/projection term (if the imatrix includes activations), 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 are two pull requests (#14891 & #15550) to merge these 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, allocating more bits to tensors estimated to matter more (e.g. llama-quantize --target-bpw 4.5678 --keep-bpw-state --imatrix imatrix.gguf baseline-model-F16.gguf 12)
5. Quantize the baseline model targeting a bpw average, treating each tensor equally instead of prioritizing some (e.g. llama-quantize --target-bpw 4.5678 --no-importance --keep-bpw-state --imatrix imatrix.gguf baseline-model-F16.gguf 12)
6. Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), HellaSwag, MMLU, Truthful QA and WinoGrande scores for each quantized model
7. Keep version with the best 𝜌PPL scores (i.e. highest Cor(ln(PPL(Q)), ln(PPL(base))))
8. Repeat until all desired quants are created

*BF16 would be preferred, but F16 performs better on Apple's GPUs

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 --keep-bpw-state option will save the above-mentioned computations to disk so that future quantizations, in the permissible bpw range 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.

   • Furthermore, the process can operate in two modes: giving priority to important tensors (default) or treating each tensor equally (setting the --no-importance option). To my knowledge, there is no computationally feasible way to determine ahead of time which modality will yield better results, and two runs per model may be needed to obtain the best quality, but the default mode usually wins.

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 quantization recipe will likely be sub-optimal.

---

Models

Bits per weight, size, perplexity and KL Divergence scores

Model	BPW	Size (GB)	μPPL	𝜌PPL	μKLD	Same Top-P
Apriel-1.6-15b-Thinker-F16	16.0006	28.9	10.328827 ±0.082711	100%	N/A	N/A
Apriel-1.6-15b-Thinker-IQ1_L	1.7500	3.2	56.217779 ±0.451822	70.15%	2.134610 ±0.004486	43.765 ±0.127
Apriel-1.6-15b-Thinker-IQ2_S	2.2500	4.1	17.718732 ±0.143206	85.86%	0.821039 ±0.002826	64.366 ±0.123
Apriel-1.6-15b-Thinker-IQ2_XS	2.1249	3.8	22.329859 ±0.168811	81.92%	1.145454 ±0.003109	59.614 ±0.126
Apriel-1.6-15b-Thinker-IQ2_XXS	2.0000	3.2	29.746083 ±0.227941	77.87%	1.478655 ±0.003549	54.132 ±0.128
Apriel-1.6-15b-Thinker-IQ3_XXS	3.0000	5.4	11.614368 ±0.090793	94.25%	0.316263 ±0.001484	77.862 ±0.107
Apriel-1.6-15b-Thinker-Q2_K	2.5000	4.5	14.935966 ±0.121114	89.08%	0.613853 ±0.002378	69.725 ±0.118
Apriel-1.6-15b-Thinker-Q3_K_L	3.7499	6.8	10.764226 ±0.085149	97.83%	0.120406 ±0.000724	86.182 ±0.089
Apriel-1.6-15b-Thinker-Q3_K_S	3.2499	5.9	11.456100 ±0.093266	96.44%	0.200764 ±0.001105	82.493 ±0.098
Apriel-1.6-15b-Thinker-Q3_K	3.5000	6.3	10.928427 ±0.088056	97.26%	0.150377 ±0.000903	84.977 ±0.092
Apriel-1.6-15b-Thinker-Q4_K_S	4.2498	7.7	10.510888 ±0.084111	98.95%	0.055504 ±0.000500	90.906 ±0.074
Apriel-1.6-15b-Thinker-Q4_K	4.5000	8.1	10.476377 ±0.084179	99.25%	0.038836 ±0.000384	92.373 ±0.068
Apriel-1.6-15b-Thinker-Q4_K_M-bartowski	4.8667	8.8	10.385404 ±0.083338	99.37%	0.032732 ±0.000359	93.084 ±0.065
Apriel-1.6-15b-Thinker-Q4_K_M-bpw	4.8666	8.8	10.447756 ±0.083977	99.45%	0.028314 ±0.000313	93.513 ±0.063
Apriel-1.6-15b-Thinker-Q5_K_S	5.2500	9.5	10.407479 ±0.083697	99.64%	0.018810 ±0.000252	94.688 ±0.058
Apriel-1.6-15b-Thinker-Q5_K	5.5000	9.9	10.411958 ±0.083741	99.74%	0.012866 ±0.000182	95.449 ±0.054
Apriel-1.6-15b-Thinker-Q6_K	6.4998	11.7	10.356608 ±0.083190	99.89%	0.004796 ±0.000159	97.435 ±0.041
Apriel-1.6-15b-Thinker-Q8_0	8.4998	15.3	10.354961 ±0.083173	99.96%	0.000659 ±0.000027	99.072 ±0.025

ARC, HellaSwag, MMLU, Truthful QA and WinoGrande scores

Scores generated using llama-perplexity with 750 tasks per test, and a context size of 768 tokens.

For the test data used in the generation of these scores, follow the appropriate links: HellaSwag, ARC, MMLU, Truthful QA and WinoGrande

Model	ARC	HellaSwag	MMLU	Truthful QA	WinoGrande	Avg Score
Apriel-1.6-15b-Thinker-IQ1_L	35.6000	42.1333	26.8000	29.4667	52.9333	37.39
Apriel-1.6-15b-Thinker-IQ2_S	52.8000	60.5333	35.6000	28.8000	56.8000	46.91
Apriel-1.6-15b-Thinker-IQ2_XS	40.9333	54.8000	30.1333	27.6000	58.4000	42.37
Apriel-1.6-15b-Thinker-IQ2_XXS	40.0000	50.5333	30.2667	25.7333	55.7333	40.45
Apriel-1.6-15b-Thinker-IQ3_XXS	58.1333	68.6666	33.7333	28.2667	67.0667	51.17
Apriel-1.6-15b-Thinker-Q2_K	49.8667	59.8666	34.5333	28.2667	62.2667	46.96
Apriel-1.6-15b-Thinker-Q3_K_L	58.5333	71.0666	38.1333	31.8667	65.3333	52.99
Apriel-1.6-15b-Thinker-Q3_K_S	60.6667	69.2000	37.0667	29.0667	66.5333	52.51
Apriel-1.6-15b-Thinker-Q3_K	62.1333	69.7333	37.6000	32.5333	65.2000	53.44
Apriel-1.6-15b-Thinker-Q4_K_S	59.3333	69.8666	37.7333	30.5333	66.9333	52.88
Apriel-1.6-15b-Thinker-Q4_K	58.2667	71.3333	37.2000	30.8000	68.4000	53.20
Apriel-1.6-15b-Thinker-Q4_K_M-bartowski	62.0000	71.3333	37.3333	31.3333	67.4667	53.89
Apriel-1.6-15b-Thinker-Q4_K_M-bpw	58.9333	71.0666	38.2667	30.8000	67.6000	53.33
Apriel-1.6-15b-Thinker-Q5_K_S	61.0667	70.5333	38.0000	30.9333	68.2667	53.76
Apriel-1.6-15b-Thinker-Q5_K	62.5333	70.5333	37.7333	30.6667	66.9333	53.68
Apriel-1.6-15b-Thinker-Q6_K	61.6000	71.6000	37.7333	31.0667	65.2000	53.44
Apriel-1.6-15b-Thinker-Q8_0	61.7333	70.4000	38.0000	30.9333	66.8000	53.57

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
Apriel-1.6-15b-Thinker-Q4_K_M-bpw	8.17 GiB	14.43 B	Metal,BLAS	12	pp512	467.49 ±3.38
Apriel-1.6-15b-Thinker-Q4_K_M-bpw	8.17 GiB	14.43 B	Metal,BLAS	12	tg128	38.30 ±1.34
Apriel-1.6-15b-Thinker-Q4_K_M-bpw	8.17 GiB	14.43 B	Metal,BLAS	12	pp1024+tg1024	61.10 ±2.58
Apriel-1.6-15b-Thinker-Q4_K_M-bartowski	8.17 GiB	14.43 B	Metal,BLAS	12	pp512	464.85 ±14.30
Apriel-1.6-15b-Thinker-Q4_K_M-bartowski	8.17 GiB	14.43 B	Metal,BLAS	12	tg128	43.63 ±1.79
Apriel-1.6-15b-Thinker-Q4_K_M-bartowski	8.17 GiB	14.43 B	Metal,BLAS	12	pp1024+tg1024	71.94 ±0.78

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.

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 (~1,200 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, 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.