remove single-card section pending further evaluation

single-card section: only path D (fp8 KV, 104k window) shown; llama.cpp mentioned briefly as usable-but-not-recommended, not a parallel option

fix: stray single tildes were rendering as unintended strikethrough on HF markdown — replaced approx-marks with ≈ throughout prose

16GB correction: the 27B does not fit in ANY runtime at usable context (Q4 weights leave no KV room) — recommend a smaller model, not offload hopes

Single-card section: honest positioning vs llama.cpp — chat rig -> llama.cpp GGUF; small multi-user API serving -> this path

Single-card max-context variant: fp8 KV doubles the pool to 113.6K -> 104k window on one 32GB card, essay throughput improves; gsm8k smoke noted

Single-card section: TP1 on 1x R9700 measured (v4) — serve command, table, 4-user ceiling; explicit 16GB does-not-fit note

Honest v4 table for THIS model (arm C): equal to FP8 within noise at realistic 6k context; FP8 edge is compute-bound shapes only

Measurement update (bench v4): replay confound, honest absolutes; relative rc gains stand

card: Terminal-Bench hard lands — local B 0.341 vs cloud ref 0.273, same 3600s clock

card: MMLU-Pro reference lands (bf16 cloud 0.804 vs local B 0.817)

card: τ² airline 0.86 / retail 0.82 + MMLU-Pro 0.817 (n=1120) land — B's class-A row complete

card: τ² B lands official — 0.94 (107/114, thinking ON, full clean run; reference parity)

card: Thinking-mode recipe — chat_template_kwargs.enable_thinking, reasoning_effort silently ignored by vLLM (τ² 0.90→0.63 trap), preflight check

card: RETRACT τ² B 0.65 — harness passed an inert reasoning knob, cell measured no-think; think-ON rerun in flight

card: τ² B lands at 0.65 (from-scratch clean pass, post-campaign window)

card: drop the GSM8K seed-repeat row (noise-band note covers it)

card: retract τ² B cell pending from-scratch rerun in a clean window (B back to ⏳)

card: HLE judged (sol) — ref 0.30 / C 0.25 / B 0.275, triangle within noise

card: τ²-telecom B cell lands (0.63 first clean pass, load caveat footnote); pin quick start + links to rc10

card: community credit in the rc10 banner (concurrency here, single-stream insight theirs)

card: rc10 performance-upgrade callout (+12-28% this model, +34% FP8 arm)

card: GSM8K think newest cells (rc10)

card: C throughput refreshed on rc10

card: B throughput refreshed on in-tree tuned GEMM configs

card: AA-LCR shows both seed runs (0.77·0.81)

card: canonical benchmark labels

card: IFEval cells promoted to newest runs (C 0.95·0.91); rule text updated to most-recent-run

card: B LCR corrected to 0.81 (B-final-tagged run; earlier 0.77 was the pre-final arm)

card: B first-run GPQA-D 0.85 and AIME25 0.97 published

card: link the public benchmark harness (Capicua25x/modelbench)

card: simplified to the release format — quick start, paired performance (short/6k), first-run accuracy table, recipe

Prove the reference's HLE empties are endpoint failures, not refusals

Quality: land HLE + Terminal-Bench, use the reference's HLE re-run, flag TB as time-limited

Correct the correction: the old 6k figures were mislabelled, not unsourced

Rebuild the throughput tables from traceable logs; withdraw the untraceable 6k figures

Correct the fp8-KV throughput column: two confounds, and the block-aligned numbers

Point serving at rc7 and document that fp8 KV needs the fp8-query path

Add calibrated KV-cache scales (51 F32 scalars, weights unchanged)

tau2 retail lands at 0.767: this build is behind on 2 of 3 domains, -9 items net — replaces the 'domain-split' framing. GPQA reference repaired to 48/60 after one of three errored docs recovered (margin +7). Withdraw the 'genuine non-answers' reading of the reference's HLE empties: no finish_reason was recorded for that run

name who serves the bf16 reference (AkashML via OpenRouter, pinned) in the results-table headers as well as the definitions, and scope the empty-response observation to this measurement rather than the provider generally

rebalance the engineering section: state what this port added (~1,230 lines, two major version lines forward-ported, a W4xA8 kernel where the borrowed one was W8A8) alongside the credit, instead of burying it under the lineage

add 'What it took to run this on RDNA4': the 0.26.1 port, the FP8-WMMA kernel and its shape dispatch, the spec-decode verify gate, the RDNA4 dispatch changes, and what targeting Quark actually required — with the lineage credited and forward-ported work marked as such

add a links block (GitHub branch + kernel file, Docker Hub, base model, port notes) and correct the intro: this build is level with the reference overall, not above it on every cell

correct the bf16 source size (55.6 GB, not ~54) and show how 22.3 GB reconciles — packed U8 + retained BF16 + scales — since 4-bit 27B naively suggests ~14 GB

say what each column is: the bf16 reference is a hosted third-party endpoint (AkashML via OpenRouter), not a local run, and its empty-response failures depress it and flatter everything measured against it

declare 4-bit explicitly and explain HF's auto '8-bit' tag and '16B' size: both come from MXFP4 living in U8, and both affect every MXFP4 repo including AMD's own

explain the '16B' sidebar figure: HF sums safetensors elements and MXFP4 packs two 4-bit weights per U8 byte, so 12.05B packed bytes = 22.7B params; the model is ~27.8B logical, unchanged from upstream

state benchmark provenance: cells ran on rc5 + two bind-mounted kernel files, byte-identical (sha256-verified) to the ones in the rc6 image the card tells you to pull

GPQA: annotate that the reference left 3 of 60 docs unanswered — its answered rate is 0.8246, so this build's +8 margin is ~+5.5, not 8