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.
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
hereThere 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:
- Convert the original model's safetensors to GGUF F16*
- Estimate the Perplexity score for the F16 model (baseline) using the wikitext-2-raw-v1 dataset, and save the logits
- Generate an imatrix from the most appropriate calibration dataset
- 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)
- 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)
- Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), HellaSwag, MMLU, Truthful QA and WinoGrande scores for each quantized model
- Keep version with the best 𝜌PPL scores (i.e. highest
Cor(ln(PPL(Q)), ln(PPL(base))))
- 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
-
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.
-
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.
-
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
-
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.
-
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.
-
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
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
WinoGrandeTokens per second benchmarks
Scores generated using
llama-bench. Standard (
llama-quantize with no optimization) Q4_K_M quantization included for comparison.
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.
