Archival mirror: gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 at revision 69274a0d8dff5dd35bcee8290612f71e03b6e981

This mirror exists so that a published measurement of ours stays checkable if the upstream revision ever stops resolving; it is precautionary, not a recovery — upstream still serves this revision today. Every model file here is byte-identical to that revision; nothing has been requantized, re-exported, or edited. Use the upstream repository for anything current.

The one deliberate difference in this repository's file layout: upstream's model card is preserved byte-identically as README.upstream-gittensor.md (sha256 65dc0fff9a6f829884fb1b3ef2ecd9b13696d13aca340406f92e80c6dc0207b3), and this file replaces it as the repo's README so the mirror explains itself. README.qwen-upstream.md (Qwen's original card, as shipped by upstream) and every other file keep their upstream names and bytes.

Why this exists

We measured this exact revision — teacher-forced KL divergence on shard 0 of our v5 held-out suite, plus resident-weight memory under our published protocol — and our receipts cite it by revision id. A Hugging Face revision id is not a durable citation: an upstream history squash makes it unresolvable (this happened to unsloth/Qwen3.8-27B-NVFP4@9c73e2da on 2026-08-15, one day after we measured it, which is why this policy is now unconditional). Xet storage is content-addressed, so this mirror costs a small fraction of its nominal size in actual transfer; what it preserves is the citation — a resolvable repo id, revision and digest table.

Attribution: all model content by gittensor-model-hub, a ModelOpt NVFP4 (W4A4, group 16, FP8-KV-calibrated) export of Qwen/Qwen3.8-27B. Licence inherited unchanged from upstream (Apache-2.0; LICENSE is upstream's byte-identical copy).

What this checkpoint is (read from its own files, not restated from anyone's card)

• config.json quantization_config.quant_method = "modelopt"; hf_quant_config.json declares quant_algo: "NVFP4", kv_cache_quant_algo: "FP8", group_size: 16, producer modelopt 0.0.1.dev1+gc4129b6e0.
• 2,399 tensors in the index: 400 body Linear modules quantized to NVFP4 (packed U8 weights + F8_E4M3 weight_scale + F32 weight_scale_2); 149 modules excluded from quantization (lm_head, embed_tokens, and every GDN conv1d/in_proj_a/in_proj_b).
• MTP draft intact: 15 mtp.* tensors, all BF16. Vision intact: 333 model.visual.* tensors, all BF16.
• model-00003-of-00003.safetensors (sha256 9ce944d5…) is bit-identical to the trailing shard of at least three other ModelOpt NVFP4 exports of this base model (vroomfondel, PassingByPixels, kristianpaul) — it is the untouched BF16 vision/embed/lm_head/MTP block that ModelOpt emits identically. The two quantized body shards are unique to this build; this is not a re-upload of any checkpoint we had measured before.

Digests

Verified locally against the pinned snapshot before upload, and re-verified after upload (sha256 over every file of revision 69274a0d; README.md upstream bytes live at README.upstream-gittensor.md here).

What we measured against this revision

Our measurement receipts live in the research repository (malaiwah/qwen38-27b-exl3): receipts/gittensor-nvfp4-rtx5090.json is the measurement receipt; the shard-0 KLD row is in receipts/cross-engine-comparator.json. The upstream card's own serving claims (resident GB, KV pool tokens, decode tok/s, 262k context on an RTX 5090) are upstream's claims, measured by upstream — this mirror preserves them; our receipts measure teacher-forced fidelity and resident weight memory only.