Recent models: tested capability and current limits
The fused K16/V8 implementation tested here does not provide a qualified serving benefit across recent architectures. Ministral 3 failed the frozen quality-preservation task. Qwen3.5-9B and Gemma 4 12B ran native calibration, but their cache layouts or attention features are outside the asymmetric implementation. These outcomes have different meanings.
The separate October 20 comparison tests Qwen3.8-27B and Gemma 4 26B-A4B with physical Transformers reference caches. Qwen passes both K16/V8 and symmetric FP8; neither Gemma candidate clears all prediction-loss bounds. It establishes no distinctive asymmetric advantage or fused serving support. See the completed comparison for quality, state memory and reproduction.
The precision/scaling reference adds matched fixed and dynamic policies on Qwen3-8B and Qwen2.5-7B-Instruct. It identifies model-dependent value from preserving keys and distinguishes answer preservation from aggregate accuracy. It does not extend optimized support to any architecture below.
| Model | Actual test | Result |
|---|---|---|
| Ministral-3-8B-Instruct-2512 | Physical K16/V8, both native controls, production checks and full task quality on H200 | Preservation not established on the frozen GSM8K task; no performance confirmation |
| Qwen3.5-9B | Native FlashInfer, FP16 compute, 32 calibration captures on H200 | Native capability passes; asymmetric shared-page layout unsupported |
| Gemma-4-12B-it | Native Triton attention, FP16 compute, 32 calibration captures on H200 | Native capability passes with its declared logit mask; asymmetric attention configuration unsupported |
Read the Ministral quality result for both task definitions, paired loss bounds and the parser limitation. A failed quality gate does not establish that all uses of quantization harm this model. The native-only rows have no asymmetric task-quality or performance measurement at all.
Qwen3.5-9B has eight full-attention layers and 24 recurrent layers. The native runtime shares each physical cache backing between three recurrent layers and one attention layer. Attention uses 16 query heads, four KV heads and head dimension 256. Reinterpreting that backing as one whole key region followed by one whole value region would violate its page layout. The current guard rejects this configuration. A correct layout port requires more than changing a head-dimension check.
The fraction of attention layers does not determine a total-memory saving: recurrent state and attention KV scale differently with sequence length. No 25% total-memory or one-quarter-of-layers saving is inferred.
Gemma 4 combines sliding and global attention with head dimensions 256 and 512 and soft-capped logits. The selected native FlashInfer backend rejected this configuration; native Triton ran the full calibration panel. Initial finite-logit checks encountered two intentional negative infinities at model-declared suppressed token IDs, 258883 and 258882. The corrected check requires that exact mask and finite logits everywhere else. This was not evidence of FP16 numerical instability. A brief native BF16 diagnostic did not qualify a BF16 asymmetric recipe.
The native capability runs used 16 documents at each of 1K and 8K input, with 128 forced output tokens per capture. They used eager execution, a 16 GiB cache budget, one GPU, and language-only inputs. They establish loading, execution and calibration under those settings, not production quality, graph qualification or throughput. Some recurrent parameters remain FP32 as required by the native implementation.
The archived record contains exact checkpoint and harness revisions, runtime policy, measured geometry, cache backing sizes and raw evidence hashes. Model definitions are available from Qwen, Google and Mistral. The supported Qwen3 recipe retains its separate installation and validation scope.