JasonW2025/Tesseract-R08-20261006

Tesseract R08-20261006

Research artifacts. Cannot be served using existing engines except our proprietary engine at the moment. Tesseract is a 126.3e9-parameter sparse language model served on a single NVIDIA DGX Spark (GB10, sm_121a). This repository is the checkpoint the serving lane actually ran: an fp8 dense stack with Tessera WINDOW trellis routed experts, an fp8 lm_head, and an MTP draft block for speculative decoding.

Research artifact, measured on one DGX Spark against this exact checkpoint. Not a claim about other hardware, other quantisations, or another serving stack.

Status — read this before quoting the flags

The checkpoint files in this repository are a bit-exact copy of the serving checkpoint: every one of the 75,313 files inside packed/ matches the checkpoint's own integrity manifest, and all 33 ple-table/ shards, 52 dense shards and 123 evidence/ records match by SHA-256. Two things are NOT the checkpoint: the checkpoint's own README.md is replaced by this card, and draft-head/ is an additional serving artifact (see Serving requirements).

The repository does not claim its own qualification. Its manifests say so, and they are published unmodified:

file	field	value
complete.json	serving_qualified	false
complete.json	quality_qualified	false
complete.json	performance_qualified	false
complete.json	status	complete_all_49_routed_owners
serving-bundle.json	status	complete_candidate_for_qualification
execution.json	whole_model_graph_qualified	false

In other words the artifact is structurally complete and self-declared unqualified: it is a candidate for qualification, and the numbers in this card are the measurements we hold against it, not vendor certification. Editing the manifests to flip those flags would break the hash chain that makes this repository verifiable, so they stay as produced.

Known exception, stated rather than hidden: complete.json's entry for config.json describes an older copy of that file (4382 bytes, 76d73213…) than the one shipped (4417 bytes, 2e186993…) — the config was updated after the manifest was emitted. A third-party run of the manifest check will report exactly that one mismatch and no others.

What was actually verified against this repository

check	result
packed/ archives vs complete.json	49/49 archives, 75,313/75,313 files, 0 mismatched
dense shards (52) + ple-table (33)	SHA-256 identical to the serving checkpoint
evidence/ (123 records)	SHA-256 identical
config.json / tokenizer set	identical; manifest entry stale, see above
serving identity of the checkpoint	d4cdd7b9a35e79c3 — fixed greedy decode
release gate on the checkpoint	pass on canary, KL, NLL, top-1, MMLU subset

packed/*.tar are upload-time archives: extract them into packed/ before serving (the tar members are layer-NN/<file>; extract with -C packed/).

Layout — packed/ ships as archives

packed/ is published as one tar per layer (packed/layer-NN.tar) because this checkpoint has 75,537 files and Hugging Face caps a repository at 20,000. Extract before serving:

for t in packed/layer-*.tar; do tar -xf "$t" -C packed/; done

That restores the directory-per-layer layout the loader reads. Everything else (config.json, the dense model-protected-*.safetensors, ple-table/, evidence/) is present as plain files.

Serving requirements

The measured serving configuration needs three things this repository does not contain in the checkpoint itself:

requirement	identity
engine	a Tessera-capable build that loads the WINDOW trellis wire; not published here
MTP draft head	draft-head/draft-head.json, sha256 878d057fe94a82c2d9237cdfb07c56786b643ef3c3256eaa42f910792993faed
draft-head binding	draft-head/provenance.json pins config sha256 2e1869932920dd1b6dc2931900dc74b1c09b3232b0a0ed2010dfba3cbfa37760 (the 4417-byte config shipped here) and tokenizer sha256 0997f410…

draft-head/ was added on 2026-11-09 because the original upload omitted it: the checkpoint does not carry the draft head, the serving recipe requires it by pinned sha256, and without it the published set cannot arm MTP d3. Its provenance file independently confirms that the 4417-byte config.json is the serving config — so complete.json's 4382-byte entry for config.json is the stale one (see Status).

Measured serving configuration, for reproduction of the numbers above: KV cache bf16 at 65,536 tokens, prefill budget 4,096, fp8 dense weights + fp8 MMA head, MTP depth 3, temperature 0, thinking off, prefix-cache granularity 1,024.

The wire

property	value
parameters	126.3e9 (5.5e9 dense + 120.8e9 experts)
layers / hidden / vocab	48 / 2,560 / 248,320
experts	top-10 of 512 per token
dense compute	fp8 (attention, dense FFN, lm_head MMA, MTP block)
routed experts	Tessera WINDOW trellis: B4 rate-4 / T3 rate-3 / W832 per-column Bresenham
bits per weight	3.9 experts, 8.4 dense, 4.1 overall (this repository's accounting)
KV cache	bf16, 65,536 tokens
on-disk	166.2 GiB, 75,537 files before packing

Memory on one DGX Spark (GB10, unified 121.63 GiB pool)

Serving plan, from the engine's own boot ledger:

component	GiB
Tessera compact routed experts (resident planes)	60.54
model weights (resident; n-gram table mmap'ed)	5.37
spec snapshot rows	3.45
layer objects (GR sites, GDN, QSA, PLE)	3.00
activation rows incl. the 2,048-row scoring plane	2.06
loader staging (pinned host, freed after the last layer)	1.81
KV cache pool	1.68
prefix cache	1.40
request slot state	0.86
vision tower weights	0.84
draft block + draft head	0.70
MoE scratch	0.35
vision workspace	0.29
draft-only physical head	0.17
engine, GEMM workspace, dense prefill bridge, n-gram staging	0.22
total plan	82.58 (79.23 device + 3.35 pinned) + 4.00 headroom

Peak observed while idling after boot: 91 GB used of 121 GB, 30 GB available, engine RSS 6.5 GiB. The 60.54 GiB of expert planes dominate; the rest is KV, snapshots and activations.

Measured on one DGX Spark

metric	value	note
cold prefill	1,543 tok/s	7.2k distinct fillers, cache miss confirmed
decode, one stream	48.3 tok/s	MTP d3, 2.54 tok/step, acceptance 0.513
decode, 8 streams	122 tok/s
TTFT	141–177 ms short	t ≈ 143 ms + n/1664
NLL vs bf16 teacher	4.3183 vs 4.2966	2,796 tokens, top-1 0.3530 vs 0.3484
MMLU	88.07%	n=570, 5-shot, temp 0
GSM8K	96.59%	n=1,319, 8-shot CoT
tool-call, 24-case hard	95.8%	tool-name accuracy 100%
TEB 3.0, hardmode	89/100	104 scenarios; deployability 76, responsiveness 45
TEB suites	GSM8K 98.0% (200), MMLU 91.7% (300), IFEval 83.5%/88.4%	each suite's own report
scored-prefill cost	0.061 ms/prompt token	was 3.905 before the chunking fix
MMLU-Pro (full test set)	67.95%	8,176/12,032, 14 categories, 5-shot answer-only, 11.4 h
KL(teacher ‖ candidate)	0.0553 nats/position	full vocabulary, 2,796 positions, frozen population

KL vs the bf16 teacher

Full-vocabulary KL over the frozen quality population: 0.0553 nats/position (154.55 nats over 2,796 positions, every one of the 248,320 logits per position). Both sides are gated, not assumed:

• the teacher logits reproduce the frozen teacher receipt's per-record NLL to 0.0000 nats on all 32 records, which also fixes the row alignment;
• the candidate logits reproduce the served API's logprobs to 1.1e-5 nats.

For scale, the same population gives a mean NLL of 12,074.1 (candidate) vs 12,013.3 (teacher), i.e. +0.0217 nats/token, and top-1 agreement of 0.3530 vs 0.3484. KL is larger than the NLL gap because it counts redistributed tail mass that never changes the argmax. Per record it splits cleanly: thinking-mode records 5.4-10.7 nats, non-thinking records 0.3-3.2.

MMLU-Pro, full test set

All 12,032 questions (14 categories), 5-shot answer-only, temperature 0, thinking off: 67.95% (8,176/12,032) in 11.4 hours. Per category it ranges from 84.2% (psychology) to 58.2% (physics). 70% of questions were scored from a single request via the top-20 logprob candidates; the other 30% fell back to one request per choice letter when a letter sat outside those candidates.

Thinking / reasoning settings

Controls: enable_thinking (alias thinking) inside chat_template_kwargs, and reasoning_effort ∈ {none, minimal, low, medium, high, xhigh, max}. Reasoning is returned in message.reasoning_content and counted in usage.completion_tokens_details.reasoning_tokens. History knobs: preserve_thinking, clear_thinking, drop_thinking.

Measured on this checkpoint, same prompt, temperature 0 — four observable states, not seven:

setting	thinking	reasoning tokens
reasoning_effort: none / enable_thinking: false	off	0
minimal ≡ low	on	56
medium	on	68
high ≡ xhigh ≡ max ≡ enable_thinking: true	on	39

Every accuracy number above was measured with thinking OFF (enable_thinking: false, and --no-think for TEB). The effort ladder is not monotonic in token count on this checkpoint, so an effort level is not a fine-grained control.

Serving identity

d4cdd7b9a35e79c3 — the SHA of a fixed greedy decode. Any change to the weights, the quantisation or the serving numerics changes it; it is the handshake this checkpoint is identified by.

What this repository is not

It is not the newest wire in the lab. A rate-3 (sub-4-bit, ~3.09 bpw) re-encode of the same official FP8 source exists internally (SUB4, October–November 2026) but its receipt is unfinished: per-layer gate results, serving gates (boot, C1/C4/MTP parity, NLL) and memory/fit numbers were never filled in, so it is not qualified and not published. Benchmark and compare against this repository only as the r08 wire.