Qift: Fixed No-Zero Quantization for LLMs

• Qift is a fixed no-zero two‐bit quantization method for rotated LLM inference that replaces standard symmetric quantization with a mirror no-zero level set.
• It leverages Hadamard rotation and per‐channel scaling to match near-zero-centered Gaussian weight distributions, reducing reconstruction error and improving model perplexity.
• The method streamlines deployment by avoiding training, learned codebooks, and zero-points while enabling shift-friendly weight application for efficient inference.

Searching arXiv for “Qift” and related usages to ground the article in current papers. Qift is an overloaded term in the arXiv literature. In the most recent and most technically specific usage, it denotes a post-training quantization redesign for rotated $\mathrm{W2A4/KV4}$ large-language-model inference, introduced as a fixed no-zero two-bit weight level set for Hadamard-rotated pipelines (Huang et al., 1 Jun 2026). In other literatures, closely related spellings or capitalizations denote distinct concepts: QIFT as Quantum Information Field Theory, a dynamical Hamiltonian encoding framework in quantum machine learning and quantum finance (Zając et al., 24 Feb 2026), and Qift as qubit fusion / qudit fission, a conversion framework between two qubits and one four-dimensional qudit in fault-tolerant quantum computation (Moussa, 2015). Because these usages are unrelated in method and domain, precise contextual qualification is required.

1. Qift in rotated low-bit LLM inference

In "Qift: Shift-Friendly No-Zero W2 Post-Training Quantization for Rotated W2A4/KV4 LLM Inference" (Huang et al., 1 Jun 2026), Qift is a post-training quantization method for the most aggressive low-bit regime studied in that paper: rotated $\mathrm{W2A4/KV4}$, meaning 2-bit weights, 4-bit activations, and 4-bit KV cache inside a QuaRot-style Hadamard-rotated inference pipeline. The method is motivated by decode-time LLM inference being memory-bound, so reducing weight and KV-cache precision directly improves deployability and throughput (Huang et al., 1 Jun 2026).

The paper situates Qift against the failure of standard two-bit weight quantization under rotated $\mathrm{W2A4/KV4}$. In conventional symmetric integer quantization, a $b$-bit weight uses

$$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$

with per-output-channel scale $s$. For $b=2$, this yields the standard W2 level set

$\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$

The paper argues that this level set is poorly matched to rotated $\mathrm{W2A4/KV4}$, because one of only four reconstruction levels is placed exactly at zero, producing a mid-tread geometry that over-allocates representation to the exact origin while underrepresenting the dense near-zero-but-not-zero bulk of rotated weights (Huang et al., 1 Jun 2026).

Empirically, the standard level set collapses badly in pure W2A4 settings. On LLaMA-2-7B with KV4 and GPTQ, perplexity rises from 5.471 FP16 to 53.849, and on LLaMA-3.1-8B it rises to 3005.556 (Huang et al., 1 Jun 2026). A key diagnostic is that even a conventional asymmetric W2 quantizer greatly improves over this baseline, which the paper interprets as evidence that the dominant issue is reconstruction-level placement rather than bit-width alone (Huang et al., 1 Jun 2026).

Qift therefore changes only the fixed four-level scalar reconstruction geometry for W2 weights. Its main proposed level set is the mirror no-zero set

$$\mathcal{G}_{\mathrm{MNZ}}=\{-1.5,-0.5,+0.5,+1.5\},$$

equivalently

$\mathrm{W2A4/KV4}$0

The paper often summarizes this as the practical parameterizations $\mathrm{W2A4/KV4}$1 or equivalently $\mathrm{W2A4/KV4}$2 (Huang et al., 1 Jun 2026). A power-of-two variant,

$\mathrm{W2A4/KV4}$3

called PoT-MNZ, supports sign-and-shift decoded weight application and is described as “shift-friendly” (Huang et al., 1 Jun 2026).

What Qift does not change is central to its definition. It retains Hadamard rotation, per-channel scaling, A4 activation quantization, KV4 cache quantization, and optional GPTQ or GPTAQ compensation. It introduces no quantization-aware training, learned codebooks, layer-specific learned centroids, group-wise grid assignment, or asymmetric zero-points. The paper explicitly positions it as training-free, learned-codebook-free, group-grid-free, and zero-point-free (Huang et al., 1 Jun 2026).

2. Distributional rationale and quantizer geometry

The source-model argument for Qift depends on rotation. The paper studies all 224 linear modules in each of LLaMA-2-7B and LLaMA-3.1-8B and reports that pretrained weights are already nearly zero-centered before rotation, while Hadamard rotation primarily Gaussianizes the standardized shape (Huang et al., 1 Jun 2026). For each output channel $\mathrm{W2A4/KV4}$4, centeredness is measured by

$\mathrm{W2A4/KV4}$5

Shape after standardization $\mathrm{W2A4/KV4}$6 is evaluated using excess kurtosis, skewness, and Q–Q error (Huang et al., 1 Jun 2026).

Across modules, rotation leaves $\mathrm{W2A4/KV4}$7 essentially unchanged at about $\mathrm{W2A4/KV4}$8, but strongly reduces non-Gaussianity. For LLaMA-2-7B, mean absolute excess kurtosis drops from 5.164 to 0.00384, Q–Q error from 0.01693 to 0.0000401, and mean absolute skewness from 0.00368 to 0.000395. For LLaMA-3.1-8B, excess kurtosis drops from 2.057 to 0.0387, Q–Q error from 0.01429 to 0.000152, and skewness from 0.00311 to 0.000527 (Huang et al., 1 Jun 2026). This motivates modeling post-rotation weights as approximately zero-centered and Gaussian-like in standardized shape.

Given such a source, the paper argues that a four-level quantizer should not allocate one reconstruction point exactly at zero. For a zero-centered Gaussian source, the four-level Lloyd-Max solution places two inner centroids around zero and two outer centroids in the tails. Qift takes that geometry as its design prior (Huang et al., 1 Jun 2026).

To formalize the shape, the paper studies the mirror no-zero family

$\mathrm{W2A4/KV4}$9

where $\mathrm{W2A4/KV4}$0 is the inner/outer centroid ratio. It evaluates

$\mathrm{W2A4/KV4}$1

with $\mathrm{W2A4/KV4}$2 pooled real rotated weights normalized by the standard W2 per-channel scale, and $\mathrm{W2A4/KV4}$3 a single global scale optimized for that $\mathrm{W2A4/KV4}$4 (Huang et al., 1 Jun 2026). The stated result is that effective 2-bit no-zero geometries cluster in the range

$\mathrm{W2A4/KV4}$5

The scan finds the best discrete point at $\mathrm{W2A4/KV4}$6 with NMSE $\mathrm{W2A4/KV4}$7, essentially identical to the Gaussian Lloyd-Max reference ratio $\mathrm{W2A4/KV4}$8 with NMSE $\mathrm{W2A4/KV4}$9 (Huang et al., 1 Jun 2026).

This analysis organizes the paper’s comparison among candidate grids. MNZ corresponds to $b$0, PoT-MNZ to $b$1, Lloyd-Max to $b$2, and NF2 to $b$3, all within the effective band. By contrast, Far-MNZ, equivalent up to scale to $b$4, has $b$5 and performs much worse, with NMSE $b$6 (Huang et al., 1 Jun 2026). The paper uses this to argue that merely removing zero is insufficient; the inner centroids must also lie close enough to the dense central bulk.

3. Operational form and deployment properties

Operationally, Qift leaves the rotated PTQ pipeline unchanged except for the fixed W2 reconstruction grid. One first applies the fixed Hadamard rotation. Then, for each output channel, one selects a scale $b$7 by the same clipping-based MSE search used in the baseline. Each weight is assigned to the nearest scaled level from the chosen Qift set, yielding RTN-style quantization or serving as the discretization step inside GPTQ or GPTAQ (Huang et al., 1 Jun 2026).

For MNZ, reconstruction is nearest-level dequantization with

$b$8

or equivalently

$b$9

The paper states that code ordering is implementation-dependent; Qift specifies the level set rather than a unique bit-pattern permutation (Huang et al., 1 Jun 2026). It also states that, for any ordered four-level scalar quantizer, decision boundaries are the midpoints between adjacent scaled levels. For MNZ, those are at $$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$0, $$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$1, and $$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$2 in the $$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$3 parameterization. For PoT-MNZ $$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$4, the decision boundaries are at $$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$5, $$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$6, and $$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$7 (Huang et al., 1 Jun 2026).

The activation and KV-cache schemes are unchanged. Activations still use symmetric A4 with clipping ratio 0.9 in the reported setup, and keys and values remain asymmetric KV4 with clipping ratio 0.95 (Huang et al., 1 Jun 2026). Qift is described as compatible with GPTQ and GPTAQ because those methods compensate errors after discretization; Qift only changes the weight discretization geometry.

The deployment rationale is particularly explicit for PoT-MNZ. Because its decoded values are fixed signed powers of two, weight application can be implemented with sign and shift rather than general multiply by irregular centroids (Huang et al., 1 Jun 2026). Relative to standard W2, memory use is unchanged because weights remain 2-bit and per-channel scales are already standard. Relative to asymmetric W2, Qift avoids stored zero-points and zero-point subtraction. Relative to learned W2 codebooks, it avoids codebook storage, codebook lookup, group-grid metadata, and training or tuning (Huang et al., 1 Jun 2026).

A plausible implication is that Qift’s principal attraction is not only reconstruction quality but also preservation of a conventional scalar quantization interface, which the paper repeatedly associates with deployment simplicity.

4. Empirical behavior on perplexity, accuracy, and residuals

The experimental study in (Huang et al., 1 Jun 2026) covers LLaMA-2-7B and LLaMA-3.1-8B. It uses 128 WikiText-2 calibration samples of length 2048, reports WikiText-2 perplexity and zero-shot downstream accuracy on ARC-Challenge, ARC-Easy, HellaSwag, PIQA, and WinoGrande, and compares SYM-INT, conventional asymmetric W2, MNZ, PoT-MNZ, Lloyd-Max, NF2, and Far-MNZ (Huang et al., 1 Jun 2026).

In pure W2A4 with KV4 and GPTAQ on LLaMA-2-7B, SYM-INT gives perplexity 12.118, while MNZ gives 9.294, PoT-MNZ 9.577, Lloyd-Max 9.265, and NF2 9.404 (Huang et al., 1 Jun 2026). Under the harsher GPTQ setting, LLaMA-2-7B drops from 53.849 with SYM-INT to 20.297 with MNZ and 18.580 with PoT-MNZ. On LLaMA-3.1-8B, pure W2A4 with KV4 and GPTAQ improves from 29.695 with SYM-INT to 19.515 with MNZ and 20.150 with PoT-MNZ; under GPTQ, the collapse from 3005.556 is recovered to 34.396 with MNZ and 32.693 with PoT-MNZ (Huang et al., 1 Jun 2026).

Downstream accuracy follows the same pattern. On LLaMA-2-7B pure W2A4 KV4, average zero-shot accuracy improves from 0.4211 for SYM-INT GPTAQ to 0.4794 for MNZ and 0.4766 for PoT-MNZ; NF2 reaches 0.4816 and Lloyd-Max 0.4793 (Huang et al., 1 Jun 2026). On LLaMA-3.1-8B pure W2A4 GPTAQ, average accuracy rises from 0.3683 for SYM-INT to 0.4064 for MNZ and 0.4140 for PoT-MNZ (Huang et al., 1 Jun 2026).

The paper also studies mixed W2/W4 precision, where the top $$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$8 sensitive layers are upgraded from W2A4 to W4A4. Layer sensitivity is measured by

$$q = \mathrm{clip}\!\left( \left\lfloor \frac{w}{s} \right\rceil, -2^{b-1}, 2^{b-1}-1 \right), \qquad \hat{w}=s q,$$9

where $s$0 is pure W2A4 perplexity and $s$1 is the perplexity after upgrading only layer $s$2 to W4A4. The average weight precision in a 32-layer model is

$s$3

Thus $s$4 gives $s$5, an iso-bit comparison against uniform W3A4 (Huang et al., 1 Jun 2026).

At $s$6 under KV4 GPTQ, LLaMA-2-7B improves from 7.825 perplexity with SYM-INT to 7.318 with MNZ and 7.316 with PoT-MNZ; W3A4 is 6.897. For LLaMA-3.1-8B, $s$7 GPTQ improves from 12.744 with SYM-INT to 11.521 with MNZ and 11.499 with PoT-MNZ; W3A4 is 10.954 (Huang et al., 1 Jun 2026). Under GPTAQ, $s$8 on LLaMA-2-7B improves from 7.719 with SYM-INT to 7.122 with MNZ, versus W3A4 at 6.509 (Huang et al., 1 Jun 2026). The paper emphasizes that at $s$9, Qift substantially narrows the gap to W3A4 while keeping half of transformer layers at 2-bit precision.

For LLaMA-2-7B under GPTQ at $b=2$0, replacing SYM-INT with MNZ closes $b=2$1 of the $b=2$2 perplexity gap to W3A4, about 55%; under GPTAQ it closes $b=2$3 of the $b=2$4 gap, about 49% (Huang et al., 1 Jun 2026). On downstream evaluation, L16 MNZ GPTAQ reaches 0.6157 average accuracy versus 0.6179 for W3A4 GPTAQ and 0.6200 for W3A4 GPTQ (Huang et al., 1 Jun 2026).

Mechanism-level evidence is also reported. In an RTN bucket diagnostic over all rotated LLaMA-2-7B weights, standard SYM-INT underuses $b=2$5, overuses zero, and concentrates error in the $b=2$6 bucket. Its total squared reconstruction error is $b=2$7. MNZ reduces this to $b=2$8, a $b=2$9 reduction, and PoT-MNZ to $\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$0, a $\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$1 reduction (Huang et al., 1 Jun 2026). In GPTQ residual analysis for $\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$2, cumulative residual ratios normalized by SYM-INT are 0.778 for MNZ, 0.795 for PoT-MNZ, 0.794 for Lloyd-Max, and 1.020 for Far-MNZ (Huang et al., 1 Jun 2026).

5. Scope, limits, and interpretive cautions

The scope of Qift in (Huang et al., 1 Jun 2026) is explicit. It is designed for rotated PTQ and is motivated by the empirical observation that rotated weights are near-zero-centered and Gaussian-like. The paper does not claim that the same fixed level set is universally optimal without rotation, or for distributions that are strongly skewed, non-centered, or non-Gaussian (Huang et al., 1 Jun 2026).

Its strongest evidence is on LLaMA-2-7B and LLaMA-3.1-8B in QuaRot-style pipelines (Huang et al., 1 Jun 2026). The method is scalar and fixed by design, so it may be less expressive than learned nonuniform quantizers or learned codebooks in settings where additional complexity is acceptable. The paper also notes that GPTAQ interactions are model-dependent: on LLaMA-2-7B GPTAQ improves $\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$3 endpoints, while on LLaMA-3.1-8B it worsens them, although Qift remains consistently better than SYM-INT in both cases (Huang et al., 1 Jun 2026).

A common misconception addressed by the paper is that pure W2A4 failure is solely a consequence of using too few bits. The conventional asymmetric W2 baseline and the ratio analysis are used to argue that reconstruction-level geometry is a major factor (Huang et al., 1 Jun 2026). Another misconception is that “no zero” alone suffices. The negative diagnostic Far-MNZ / $\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$4, with inner/outer ratio $\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$5, is included precisely to show that a no-zero design can still perform poorly when the centroid ratio is mismatched (Huang et al., 1 Jun 2026).

This suggests a more general methodological lesson: for aggressive PTQ regimes, source-model alignment of the scalar reconstruction grid can matter as much as nominal bit-width. That statement is interpretive, but it is directly motivated by the paper’s contrast between SYM-INT, asymmetric W2, and near-Lloyd no-zero grids.

6. Other meanings of Qift and QIFT in the arXiv literature

The term is not unique to LLM quantization. In "The Inverse Born Rule Fallacy: On the Informational Limits of Phase-Locked Amplitude Encoding" (Zając et al., 24 Feb 2026), QIFT is explicitly expanded as Quantum Information Field Theory. There it is presented not as an amplitude-encoding trick but as a broader dynamical framework in which data acts as a generator of quantum dynamics through a data-dependent Hamiltonian (Zając et al., 24 Feb 2026).

That paper contrasts the map

$\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$6

or $\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$7, with a Hamiltonian-based encoding. It criticizes the former as “phase-locked” and “phase-deaf,” restricting the data manifold to the positive real orthant and thereby suppressing relative phase and non-commutativity (Zając et al., 24 Feb 2026). The proposed QIFT encoding uses an effective Hamiltonian

$\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$8

with

$\mathcal{G}_{\mathrm{sym}}=\{-2,-1,0,+1\}.$9

and a symmetric second-order Trotterized implementation

$\mathrm{W2A4/KV4}$0

(Zając et al., 24 Feb 2026). In that usage, “QIFT” belongs to quantum machine learning and quantum finance rather than model compression.

An older and again unrelated usage appears in "Quantum circuits for qubit fusion" (Moussa, 2015), where Qift refers to qubit fusion and qudit fission. The central map is

$\mathrm{W2A4/KV4}$1

with adjoint $\mathrm{W2A4/KV4}$2 performing qudit fission (Moussa, 2015). The framework exploits the identification

$\mathrm{W2A4/KV4}$3

together with a mismatch between qudit and qubit Pauli/Clifford structures, yielding the nesting

$\mathrm{W2A4/KV4}$4

(Moussa, 2015). In this context, Qift is a fault-tolerant code-conversion and non-Clifford resource mechanism, not a quantization algorithm or a Hamiltonian data-encoding theory.

These different expansions explain why the bare token “Qift” is ambiguous in bibliographic and technical contexts. The recent LLM paper (Huang et al., 1 Jun 2026) uses it as a proper method name without acronym expansion, whereas (Zając et al., 24 Feb 2026) uses uppercase QIFT as an explicit acronym, and (Moussa, 2015) uses the spelling in connection with “qubit fusion.”

7. Terminological status and contemporary significance

As of the cited arXiv record, the most current high-visibility usage of Qift is the LLM quantization method in (Huang et al., 1 Jun 2026). In that setting, the term denotes a minimal redesign of 2-bit weight reconstruction levels for rotated PTQ: instead of the standard $\mathrm{W2A4/KV4}$5, it uses fixed no-zero mirror levels—primarily $\mathrm{W2A4/KV4}$6, equivalently $\mathrm{W2A4/KV4}$7, or the shift-friendly $\mathrm{W2A4/KV4}$8—while keeping per-channel scaling and the rest of the rotated $\mathrm{W2A4/KV4}$9 pipeline unchanged (Huang et al., 1 Jun 2026).

Its significance within that paper is twofold. First, it provides empirical evidence that pure W2A4 collapse in rotated pipelines is partly a reconstruction-geometry problem rather than merely a bit-budget problem. Second, it offers a fixed, metadata-light alternative to learned W2 codebooks, group-specific grids, and zero-point-based asymmetric schemes (Huang et al., 1 Jun 2026). A plausible implication is that Qift occupies a specific design point: less expressive than learned quantizers, but simpler to deploy and substantially stronger than standard symmetric W2 in the rotated regime.

Because the same string also denotes unrelated concepts in quantum information and quantum computing (Zając et al., 24 Feb 2026, Moussa, 2015), scholarly usage benefits from immediate domain qualification, such as Qift quantization, QIFT (Quantum Information Field Theory), or Qift/qubit fusion. Without that qualification, the term does not identify a single cross-disciplinary concept.