A six-layer DFlash-family draft model for speculative decoding with Qwen3.8-27B. This repository contains BF16 source weights; Atlas quantizes the dense draft projections to NVFP4 when loading the speed profile.

This is a draft model, not a standalone language model. A target model must verify every proposed token.

On one DGX Spark / GB10, the answering optimized-qwen target and this drafter measured the following on Weschera's fixed MinHeap prompt, single stream, temperature 0, reasoning disabled, 400 output tokens:

Full-vocabulary rates were 70.6411, 72.1877, 72.2574, 72.0858, and 72.1689 tok/s. All responses had stable output SHA-256 f51d8358ea2a5c63353ca00a29208ae2cccd3039b070043cad514cc4af9761c4. The complete result is included at benchmarks/weschera-minheap-400tok-fullvocab-nvfp4kv-5x.json.

A post-publication end-to-end repeatability gate then ran the same request ten more times against the unchanged server. All ten responses retained the exact stable output hash above. Decode throughput had median 72.321 tok/s, mean 72.151 tok/s, standard deviation 0.712 tok/s, and 0.987% coefficient of variation (range 70.799-72.929 tok/s). Thus the token output is deterministic for this probe, while throughput is repeatable within normal runtime variance, not numerically deterministic. The result is included at benchmarks/weschera-minheap-400tok-fullvocab-nvfp4kv-e2e-10x.json.

NVFP4 target KV changed the target output relative to BF16 and FP8 KV. The 72.1689 figure is therefore a deterministic throughput result for this fixed probe, not evidence that NVFP4 KV is output-equivalent or broadly quality-equivalent to BF16. Every response reached the 400-token cap, so the probe also does not certify task completion. Use BF16 KV for reference behavior and separately qualify NVFP4 KV on your workloads.

block_size=16 means trained_drafts = block_size - 1; use gamma 15. Atlas refuses widths above the checkpoint contract.

The included serve-atlas-gb10-72.sh records the measured single-stream profile without requiring changes to the Atlas Git repository:

Then run the fixed five-repeat harness from the Atlas performance branch:

The measured binary SHA-256 was ebe7a7c8408bcd2d2aa492273aba8bc608c625e3e4c6c22831d014824c44f7d5. Performance is workload-, engine-, binary-, and hardware-specific.

The exact three-file runtime export was retained without its optimizer state, training-state sidecar, corpus manifest, or immutable training manifest. This publication therefore does not claim an exact reproducible training lineage. Neighboring campaign notes are not substituted for missing artifact provenance.

What is verified is the published runtime artifact itself: file hashes, tensor count/shape contract, successful Atlas loading, deterministic target-verified decoding, acceptance telemetry, and the benchmark JSON above. The base target family is Apache-2.0 Qwen3.8-27B.

Do not infer broad quality, safety, long-context, or training-corpus claims from the MinHeap throughput result.