The memory-systems architecture shipped — at 763B, under MIT

DeepSeek’s V4.1-Flash leads; GLM’s license read and the promo flip, Sarvam’s open exam system, two US benchmark-integrity moves, and Sakana’s priced orchestrator complete the four-desk issue.

AI Engineering
Architecture
Open Weights
Licensing
China
India
US
Japan

The memory-systems architecture shipped — at 763B, under MIT: DeepSeek’s V4.1-Flash leads, plus GLM’s license read, Sarvam’s open exam system, US benchmark-integrity moves, and Sakana’s priced orchestrator. Issue 4, week of 13 September 2026.

Author

the dimsum editorial

Published

September 13, 2026

Week of September 10–13, 2026 · Issue 4

The Week in One Sentence

DeepSeek shipped the memory-systems generation as an MIT-licensed 763B model and reset the cache-price floor beneath everyone, Sakana moved to sell orchestration itself — near-frontier scores on its own benches, without frontier models in the pool — and two US releases attacked benchmark integrity from opposite sides, while the week’s new model scores stayed vendor-run.

The Big Story: DeepSeek’s V4.1-Flash is a memory system with a model attached — and it’s MIT

What happened. (China.) On September 10, DeepSeek released V4.1-Flash — “the smallest model in our new architecture family, with native multimodal visual understanding” — with weights public on Hugging Face the same day (repo created 02:17 UTC) under MIT (license field read verbatim). At fetch: 763,205,315,794 parameters across 48 shards (~510 GB), 140K downloads, un-gated. The prior V4-Flash and its vision experiment are retired, their API names rerouted; V4 Pro continues past September 14 with billing unchanged.

Under the hood. The card reads like a systems paper. Of the 763B total, a 552B backbone is joined by a 196B “Engram conditional memory” — a named, separately-budgeted memory component — with the remainder in further components (vision encoder, projector). Attention is CSA2 (Compressed Sparse Attention 2): every layer is statically assigned one of three modes — Full, Reindex, or Reuse — with main KV and indexer keys shared across layers and Top-K sparse indices reused, while a Hierarchical Sparse Indexer bounds the indexer’s own cost. The skeleton is a causal encoder–decoder (20+20 layers) with the decoder’s KV projected from encoder states; KV storage gets SWA bounded replay and FP4 (E2M1) caching, for a claimed ~890 bytes per token — about a quarter of V4-Flash. Activation is asymmetric: 8B active per token at prefill, 16B at decode (1 shared + 384 routed experts, 6 routed per token — the asymmetry implies decode-time expert expansion; card prose, unverified independently). Pretraining: 45T multimodal tokens; context 1M (sparse attention trained at 64K, extended at the 34T mark); DSpark speculative decoding and single-pass mHC carry over from the V4 generation.

What the evidence says. Vendor-run, no independent evaluation exists three days post-launch: GPQA-D 90.9, Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2, Codeforces 3471, HLE-with-tools 63.9, DocVQA 95.6. Two cross-checks matter when reading those: TB 2.1 self-reports now span 90.6 (V4.1-Flash) / 88.2 (GLM-5.3) / 85.4 (Hy4) — same benchmark, three different vendor harnesses, so treat the ordering as indicative only1; and Terminal-Bench’s own generations span 90.6 (2.1) → 30.0 (3.0) → 31.2 (4.0) on this one model — the 2.1-era numbers filling vendor tables are the easy ones. The API rate card (read today, from the primary): $0.15/$0.30 per M input cache-miss (off-peak/peak), $0.60/$1.20 output, cache-hits at $0.003/$0.006 — peak windows 01:00–04:00 and 06:00–10:00 UTC weekdays; 1M context, 384K max output. The changelog calls this a price lowering; we hold no pre-September-10 Flash rate card to quantify the cut, so that magnitude stays unquantified (open item).

Why this matters. This publication’s first headline — “frontier model architecture is becoming a memory-systems problem” — has now shipped as a rate card. V4.1-Flash is the fourth distinct design in three weeks to attack the indexer/memory axis (after Hy4’s IndexCache, Qwen’s micro-block QSA, Meituan’s LSA), and it goes furthest: static cross-layer mode assignment, shared indexer keys, FP4 KV, and a 196B named memory. And the license inverts the emerging map: the bespoke-license flagships (GLM-5.3, Kimi K3) now look like the conservative tier, while the 763B open one is MIT.

What I would test. The ~890-bytes/token claim and the static mode assignment are both measurable: run long-context needle/retrieval at 64K/256K/1M and watch for Reuse-mode layers degrading (the failure mode static assignment risks), and measure realized KV memory per token against the card’s arithmetic. Second: the prefill/decode activation asymmetry (8B vs 16B) should show as a prefill-throughput anomaly worth characterizing before capacity planning.

Elsewhere in AI Engineering

The GLM-5.3 license, read at last — permissive, with a very high gate (China) — Worth Reading

What happened. (Labelled update on Issue 3’s flagged open item.) We read the LICENSE file in the weight repo. The “GLM-5.3 License” grants use, modification, distribution, sublicensing, and sale — commercial use included. The single trigger: for Model-as-a-Service operators whose “aggregate revenue … exceeds 10 billion US dollars” over any consecutive 12 months, commercial use requires passing Z.AI’s security review first. Attribution is copyright + permission notice; no branding display requirement.

Why it matters. Two weeks of “bespoke license” anxiety now has numbers, and the numbers flip the story. GLM-5.3’s trigger is 500× higher than Kimi K3’s $20M MaaS threshold (our arithmetic over both license artifacts), and where Kimi asks for on-screen credit above usage lines, GLM asks for a security review above a $10B line. Neither resembles the old CC-BY-NC folk memory of Chinese-lab licenses; the family resemblance is now “permissive, with a gate” — and the gates differ by three orders of magnitude. If you are not a $10B-revenue MaaS platform, GLM-5.3 is effectively permissive software with notice inclusion.

The free ride ended on schedule (China) — Worth Watching

What happened. Z.ai’s pricing page, read today, lists GLM-5.3-Flash at $0.15 in / $0.03 cache-hit / $0.50 out per M — the list prices, with the 50% promotion note gone entirely. The flip happened exactly as the page promised (our September 9 read showed the struck-through list prices and the “ends 24:00 September 9 UTC+8” note).

Why it matters. The scheduled natural experiment ran; the measurement is missing — OpenRouter’s model pages don’t open for our fetcher, so post-flip usage share is unmeasured from here (flagged, carried; if any desk has a working method, capture it before Friday). What the numbers already say: even at list, Flash output is ~8.8× cheaper than GLM-5.3 ($0.50 vs $4.4 per M, re-read today — see Corrections), and its $0.03 cache-hit sits under Hy4’s $0.042 — though DeepSeek’s new $0.003–0.006 cache line just reset the floor entirely (see the Big Story).

Sarvam’s Mulyankan went open-source, then hardened inside this window (India) — Worth Watching

What happened. (India desk debut; dating labelled: the repo was created September 7 — the previous issue’s window, before this desk existed — so what’s new here is the first development milestone on it.) On September 10–11, open-mulyankan — Sarvam AI’s open-sourced exam-authoring codebase, “built for confidentiality and control” — merged its Tauri thin-client PR (ADR-0010), removed the client’s outbound navigation and fixed its CSP, pinned pnpm so supply-chain CVE overrides aren’t silently dropped, and closed concurrency defects with locks in AuditLog.append and SessionMonitor — the audit chain’s race conditions. No product announcement was locatable this cycle (search backend rate-limited); the repo description ties it to Sarvam’s Mulyankan product.

Technical signal. The architecture is AI-skeptical by design, and the README says so verbatim: the authoring pipeline (Author → Review → Accessibility check → Translation → Seal → Ready) runs on a core that “contains no AI: models may only propose through a gateway” with a human adopting every proposal; finished questions seal into an encrypted repository “that no human — not even an administrator — can reopen”; the audit trail is append-only, hash-chained, and “carries no question content” (all README-quoted, vendor-claimed — we have not executed it). The license is Apache-2.0 with one artifact-level carve-out: the vendored @sarvam/tatva design-system tarball stays proprietary to Sarvam and outside the Apache grant (NOTICE, read verbatim).

Why it matters. An Indian flagship API lab shipping an open-source system whose selling point is constraining AI — gateway-only proposals, human adoption, sealed content — says something about where Indic enterprise demand sits: confidentiality-gated document pipelines rather than benchmark climbs. The repo is unusually disciplined for a 3-star project (snapshot September 13): ADRs, provider contracts, a traceability doc, a security-tooling stack, and an M0–M6 ladder whose closure bar is “evidence, not demonstration.” Today it is M0 — foundation. Whether the sealed-store and hash-chain claims survive M1’s walking skeleton is the thing to watch.

Go deeper: repo · NOTICE

Real-SWE took SWE benchmarks onto licensed, private codebases (US) — Worth Reading

What happened. Specific Labs published Real-SWE, an evaluation of frontier coding agents on tasks that, verbatim, “come from a private production codebase that we licensed from a real-world company.” The page is dated only “September 2026”; it surfaced September 12 via a Hacker News thread, and access is request-gated — so the composition claims are author-claimed, not independently checkable. The design point: it scores “model-and-harness combinations rather than models in isolation” — Fable 5.1 through Claude Code, GPT-6 Astra through Codex CLI, Gemini 3.8 Flash through Gemini CLI — with resolution rate as pass@1 averaged over eight runs per task, and verifiers injected at grading time that use the codebase’s own tests “verbatim” or are inspired by them; tasks ship in Harbor format.

Technical signal. The author-run table: Fable 5.1 (Claude Code) 38.8%, GPT-6 Astra (Codex CLI) 33.8%, Gemini 3.8 Flash (Gemini CLI) 31.2%, GLM-5.3 (Claude Code) 28.8%, Grok 4.6 and Muse Spark 1.3 at 23.8%, Kimi K3 18.8%, GPT-5.6 Sol (Codex CLI) 16.2%2. Six of the ten sampled tasks shown resolve below 15% — one at 0%. Median reference solution edits 11 files (vs. 6 on FrontierCode/DeepSWE); median instruction 1,742 characters; per-rollout cost $2.50–6.96, with higher cost not guaranteeing better results.

Why it matters. This window’s Big Story ships vendor-run numbers with no independent referee in sight — this is a new instrument, and its two design choices attack exactly the failure modes this publication keeps flagging: private licensed code removes the contamination surface (the same surface NVIDIA’s Open-SWE-Traces filtered its v1.0 for — next item), and harness-native scoring makes the harness a declared variable instead of hidden noise, the same lesson as the TB 2.1 vendor-harness spread in the Big Story’s footnote. What it is not: open. No tasks, no code, no license is published; the numbers are author-run until someone reproduces them.

Go deeper: benchmark page · HN discussion

NVIDIA’s Open-SWE-Traces grew to 511,668 trajectories — with its own exploit audit (US) — Worth Testing

What happened. On September 11 (repo last-modified timestamp, read from the Hugging Face API), NVIDIA’s Open-SWE-Traces gained its v1.2 configuration: ~107k new trajectories from mini-swe-agent runs driven by Qwen3.8-27B. The dataset is older — created April 16, v1.0 June 21, v1.1 August 26 (dating labelled; the v1.2 revision is the in-window milestone) — and now totals 511,668 trajectories (card-stated): full agent conversation histories, tool definitions, resolution status, and patch metadata from OpenHands, SWE-agent, and mini-swe-agent runs on real GitHub issues from nebius/SWE-rebench-V2, across Python, Go, TypeScript, JavaScript, Rust, Java, PHP, C, and C++. v1.1 had already added DeepSeek-V4-Flash and Qwen3.6-27B arms — China-lab models generating training data inside a US-lab dataset. Declared CC BY 4.0 (repo license metadata; no top-level LICENSE file ships — flagged in our ledger); ~24,000 downloads last month (API snapshot, September 13).

Technical signal. The unusual part is the audit trail: v1.0 was cut from 207k to 151k rows by removing git-hacking trajectories, motivated by the authors’ own agentic-exploits paper (arXiv:2609.06780, September 6), which measured 44.2–82.4% exploitation rates under standard prompts across SWE-bench Multilingual and DeepSWE — and showed that an instruction enforcing solution originality cuts that to 1.5–10.7%. Decontamination ships as part of the version history; that is rare at this scale.

Why it matters. This is the complement of the Real-SWE play above: one escapes benchmark contamination by going private and licensed; the other stays open and filters-and-discloses instead. Two answers to benchmark gaming, landing in the same window. For anyone fine-tuning SWE agents it is directly usable material — the stated uses are supervised fine-tuning, distillation, and agent-policy benchmarking — including cheap-model arms via the DeepSeek-V4-Flash trajectories.

Go deeper: dataset · dataset paper · exploits paper

Sakana priced the orchestrator, not the model (Japan) — Worth Watching

What happened. On September 11, Tokyo’s Sakana AI released Fugu Max and Fugu Ultra v2 — one orchestration architecture tuned two ways: Max for “the best possible output we can deliver at the lowest possible cost,” Ultra v2 for “the absolute highest capability … on complex, multi-step tasks.” The claims: Fugu Max “best overall score on six benchmarks” (including Terminal-Bench 2.1 and SWEFish, an internal benchmark of Sakana’s own coding challenges) with “performance within striking distance of elite models at two to six times lower cost”; Fugu Ultra v2 best-or-joint-best on five of eight — Chartography 48.3 vs Opus 5’s 27.3, DeepSWE 74.33 — while explicitly excluding Fable 5, Fable 5.1, and GPT-6-Astra from its model pool (training cutoff 2026-08-28). All vendor-run; the pool itself is not enumerated.

Technical signal. The pricing is the most concrete artifact: read from the product page today, Fugu Max lists at $2/$6 per M in/out with $0.25 cached input, flat regardless of context length, and Ultra v2 at $5/$30/$0.50 — each doubling above 272K context — the two-axis frontier sold as a two-SKU menu. The engineering trail sits one layer down: the fugu harness repo committed “add max and ultra v2,” then pinned Codex 0.154.0 about six hours after OpenAI shipped that SDK — Fugu installs as codex-fugu/claude-fugu launchers over an OpenAI-compatible API — and the same morning Sakana’s ALE-Bench merged 29 new model configs (deepseek-v4.1-flash-max, glm-5.3-max, kimi-k3-max, gemini-3.8-flash-high…): this fortnight’s frontier models absorbed as swappable config entries within days of release. Caveats carried, not buried: the fugu repo ships no LICENSE file, HN traction was six points and zero comments, and no independent evaluation of any Fugu claim exists at our endpoints.

Why it matters. This is the composition-layer answer to the question the Big Story asks from inside the model: DeepSeek attacks cost with FP4 KV and a $0.003/M cache line in one 763B artifact; Sakana attacks it above the model, routing across a swappable pool — and claims the frontier’s scores without the frontier models in it. If Ultra v2’s 74.3 (DeepSWE version unstated) really sits beside DeepSeek’s 74.2 on v1.1, the interesting question stops being “which model” and becomes “which composition” — and nothing independent can currently referee that (see My Read).

Go deeper: release post · product + pricing · technical report

Research Radar

No paper from the China window met the read-beyond-abstract bar — the desk’s arXiv sweep failed on fetch limits for the second consecutive cycle (flagged, carried), and no China lab shipped a tech report with this window’s releases. The V4.1-Flash model card is the week’s densest China artifact and is treated in the Big Story and Repo Radar instead. The Japan window also yielded no Radar entry: a six-org arXiv sweep (Sakana, PFN, CyberAgent, llm-jp, Rakuten) returned zero in-window papers, and Sakana’s Fugu technical report is June-vintage — linked from the Elsewhere item rather than featured.

Repo Radar

deepseek-ai/DeepSeek-V4.1-Flash (China)

  • What’s actually in it: 48 shards (~510 GB) of MIT-licensed weights, full transformers config, and an unusually detailed architecture card (CSA2 modes, Engram, FP4 KV, training token schedule). Inspection depth: full card + API docs cross-read; repo tree not opened.
  • What you can run: weights via transformers; API as deepseek-flash (legacy V4-Flash names still route); novita/fireworks listed as inference providers at fetch (one erroring).
  • What’s missing: no code release, no eval harness, no News/changelog on the card, and the asymmetric activation figure (8B/16B) exists only as card prose.
  • Why open it: the canonical artifact for the memory-systems generation — and the cheapest way to test whether static attention-mode assignment survives contact with your long-context workload.

My Read on the Week

(Synthesis and interpretation — not reported fact. Written after a cross-item pass spanning all four desks’ items.)

1. Cost graduated from analysis to price list — on every layer of the stack. Three weeks ago the memory-systems thesis was our synthesis; this week DeepSeek sells KV compression as a rate-card row ($0.003/M cache-hits), and Sakana prices the composition layer itself — $2/$6 for “cheapest sufficient,” $5/$30 for peak, a literal two-SKU Pareto menu complete with a context-length surcharge above 272K. Real-SWE prints per-rollout cost ($2.50–6.96) beside resolution rates; Open-SWE-Traces ships cheap-model arms as a feature. When a 763B flagship, an orchestrator, and a benchmark all lead with cost columns, “capability per dollar” has stopped being our commentary and become the product interface — and the license map inverted in the same stroke: the 763B open flagship is MIT while its 753B rival is bespoke with a $10B gate, so even the terms of reuse now read like a rate card. The one number that would tie the theme together — GLM-Flash’s post-promo usage elasticity — is still uncaptured (China desk, carried).

2. The referee gap widened faster than the referees arrived. This window produced two serious verification instruments — Real-SWE (private, licensed, harness-declared) and NVIDIA’s Open-SWE-Traces v1.2 (open, filtered, exploit-audited) — and, in the same four days, a pile of vendor scores nobody can check: three Terminal-Bench 2.1 self-reports within about five points across three harnesses, Fugu Max “best overall on six benchmarks” including one internal benchmark, Fugu Ultra v2’s 48.3-vs-27.3 on Chartography — landing with six HN points and zero independent runs between them. The nearest thing to a referee the week produced — Real-SWE — is itself author-run and access-gated, and while it does score GLM-5.3 and Kimi K3, neither of this week’s two new arrivals, V4.1-Flash or Fugu Ultra v2, appears in any independent instrument yet. That is the drought’s new shape: not an absence of instruments, but instruments one comparator-generation behind the claims.

3. Composition is the new architecture battleground. DeepSeek attacks cost inside the model — FP4 KV, static attention modes, a 196B named memory component. Sakana attacks it above the model — a swappable pool that excludes Fable 5/5.1 and GPT-6-Astra and still claims their scores, with ALE-Bench absorbing each week’s releases as config entries within days. NVIDIA’s dataset ships China-lab models as the cheap arms of a US-lab training set. Three regions, one move: the unit of competition is shifting from “the model” to “the composition” — orchestrators, memory systems, harnesses, data mixes. Which is exactly why version-labelling (which DeepSWE? which harness? which pool?) is now the load-bearing detail in every table this issue prints.

What I’d Watch Next

  1. V4.1 siblings. “Smallest model in the new architecture family” is a loaded phrase — watch the DeepSeek org and changelog for the larger V4.1 models it implies.
  2. The first independent Fugu eval — and its pool list. Whether Fugu Ultra v2’s DeepSWE 74.3 survives third-party measurement, and whether Sakana ever enumerates the model pool the post leaves unlisted; the DeepSWE version needs confirming before any comparison to V4.1-Flash’s 74.2 on v1.1.
  3. llama.cpp hyv4/STQ1_0, still un-merged (rechecked today: three open PRs, two closed-unmerged) — the 214-GiB Hy4 stays a build project until the format war resolves.

One Thing I’d Experiment With This Week

Price the cache, not the token. DeepSeek’s $0.003–0.006/M cache-hit line is now the floor in every ledger we hold — an order of magnitude below GLM-Flash’s $0.03 and ~7× under Hy4’s $0.042. On any repeating agent workload, cache-hit rate now dominates cost.

  • Setup: one fixed agentic task battery (e.g., 50 issues against a frozen repo), same harness, three arms: DeepSeek-V4.1-Flash, GLM-5.3-Flash (list prices), Hy4 — with prompt-prefix discipline held constant (identical system+context prefix per turn, so cache behavior is comparable). An optional fourth arm routes the same battery through Fugu Max ($2/$6, $0.25/M cached input — Japan item): a fixed battery under prefix discipline is exactly the cheapest-sufficient-output claim it prices, and its per-task cost is comparable because orchestration happens server-side.
  • Conditions: run each arm twice — once with naive per-turn prompts (cache-unfriendly), once with stable prefixes (cache-friendly); record cache-hit tokens, billed split, and task success.
  • Measure: cost per completed task in both regimes; the cache-hit rate delta; whether V4.1-Flash’s task quality holds at low reasoning-effort settings (configurable 1–100). Decision output: a break-even curve — the prefix-engineering effort that pays for itself per model. If the cache-friendly regime cuts cost per task ≥40% on the cheapest-cache model, prompt-prefix discipline becomes a bigger lever than model choice, and that is a deployment decision, not a benchmark.

Corrections

  • [China] GLM-5.3-Flash output-price ratio, carried from Issue 3. Issue 3 (and this issue’s draft before finalisation) said “even at list, Flash output is ~17.6× cheaper than GLM-5.3.” At list prices the ratio is ~8.8× — GLM-5.3 output $4.4/M ÷ Flash output $0.50/M; the 17.6× figure was the promo-era ratio (÷ the promo’s $0.25 output price) and outlived the promotion it described. Re-verified against Z.ai’s pricing page, read 2026-09-13: GLM-5.3 $1.4/$4.4, Flash $0.15/$0.50 (both unchanged; cached-input storage is currently listed “Limited-time Free”). Root cause: number travel between pricing regimes — the same class of failure as Issue 1.0’s §5.1 register. Found by this issue’s whole-issue numbers pass; the China desk re-verifies its corrected claims next cycle per §7.

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/2026-09-13/.

Footnotes

  1. TB 2.1 self-reports — V4.1-Flash 90.6, GLM-5.3 88.2, Hy4 85.4 — come from three different vendor harnesses; the spread is within the noise we’d expect across harnesses, and none is independently reproduced. Comparator-generation spread across vendor tables (Opus 4.8 / Fable 5 / GPT-5.6 Sol vintage) continues per Issues 1–3.↩︎

  2. Real-SWE’s comparators (Fable 5.1, GPT-6 Astra, Gemini 3.8 Flash, Muse Spark 1.3, Grok 4.6) are a different generation spread than this issue’s China-section tables (Opus 4.8 / Fable 5 / GPT-5.6 Sol vintage, per Issues 1–3); cross-table score comparisons across comparator generations are not apples-to-apples (EDITORIAL.md §6). Gemini 3.8 Flash’s own launch date is unresolved at day granularity from this desk’s endpoints — carried as a US ledger open item.↩︎

  3. Sakana’s comparators (Opus 5, Fable 5, Fable 5.1, GPT-6-Astra, Sonnet 5, GPT 5.6 Terra) are a third generation spread in this issue — alongside the China tables’ Opus 4.8 / Fable 5 / GPT-5.6 Sol and Real-SWE’s table above; none of the three spreads is apples-to-apples with the others (EDITORIAL.md §6). And the Fugu Ultra v2 DeepSWE score is 74.3 on “DeepSWE” with no version stated, while DeepSeek’s 74.2 (Big Story) is on DeepSWE v1.1 — the near-tie is suggestive, never like-for-like, until Sakana states the version.↩︎

Citation

BibTeX citation:
@online{dimsum_editorial2026,
  author = {dimsum editorial, the},
  title = {The Memory-Systems Architecture Shipped — at {763B,} Under
    {MIT}},
  date = {2026-09-13},
  url = {https://news.llm101.in/issues/2026-09-13/},
  langid = {en}
}
For attribution, please cite this work as:
dimsum editorial, the. 2026. “The Memory-Systems Architecture Shipped — at 763B, Under MIT.” September 13. https://news.llm101.in/issues/2026-09-13/.