Parsing goes omni — with a no-EU clause
Tencent’s Youtu-Parsing-Omni takes the first same-version parsing comparison in our ledger (96.96 vs 95.38) across audio and video too — under an Apache-based license whose one modification is territorial; and a discharged owed-read shows the harness beating the model.
Youtu-Parsing-Omni (5B, Oct 9) scores 96.96 on OmniDocBench v1.6 — directly comparable with WeVisDoc’s 95.38 at last — across documents, audio, and video, licensed Apache-based with zero thresholds but a “not intended for use within the European Union” clause; the owed LongCat read shows the harness (+11.01) outcontributing the model (+4.09); Sakana’s MASS closes that loop by alternating workflow search with post-training — published with the benchmarks it lost. Issue 12, coverage window 8–11 October 2026.
Week of October 8–11, 2026 · Issue 12
The Week in One Sentence
Tencent’s omni-modal parser gave the ledger its first same-version benchmark fight under a license whose single modification is a map — not for the EU — LongCat’s owed read measured the harness outranking the model it serves, Sarvam and Anthropic both cut cache prices within days of each other, and Sakana closed the loop by publishing a self-improvement cycle that trains the model on its own multi-agent workflows — decline tables included.
The Big Story: Youtu-Parsing-Omni — the third horse is omni-modal, and its license carves a map
What happened. (China.) Tencent released Youtu-Parsing-Omni (5B; created October 9, weights and technical report live) — a parsing model that doesn’t stop at page images: documents, natural images, charts and flowcharts, geometry figures, audio, and video all become structured JSON (layout, OCR/ASR, tables, formulas, bounding boxes, timestamps, captions, report fields) via ten task prompts. Code, a vLLM plugin, and the OmniDocBench evaluation code ship with it. Vendor benchmarks: OmniDocBench v1.6 overall 96.96 — claimed SOTA — with Table TEDS 96.79 and Formula CDM 96.80; on the broader OmniParsingBench, 75.08 average: best open-weight, second overall to Gemini-3-Pro1. The odd engineering corner: it runs through a custom code path (trust_remote_code=True), and no inference provider serves it yet.
What the evidence says. The number that matters is the comparison we could never make before: WeVisDoc scored 95.38 on the same OmniDocBench v1.6 (Issue 6). Same version, both vendor-run — +1.58, the first directly comparable parsing scorepair this publication has recorded, with the standing caveat that both numbers are self-reported and the independent leaderboard check joins the October 14 backlog. The scope difference matters more than the margin: WeVisDoc parses page images to Markdown; Youtu parses anything with a timeslot or a shape — its audio sub-score (78.08) and chart sub-score (95.02) on OmniParsingBench point at meeting-transcription and diagram pipelines, not just document stacks. Z.ai’s GLM-OCR (MIT) remains on v1.5 (94.62) — the three-way race stays notional until re-scored.
Under the license — the actual headline. Read from the LICENSE artifact: “License Terms of Youtu-Parsing (based on Apache-2.0 with modifications),” © Tencent 2026. The permissions are broad — use, modify, sublicense, distribute, commercial — and there are no revenue thresholds, no MAU gates, nothing financial at all. The one modification is territorial, quoted verbatim: “Youtu-Parsing IS NOT INTENDED FOR USE WITHIN THE EUROPEAN UNION.” That is a genuinely new pattern in our license ledger: since Issue 3 we have tracked per-artifact gates as money (Kimi’s $20M, GLM’s $10B security review); this is the first geography. We quote it as intent language — the clause says “not intended,” not “prohibited” — and we do not assess its legal effect; any EU-deploying reader should take the artifact to counsel. But as signal: the license map’s newest mutation is regional, not financial.
Why this matters. The enterprise-document niche keeps compounding — parsing now has three Chinese open-weight entries in six weeks (GLM-OCR, MIT; WeVisDoc, Apache-2.0; Youtu, Apache-based-plus-geography), and each license is a different experiment in openness. And the parsing race is converging on a ceiling (95–97 on the same benchmark) just as the differentiator shifts to modality coverage — audio/video parsing is where the next gap will open, and Youtu is first there with weights.
What I would test. The audio and video paths, hard: run meeting and lecture recordings through the ASR-plus-structure pipeline against your current whisper-then-parse stack; then the EU-relevant question for European readers is moot-by-license — test instead whether WeVisDoc’s 1.58-point deficit on documents is real by running both on your own degraded scans (Issue 7’s finding: the degraded track is where parsers actually fail).
Elsewhere in AI Engineering
The owed read: LongCat-DeepResearch — the harness outranked the model (China) — Worth Reading
What happened. (Explicitly labelled: this paper is dated September 28 — pre-window — read this week to discharge the ledger debt from the disrupted cycles.) Meituan’s LongCat-DeepResearch technical report, read in full: a multi-agent research harness built around an executable ResearchSpec (independent Planning Writers → Judge/Critic/Reviser → parallel Researchers with one section each and independent evidence contexts → Global and Local Editors that revise sections instead of regenerating reports). No weights — a harness paper.
Why it matters. DeepResearchBench 55.25 edges ChatGPT-DeepResearch’s 54.95 (a 0.30 margin, labeled), ResearchRubrics runs +5.62 clear, and the in-house benchmark honestly loses to ChatGPT-DR (76.04 vs 76.59) — counter-evidence carried. But the number that earns the item is the config study: previous model on the old harness 47.04 → previous model on the new harness 58.05 → current model on the new harness 62.14. The harness contributed +11.01; the model, +4.09. For three issues this desk has argued the interesting layer moved between models (selection, communication); here is a vendor’s own table saying the scaffold outranks the engine — and providing the audit pattern (config studies) that every harness claim should now require.
Sarvam repriced the model stack — and added cache-affinity routing (India) — Worth Reading
What happened. (India. Labelled update on Issues 6 and 8 coverage.) October 10, per the docs changelog: language-model pricing updated across the board, and the pricing page — re-read today — confirms the rows: sarvam-105b and its conversations variant at ₹15 input / ₹5 cached / ₹60 output per 1M (Available), gemma4 (Gemma 4 31B) at ₹35 / ₹10 / ₹90 (Beta), and glm5.3 at ₹126 / ₹20 / ₹390 (Beta) — a quiet cut against the ₹126 / ₹23.4 / ₹396 this publication published on September 20 (cached −₹3.4, output −₹6; our arithmetic over the two dated page-reads). The new row that matters most: deepseekv4.1-flash at ₹25 / ₹0.5 / ₹105 (Beta, reasoning billed as output) — DeepSeek’s September flagship, covered in this publication’s Issues 4, 6, and 10, now has an Indian rate card. The same changelog day brought session affinity: Chat Completion V2 accepts an x-session-affinity header on gemma4, letting later turns “reuse cached prompt context for lower latency,” with affinity IDs scoped to the API subscription.
Why this matters. Two moves, one thesis: cache economics. The ₹0.5/M cached price on the hosted DeepSeek is — at typical 80–85 INR/USD rates, our arithmetic, labeled as such — roughly $0.006/M, the top of DeepSeek’s own celebrated $0.003–0.006 cache-hit range (Issue 4): the axis this publication has tracked on Chinese rate cards all quarter is now priced at par on an Indian one. And session-affinity routing is the request-side complement — cache pricing only pays when turns actually hit the cache, and a subscription-scoped affinity header makes prefix reuse automatic instead of behavioral. Caveats labeled: the three headline rows are Beta tier; this rate card changed between our own artifact reads four weeks apart; and the changelog’s “now cost” implies prior prices we can verify only for glm5.3 (our own September artifact).
Anthropic’s week: a cache-read cut, a Haiku with teeth, and server-side agent programs (US) — Worth Reading
What happened. Two changelog entries, dates labelled. (October 7 — landed after last issue’s desk runs; caught up here.) Claude Haiku 5.5 shipped (claude-haiku-5-5, 1M context, 128K output, adaptive thinking plus the effort parameter — no per-MTok price stated in the entry) with a wide breaking set versus Haiku 4.5: temperature, top_p, top_k, manual budget_tokens, and assistant prefill can each return 400; the computer_20250124 tool errors (migrate to computer_toolset_20260801, or computer_20251124 on Bedrock); thinking-block replay after changing system/tools can error; structured outputs are unavailable on Bedrock; and adaptive thinking is on by default. The same entry cut Sonnet 5.5’s prompt-cache reads from $0.20 to $0.10 per million tokens — “0.05x the base input price instead of 0.1x,” writes and all other prices unchanged. October 9 (in-window): Managed Agents gained dynamic workflows in beta — an agent can write “a program that runs many agents in phases and combines their results,” executed server-side as background runs (managed-agents-2026-04-01 header, workflows: {"type": "enabled"}, workflow_run.* events).
Why it matters. The cache-read halving is this quarter’s pricing story in miniature — the prefix is the product, and Anthropic just made re-reads half price on the mid-tier (read alongside India’s item this issue, three vendors moved on cache economics in one window). The Haiku breaking set continues the 5.5 generation’s pattern: every launch ships migration work. And server-side multi-agent programs are the harness-vs-model question (this issue’s Big Story companion) turning into a hosted primitive — if the scaffold outranks the engine, the vendor selling the scaffold as infrastructure is selling the right layer.
Go deeper: changelog
Microsoft open-sourced AesCode: document generation with verifiable rewards (US) — Worth Testing
What happened. Microsoft published AesCode-32B and an 8B sibling (repo created September 29; surfaced on the org October 11 — both dates per house rule): image-text-to-text models, Apache 2.0, that “generate information-rich visual artifacts such as slides, posters, and dashboards as HTML/CSS,” using an image generated from the same prompt as an aesthetic reference. Fine-tuned from Qwen3-VL-32B-Instruct (33.36B params, ~66.7 GB BF16). The interesting engineering is the training loop: 3,000-demo cold-start SFT, then GDPO reinforcement learning over 7,408 prompts with seven reward channels — execution, text, boundary, tablechart, layout, whitespace, design — scored by rendering the candidate HTML in a sandboxed Playwright browser. Code ships at github.com/microsoft/AesCode; the paper is under review.
Technical signal. On the vendor’s 300-infographic benchmark: AesCode-32B Overall 85.89 versus GPT-5.5 at 81.28, Claude Opus 4.8 at 80.39 (both comparator generations enter our tables here for the first time, per §6), and the backbone at 61.10 — +24.8 from fine-tuning. The failure-mode numbers are the useful part: severe boundary failures on 4.3% of samples against 34.7% for GPT-5.5, and “Style is the shared ceiling for every system.” Single-vendor instrument, self-reported — labeled.
Why it matters. Renderable-output generation with execution-grade rewards is the same verification pattern this publication keeps endorsing — LongCat’s config studies, Draft-KV’s deranged-message control — applied to a boring, spendable workload (slides, posters, dashboards) at Apache-2.0. For anyone building document pipelines on the open parsing/retrieval stack covered in earlier issues, this is the generation side of the niche arriving as weights.
Go deeper: model card · code
Sakana’s MASS: recursive self-improvement, published with its own failures (Japan) — Worth Reading
What happened. (Japan. Dating: paper arXiv:2610.12176 v1 submitted October 8; the repo was created September 29 but absent from this desk’s October 7 sweep — code and paper went public together this window.) Sakana AI released MASS — Multi-Agent Self-Supervision — aimed at recursive self-improvement’s supervision bottleneck: on non-verifiable, open-ended tasks a single model instance struggles to critique its own reasoning, so MASS has the model iteratively propose, execute, and self-evaluate multi-agent workflows — evolutionary search over computational-graph-like orchestrations (role assignments, information routing) under structural guardrails — then fine-tunes on the winning workflows’ self-generated trajectories, and repeats with the updated model as executor, optimizer, and in-loop evaluator. Two cycles ran on a Qwen3.6-27B base with qwen-code 0.20.0 as the coding-agent runtime: a Tokyo lab’s self-improvement loop running on Chinese open weights.
Technical signal. The README ships results with standard deviations over three trials — and ships the declines: ScienceAgentBench 28.8 → 31.7, MLR-Bench 1.71 → 2.25 with delivered papers 57% → 76%, AstaBench E2E-Bench-Hard 0.053 → 0.067 with reports delivered 50% → 76% — but Terminal-Bench 2.0 falls 0.375 → 0.352 and SWE-bench Verified 0.626 → 0.611 across the same cycles: the loop buys open-ended research capability and quietly spends coding-agent precision. The paper claims 1.2–1.6× higher performance per output token across four open-ended benchmarks, and that a student trained on multi-agent traces beats a single-agent student trained on 1.4× more tokens. The repo carries the pipeline — workflow search, LoRA post-training with BF16/FP8 export, twelve synthetic tasks, evaluation adapters — though not the full benchmark configurations or the weights, and NOTICE.md says it verbatim: “No public open-source license has been assigned to the MASS-specific code in this repository.”
Why it matters. This issue’s companion piece writes itself: LongCat’s config study measured the harness outcontributing the model (+11.01 vs +4.09, above); MASS is the industrialization of that finding — stop choosing between harness and model, alternate them, and let each loop generation improve the other’s substrate. It is also Sakana’s orchestration line (Fugu, Issue 4) moving from routing between models to retraining on the routing’s output. Two flags for the file: gains and losses move in opposite directions by task family — deploy accordingly — and the code is unlicensed by explicit statement, so read before reuse.
PLaMo Translate went on-device — llama.cpp and MLX backends with Metal (Japan) — Worth Testing
What happened. (Japan. Dating labelled: v1.1.0 — the llama.cpp and MLX release — landed October 7 at 12:47 UTC, hours after this desk’s October 7 morning sweep recorded the org quiet; v1.1.1 shipped October 10, in-window.) PFN’s plamo-translate-cli (Apache-2.0, 353 stars) reached the shape engineers actually want: local translation via plamo-2-translate with native llama.cpp and MLX backends, Metal-accelerated on Apple Silicon, pip-installable on Python 3.10–3.14, backed by verified MLX 8-bit and BF16 model releases — the release log records each verification — with v1.1.1 fixing pip installs under Transformers 5 using tested distribution wheels. Sixteen-plus languages, Japanese and English at the center.
Why it matters. The tool is open; the engine is not, on familiar terms: plamo-2-translate downloads behind a gate under the PLaMo Community License — the ¥1B-revenue-capped, naming-requiring terms this publication read from the vendor’s shipped bundle in Issue 9, now demonstrably governing the PLaMo-2 vintage too. The honest deployment read is “Apache-2.0 CLI, community-licensed engine”: below ~US$7M revenue you are licensed; above it, the contact form. For everyone below the gate, this is the cheapest on-device Japanese↔︎English pipeline from a major Japanese lab — and the verified-quantization discipline, recorded per release, is the part other open tool releases should copy.
Go deeper: repo · plamo-2-translate
Research Radar
The window’s featured research story is the owed LongCat-DeepResearch read above (full notes in the packet). No in-window China paper met the bar; N-OPSD remains unreadable (HTML 404, third attempt — October 14 backlog with the search reset).
Repo Radar
TencentCloudADP/youtu-parsing (China)
- What’s actually in it: the omni parsing pipeline (
youtu_parsing_omni), a vLLM plugin, released OmniDocBench evaluation code, and the technical report PDF. Inspection depth: card + license + repo listing; code not executed. - What you can run: the ten task families end-to-end via vLLM — with
trust_remote_code=True(customYoutuVITAForCausalLMclass) as the security-review surface; no provider deployment yet. - What’s missing: community quantizations (none at fetch); an independent OmniDocBench leaderboard confirmation; and clarity — legal, not technical — on the EU clause’s effect.
- Why open it: the first parsing-specialist in our ledger to cover audio and video in one model — and the license is the deploy-time read.
My Read on the Week
(Synthesis and interpretation — not reported fact. Written after a cross-item pass spanning all four desks’ items.)
1. The harness graduated from measurement to method. LongCat’s config study put numbers on the scaffold outranking the engine (+11.01 vs +4.09); Anthropic shipped server-side multi-agent programs as a hosted primitive the same week; and MASS is the full industrialization — alternate workflow search with post-training, and let each loop generation improve the other’s substrate, on an open Chinese base. The verification demand upgrades with it: this publication’s rule that a harness claim without a same-model ablation is unverified now extends to self-improvement claims without decline tables — and Sakana shipped theirs unprompted, which is exactly what the rule should incentivize.
2. The license map’s mutations are now multi-axis — money, territory, and named absence. Youtu’s no-EU clause is the first geography in our ledger (after Kimi’s $20M and GLM’s $10B money-gates); PLaMo’s ¥1B cap now demonstrably reaches back to govern the PLaMo-2 engine inside an Apache-2.0 tool; EmbeddingGemma 2 declared Apache with no license file at all; MASS states in its NOTICE that no license is assigned; AesCode ships clean Apache-2.0 with weights. Five artifacts, five different openness regimes in one window. The read-the-artifact rule now explicitly covers what isn’t in the repo — the absent file is a fact.
3. Cache economics went global and local in the same week. Sarvam priced hosted DeepSeek cache at ₹0.5/M and made prefix reuse automatic with session affinity; Anthropic halved Sonnet cache reads to $0.10/M; and PLaMo Translate took the other end — the whole model resident on Apple Silicon, where the prefix is free because nothing leaves the machine. From rate cards to Metal, the same thesis: whoever controls re-read costs controls the workload. The parsing race converging on a 95–97 ceiling while modality coverage opens the next gap is the same story in miniature — the differentiators moved off the leaderboard.
What I’d Watch Next
- Dated deadlines, three of them — GPT-5.5 retires October 14; Google’s omni-flash-preview now reads as October 22 (our September read had the entry absent — verify endpoints, not memories); and MASS’s unlicensed NOTICE says watch for terms before any reuse.
- The October 14 backlog sprint — search resets October 13: Zhenwu V900’s primary, MarathonMemBench, N-OPSD, the independent OmniDocBench check, the ExtractBench cross-read, the PLaMo license browser-render, and SAS’s first open issue (seven checks, one community contact).
- MASS’s first external reproduction — the README’s own ± standard deviations make a falsification run cheap, and the Terminal-Bench/SWE declines are the rows to attack first; also whether the L¹/L² checkpoints ever ship.
One Thing I’d Experiment With This Week
Run the harness-ablation on your own agent before you buy anyone’s model claim. LongCat’s config study is the template, your stack is the subject — and MASS’s cycles are this experiment scaled into a training loop, which is the strongest argument for running it at least once by hand.
- Setup: freeze your current model. Keep your current harness as arm one. Build the cheapest possible “new harness” arm — a written plan artifact (the ResearchSpec pattern: merged, critiqued, revised), per-task independent contexts instead of a shared history, and section-level edits instead of full-output regeneration. No new model.
- Conditions: one fixed task battery you already run in production (research reports, code review sweeps, whatever your agents do); three runs per arm, same budget, same tools.
- Measure: output quality per task, tokens consumed, and the delta between arms. Decision output: your own harness-vs-model attribution number. If the harness arm moves quality more than your last model upgrade did, your next budget line is engineering, not inference — and you will have replicated the season’s most important result on your own workload before anyone else publishes theirs.
All vendor-reported figures are labeled as such; if independent evidence doesn’t exist yet, that’s stated rather than papered over. This issue’s desk ledgers live under research/2026-10-11/.
Footnotes
Gemini-3-Pro enters our comparator tables for the first time (OmniParsingBench, vendor-reported); per §6 it is not comparable to the Gemini 3.8 Flash vintage in earlier issues’ tables, and LongCat’s DRB comparators (ChatGPT-DeepResearch, Claude, Gemini — versions unstated in table headers) carry the same spread caveat.↩︎
Citation
@online{dimsum_editorial2026,
author = {dimsum editorial, the},
title = {Parsing Goes Omni — with a {no-EU} Clause},
date = {2026-10-11},
url = {https://news.llm101.in/issues/2026-10-11/},
langid = {en}
}