Three checkpoints: K3.25, K4 and K4-Core-K6

Two producers, two calibration mixes. tpurtell's K3.25 and K4-Core-K6 share one calibration set weighted to code, maths and function calling; the calibration windows published for brandonmusic's K4 add legal text. Going from K3.25 to either 4-bit checkpoint removes most of the confident mistakes on the teacher panel. Between the two 4-bit checkpoints neither is closer to BF16 on both panels: K4 on legal and reasoning text, K4-Core-K6 on tpurtell's fidelity set.

At a glance

All three on CuteAFD v4.0.0, two RTX PRO 6000 Max-Q, thinking on

	K3.25	K4	K4-Core-K6
Producer	tpurtell	brandonmusic	tpurtell
Routed experts	3.25-bit	4-bit	4-bit
Coreattention, shared expert, output head	FP8 (Core-FP8 build; the vLLM fork serves the v1 build, same experts)	BF16 (CuteAFD loads it as FP8)	6-bit
Size on disk	144 GB (Core-FP8) · 146 GB (v1)	176 GB	166 GB
Calibration set	agentic-quant-panel-v1	windows published in the producer's K4 evidence set (see below)	agentic-quant-panel-v1
Context roomKV pool, tokens	5,799,424	1,931,264	2,240,512
Speed on CuteAFD v4	not compared here	reference	5-15% slower than K4
Top token differs from BF16per 1,000 tokens, panel P-A	68	53	53
…where BF16 was ≥90% sure	0.61	0.10	0.18
Multi-turn tool callingtool-eval-bench, points of 176, CuteAFD v4	163	158	160

What they were calibrated on

Calibration is the text a quantizer looks at while deciding how to round. A checkpoint can keep more precision on text like its calibration set. Both producers give about a third to code; the visible difference is legal text (in the windows published for K4) against maths and function calling (tpurtell).

Share of the calibration set by content

Content	tpurtell K3.25, K4-Core-K6	brandonmusic K4
Code	33%	31%
General	24%	32%
Legal	none	27%
Math	20%	none
Reasoning topics	12%	10%
Function calling	10%	none

tpurtell: agentic-quant-panel-v1, 1,426 single-turn prompts, 1.08M tokens, shares of tokens; general is multilingual Wikipedia; its reasoning category is questions on reasoning-related Wikipedia topics (the card calls the name historical); the card says a dedicated legal allocation was removed. brandonmusic: K4's model card does not describe its calibration. His K4 evidence set (GLM-5.3-Flash-ShapleyMCG-K4-Evidence, calibration/windows/panel.json) lists 640 windows of 2,048 tokens labelled fit, conditional-fit, selection and confirmation, plus the 25 sealed windows used here as panel P-A; shares above are of those 640 windows. We read them as K4's calibration windows; the card does not say so. The two reasoning categories share a name, not content.

How often each differs from BF16

Read the highlighted row: on panel P-A, confident misses fall about sixfold from K3.25 to K4 and about threefold to K4-Core-K6. P-A is drawn from K4's producer's corpus; P-B, below, also puts K3.25 furthest.

Per 1,000 tokens, CuteAFD v4.0.0 (vLLM fork in grey where it runs the checkpoint)

	K3.25	K4	K4-Core-K6
Panel P-A, 51,175 positions, full vocabulary
Top token differs	68vLLM 65	53vLLM 47	53
…where BF16 was unsureits top token under 50%	56vLLM 53	46vLLM 41	46
…where BF16 was ≥90% sure57% of all positions	0.61vLLM 0.59	0.10vLLM 0.06	0.18
Large missesKL above 1 nat	5.0vLLM 4.6	2.1vLLM 1.7	2.5
Panel P-B, generated positions, top-20 measures
Top token differs	67vLLM 62	55vLLM 52	50
Large missesKL above 1 nat	3.2vLLM 3.5	1.7vLLM 1.6	1.9

A difference where BF16 itself was unsure is usually a different but reasonable word. A difference where BF16 was confident is a likely mistake; in context these are mostly specific recall (statute numbers, names, dates). The same checkpoint on the two engines differs in 40 (K3.25) and 43 (K4) top tokens per 1,000, almost all where BF16 was unsure.

K4 and K4-Core-K6

Same 4-bit expert budget, different producer, calibration and core. On the panel drawn from brandonmusic's corpus they tie on average, with K4 closer on legal and reasoning-termination text. On tpurtell's set K4-Core-K6 is closer. K4-Core-K6 also leaves 16% more room for context and runs 5-15% slower (a later rerun puts the decode slowdown in its 6-bit core). Its code replies thought longer in one run (1.43x, read after the fact); rerun with new seeds, 1.27x and inconclusive, with no measurable part from the routed experts (Thinking length).

K4-Core-K6 against K4: where each is closer to BF16Difference in KL, K4-Core-K6 minus K4, both on CuteAFD v4.0.0. Bar left of the line = K4-Core-K6 closer, right = K4 closer; dashed = not resolved

	Difference	95% interval
Panel P-A (same corpus as the windows published for K4)
All 25 windowsfull vocabulary	+0.0016	[−0.0031, +0.0057]	not resolved
Legal text	+0.0097		resolved
Reasoning-termination text	+0.0062		resolved
General text	−0.0081		not resolved
Code and agentic text	+0.0004		not resolved
Panel P-B (tpurtell's fidelity set)
Generated positionstop-20 measure	−0.0049−14%	[−0.0076, −0.0026]	resolved

Per-domain rows: 6-7 windows each, window bootstrap within the domain. P-A shares its corpus, domains and reference model with the calibration windows published for K4 (its 25 windows are document-disjoint from them), so it can favour K4; P-B's relation to either calibration set is not known. A direction between checkpoints is claimed only when both panels agree: here they do not, so the result is panel-dependent.

K4 and K3.25

K4's distance from BF16 as a share of K3.25's (top-20 KL; lower = K4 closer)

	vLLM fork	CuteAFD
Panel P-A	0.53	0.58	K4 closer
Panel P-B	0.69	0.75	K4 closer

Full vocabulary, P-A, CuteAFD v4: K4 − K3.25 = −0.0195 [−0.0257, −0.0135]. K4 is closer on both panels and both engines, by more on P-A, the panel that can favour it.

Measures

KL: how far the checkpoint's next-token probabilities sit from BF16's, in nats; 0 = identical.
Top token differs: the checkpoint's most likely next token is not BF16's.
P-A: 25 sealed windows of 2,048 tokens with full-vocabulary BF16 teacher logits (general, legal, code/agentic, reasoning-termination). P-B: tpurtell's fidelity set, BF16 top-k rows, generated positions.
Resolved: the 95% interval (window bootstrap) excludes 0.

Numbers: SUMMARY.json; engine results on CuteAFD v4.0.0.