nerkyor

/

Qwen3.6-35B-A3B-DSV4Pro-SFT-GPT56Sol-RL-Agent-GGUF

Like 28

Thanks for the suggestion. Our existing LynnStyle Q3 and Q4 are already mixed-precision quantizations, not uniform Q3/Q4 K-quants.
“Q3” and “Q4” describe the approximate size/bit-budget tier. Internally, different tensor groups use different precisions: sensitive tensors are preserved at higher precision, while more tolerant tensors—especially selected routed-expert matrices—are quantized more aggressively. The goal is to retain reasoning, coding and agent behavior within a given storage budget.
APEX follows the same broad mixed-precision principle, but supplies a standardized set of MoE-oriented profiles:
APEX I-Compact: approximately 16.1 GB
APEX I-Quality: approximately 21.3 GB
APEX I-Balanced: approximately 23.6 GB
APEX generally keeps shared experts and other sensitive dense tensors at higher precision, while applying a layer-dependent precision gradient to routed experts. Its I- profiles also use an importance matrix calibrated on diverse workloads such as chat, reasoning, coding and tool use.
Therefore, APEX would be an additional mixed-precision recipe, not a correction or replacement for our existing quants. The names are not directly comparable: for example, APEX I-Quality is not automatically better than LynnStyle Q4 simply because it is called “Quality.”
We can evaluate APEX I-Compact and I-Quality from the same F16 source, then compare them with LynnStyle using the same runtime and the same MMLU, GPQA, LiveCodeBench and Agent100 harnesses. We would publish additional tiers only if the exact-file results justify their larger size or show a useful quality/runtime trade-off.

Thanks for the suggestion. Our existing LynnStyle Q3 and Q4 are already mixed-precision quantizations, not uniform Q3/Q4 K-quants.
“Q3” and “Q4” describe the approximate size/bit-budget tier. Internally, different tensor groups use different precisions: sensitive tensors are preserved at higher precision, while more tolerant tensors—especially selected routed-expert matrices—are quantized more aggressively. The goal is to retain reasoning, coding and agent behavior within a given storage budget.
APEX follows the same broad mixed-precision principle, but supplies a standardized set of MoE-oriented profiles:
APEX I-Compact: approximately 16.1 GB
APEX I-Quality: approximately 21.3 GB
APEX I-Balanced: approximately 23.6 GB
APEX generally keeps shared experts and other sensitive dense tensors at higher precision, while applying a layer-dependent precision gradient to routed experts. Its I- profiles also use an importance matrix calibrated on diverse workloads such as chat, reasoning, coding and tool use.
Therefore, APEX would be an additional mixed-precision recipe, not a correction or replacement for our existing quants. The names are not directly comparable: for example, APEX I-Quality is not automatically better than LynnStyle Q4 simply because it is called “Quality.”
We can evaluate APEX I-Compact and I-Quality from the same F16 source, then compare them with LynnStyle using the same runtime and the same MMLU, GPQA, LiveCodeBench and Agent100 harnesses. We would publish additional tiers only if the exact-file results justify their larger size or show a useful quality/runtime trade-off.

will provide a mtp version?

Thanks for the suggestion. Our existing LynnStyle Q3 and Q4 are already mixed-precision quantizations, not uniform Q3/Q4 K-quants.
“Q3” and “Q4” describe the approximate size/bit-budget tier. Internally, different tensor groups use different precisions: sensitive tensors are preserved at higher precision, while more tolerant tensors—especially selected routed-expert matrices—are quantized more aggressively. The goal is to retain reasoning, coding and agent behavior within a given storage budget.
APEX follows the same broad mixed-precision principle, but supplies a standardized set of MoE-oriented profiles:
APEX I-Compact: approximately 16.1 GB
APEX I-Quality: approximately 21.3 GB
APEX I-Balanced: approximately 23.6 GB
APEX generally keeps shared experts and other sensitive dense tensors at higher precision, while applying a layer-dependent precision gradient to routed experts. Its I- profiles also use an importance matrix calibrated on diverse workloads such as chat, reasoning, coding and tool use.
Therefore, APEX would be an additional mixed-precision recipe, not a correction or replacement for our existing quants. The names are not directly comparable: for example, APEX I-Quality is not automatically better than LynnStyle Q4 simply because it is called “Quality.”
We can evaluate APEX I-Compact and I-Quality from the same F16 source, then compare them with LynnStyle using the same runtime and the same MMLU, GPQA, LiveCodeBench and Agent100 harnesses. We would publish additional tiers only if the exact-file results justify their larger size or show a useful quality/runtime trade-off.

will provide a mtp version?

Thanks for asking. At the moment, we do not plan to provide an MTP version for this 35B release.
In our 35B llama.cpp tests, the MTP speedup remained below 20%, while speculative decoding also had an observable impact on output quality. We therefore do not consider the current trade-off strong enough for a recommended public MTP sidecar; the bare model remains the reference inference path.
If future llama.cpp or MTP updates materially improve the acceptance rate while preserving output quality under the same evaluation suite, we will retest and may reconsider.