Compression became the deployment story for frontier open weights

Tencent’s official 2.38-bpw Hy4 build, Qwen’s Apache-2.0 driving foundation model, and the first independent rankings — all within a week of launch.

Compression became the deployment story for frontier open weights: Tencent’s official 2.38-bpw Hy4 build, Qwen’s Apache-2.0 driving foundation model, and the first independent rankings. Issue 2.1, week of 6 September 2026.

Week of August 31 – September 6, 2026 · Issue 2.1 (corrected 2026-09-09 — see Corrections)

The Week in One Sentence

One week after three frontier open-weight launches, the ecosystem shipped its consequences: Tencent put an official 2.38-bit build of the 770B Hy4 preview into a 214-GiB artifact, the first independent arena ranking arrived (Hy4 top-10, ahead of Qwen3.8-Flash-Next), Qwen opened a new front by releasing an Apache-2.0 vision-language foundation model for autonomous driving — and GLM-5.3-Flash’s 50% introductory pricing runs out on September 9.

---

The Big Story: Tencent ships an official 2.38-bpw Hy4 — and the artifact is the release

(An update on Issue #001’s coverage — and a correction to it; see the correction appended to Issue #001.)

What happened. In the final days of August, Tencent’s Hunyuan team published what we carried last week as “a community GGUF quantization” — it wasn’t. It is Tencent’s own release: an official post (“We compressed Hy4-preview from 1.5TB to ~200GiB GGUF… Meet MIX-STQ1_0… Accuracy barely moves vs BF16”) plus an artifact in the AngelSlim repo. This week brought the verification layer: an unusually good independent teardown by traictory (Aug 31), then rankings coverage (SCMP, Sep 2).

Under the hood. The artifact is Hy4-preview-STQ1_0.gguf: 213.66 GiB at 2.38 bits per weight — roughly 7× smaller than the ~1.56 TB BF16 checkpoint — with a conventional 435.2-GiB Q4_K_M build alongside it as the safer default. The method, per Tencent (relayed with the artifact inspected by traictory): calibration data assigns each layer a bit-width; the routed-expert gate and up projections — 97.7% of all parameters — run at 1.3125 bpw in a sparse-ternary format (STQ1_0: one of every four weights forced to zero, 3:4 sparsity, shared scale per 256 weights) on 29 layers and at 2.0625 bpw (IQ2_XXS) on the other 48. Tencent claims its encoder cuts weighted squared error ~90% versus the upstream quantizer on 1,200 expert rows — vendor-claimed, not reproduced. The per-layer recipe files are published.

What the evidence says. Vendor-run, single-run, no error bars — four deltas vs BF16: MCP Atlas 83.7→83.2, SWE-Bench Multilingual 82.9→81.3, MRCR 81.3→81.1, IFBench 73.5→72.5; retention 98.1–99.8%. Read that as: the worst drop (1.6 points) lands on coding, the workload you’d actually buy a 770B model for. The suite tests no long context (advertised 1M), no MMLU/GSM8K, no multimodal, no agentic sweeps. Demand is real though: >26,000 downloads of the GGUF repo in its first days versus ~2,100 for the BF16 original.

The runtime reality. This is not a download-and-run release: “Neither file runs on stock llama.cpp. The hyv4 architecture is not upstream. Apply the patches” — a pinned commit plus two patch files (quant + CUDA kernels). Full residency needs ~214 GiB of VRAM; Tencent measured 204.56 t/s prefill and 20.47 t/s decode on 8×H20. Both figures are vendor-measured; we did not execute the recipe (documented-not-verified).

Why this matters. Issue #001 closed with “compression tooling is suddenly the practical bottleneck for actually deploying any of them.” That took three days to become literal: the most-downloaded form of last week’s flagship is a sub-240-GiB ternary artifact with bespoke kernels. The release also sharpens what “open weights” now means — the artifact is the release, and the engineering surface moved from the model repo into quantization recipes and llama.cpp patches.

What I would test. Exactly the gap traictory names: side-by-side BF16 / Q4_K_M / STQ1_0 on a fixed coding battery at 8K and 128K context. The vendor’s 1.3-point SWE-Bench delta is a single run at (presumably) short context; whether 2.38 bpw holds at 128K is the difference between “usable coding model” and “demo.”

---

Elsewhere in AI Engineering

Hy4’s first independent rankings — and the “ecosystem-driven training” thesis — Worth Reading

What happened. On Tuesday Sep 1, Hy4 preview ranked 8th globally on Code Arena’s WebDev leaderboard, behind Anthropic’s Claude Fable 5¹ and ahead of Alibaba’s Qwen3.8-Flash-Next (9th); the previous Hunyuan generation, Hy3, sat 34th the same day (SCMP, Sep 2). SCMP also relays DeepSWE scores of 64.3 (Hy4) vs 62.7 (DeepSeek-V4 Pro) and 56.6 (Qwen3.8-Max) — the 64.3 matches the model card we verified last week.

Why it matters. Two things. First, the independent layer is finally moving: at launch, the only independent signal was BenchLM’s estimate-marked #7/228; four days later there is a top-10 arena position — still one leaderboard, relayed via press, and we did not open the arena ourselves. Second, Goldman Sachs analysts (relayed by SCMP) named the mechanism: Tencent ships preview models inside its own products, collects usage data, and feeds it into later training rounds — a “product-plus-model” closed loop they call particularly relevant for coding workloads. That is the same loop Tencent described in its supervised self-improvement disclosure (below), running on two surfaces: the model’s training and Tencent’s product fleet.

The Hy4 self-improvement loop, one week later — supervised, disclosed, still unverified — Worth Reading

What happened. TechTimes (Aug 31) details what the “participated in its own training” claim actually means: two supervised buckets — training-pipeline participation (proposing methods, evaluating data strategies, refining eval frameworks) and serving-stack optimization (the model analyzing its own inference bottlenecks and multi-round optimizing operator fusion and inter-GPU communication, which is where the 31.8% throughput claim comes from). Human engineers made the architectural decisions; the Cloud Security Alliance’s June 2026 report is cited to distinguish this from autonomous weight-rewriting, while warning AI-built-AI pipelines become high-value adversarial targets. Tencent reports no incidents.

Why it matters. The claim upgrades from marketing-grade to mechanism-grade — but not to verified: the 31.8% remains self-reported, with no independent reproduction. Also useful reality checks in the same piece: OpenRouter listings show ~26–38 t/s with sub-4-second P50, and BenchLM still has evidence for only 28 of 406 Hy4 benchmark slots (Agentic 9th, Coding 11th; scores marked “Estimated”). One number we could not cross-confirm from a primary: Terminal Bench 2.1 at 85.4 — flagged, not asserted.

GLM-5.3-Flash’s introductory pricing ends September 9 — Worth Watching

What happened. Read from Z.ai’s pricing page today: GLM-5.3-Flash lists at $0.075/M input, $0.25/M output, $0.015/M cache-hit (cache miss billed at input rate; cache storage “limited-time free”) — a 50% promotion that ends 24:00 on September 9, 2026 (UTC+8), after which list prices double. GLM-5.3 itself sits at $1.4/$4.4 in/out.

Why it matters. The Ox Alpha usage record was built on free; the retention question gets answered at $0.15/$0.50 — still dramatically cheaper than GLM-5.3, and it lets us check Z.ai’s launch positioning (“GLM-5.2-class capability at one-tenth the price”): against GLM-5.3’s list, Flash is ≈17.6× cheaper on output and ≈18.7× on input (our arithmetic) — the claim was conservative, not inflated. The cache-hit line ($0.015/M promo, $0.03/M list) is the number that matters for agent loops, and it undercuts Hy4’s $0.042/M from day one.

---

Research Radar

Qwen-Drive-1.0: an untouchable VLM, an inspectable probe, and an expert planner

arXiv:2609.00111 — Zhou et al. (16 authors, Qwen team; v1 Aug 31, 2026). Weights: Qwen-Drive-1.0-4B under Apache-2.0, ~9.1 GB, weights repo created Aug 27, updated Sep 2.

• Problem. Driving stacks and VLMs live apart: one gets 3D structure without language, the other language without inspectable geometry.
• Core idea. Keep a pretrained VLM — Qwen3.5-4B, “entirely untouched” — and attach two external modules: a BEV perception head (joint 3D detection, semantic occupancy, BEV map segmentation, via depth-based view transformation; works across 6-camera nuScenes and 8-camera OpenScene rigs with no rig-specific embeddings) and a Planning Expert — a ~1.1B 32-layer diffusion transformer using flow matching to emit 50 waypoints (5 s at 10 Hz) in a 10-step Euler solve, conditioned on cached VLM attention keys, navigation, ego state, and optional textual reasoning (Zhou et al. 2026). The head doubles as a probe: does a general VLM’s representation already carry enough 3D to anchor BEV tasks?
• Evidence (vendor-run). nuScenes mAP/NDS 43.95/42.83; OpenScene 43.45/44.16; NAVSIM PDMS 90.7 (91.4 best-of-6); WOD-E2E test RFS 7.91; driving VQA LingoQA 77.8; general VLM capability retained (MMBench 85.5, MMMU 72.7). Training: four stages from frozen-VLM head pretraining through RL (group-relative policy gradient; 15K NAVSIM + 15K PAI-AV + 479 WOD-E2E scenarios). Decode settings near-greedy, and LingoQA was re-scored with Qwen-Plus as judge — both stated on the card, both worth discounting for.
• Limitations (authors’). Sparser temporal sampling than the dense-history baseline may limit short-horizon responsiveness; near the NAVSIM ceiling, gains “may increasingly reflect adaptation to the scoring function.”
• Why I bookmarked it. It’s the packaging bet, not any single ingredient: VLM untouched + inspectable BEV interface + expert planner, shipped Apache-2.0 at ~9 GB — into a domain (embodied/driving) where open weights were thinnest. And the license is the story inside the story: Qwen shipped Flash-Next under qwen-community-1.0 ten days ago and this under Apache-2.0 — per-artifact licensing, confirmed twice in one fortnight.
• Read this if… you build multimodal systems that must reason about physical scenes, or you track where the open-weights firehose points next.

---

Repo Radar

AngelSlim/Hy4-preview-GGUF (ModelScope)

• What’s actually in it: the STQ1_0 (213.66 GiB, 2.38 bpw) and Q4_K_M (435.2 GiB) artifacts plus published per-layer recipe files. Inspection depth: via traictory’s independent teardown + listing; we did not open the repo directly.
• What you can run: a llama.cpp build from a pinned commit with two patch files, on ≥214 GiB VRAM; vendor-measured 204.56 t/s prefill / 20.47 t/s decode on 8×H20. Documented-not-verified by us.
• What’s missing: stock-llama.cpp support (hyv4 isn’t upstream), any long-context/math/multimodal/agentic benchmark, error bars.
• Why open it: the canonical reference for sub-240-GiB frontier serving — the recipes are the transferable IP.

Qwen/Qwen-Drive-1.0-4B (Hugging Face) + QwenLM/Qwen-Drive-1.0

• What’s actually in it: unsharded BF16 safetensors for the root VLM, planner-sft/, planner-rl/, and the perception/ head (4,539,265,536 params total); code, demo data, and docs on GitHub with pip install -e . and a bundled demo (four planning scenes, six perception frames). Inspection depth: model card + paper; GitHub tree not opened.
• What you can run: the demo pipeline end-to-end; SFT and RL planner checkpoints ship separately, which is unusual and welcome for ablation.
• What’s missing: context length is not stated on the card; no training data release.
• Why open it: the cheapest entry point into embodied VLM experimentation anyone has shipped this year — at 9.1 GB it runs where the 770B-class models can’t.

---

My Read on the Week

(Synthesis and interpretation — not reported fact.)

1. The deployment interface for frontier open weights is now a quantization recipe, not a checkpoint. Last week’s thesis — “the repo is a manifest; the real artifact is serving-stack support” — compressed one step further: the artifact is a GGUF plus kernel patches, and adoption data says so (26K GGUF downloads vs 2K BF16). “Is it open?” has become “how many bits, on which patched runtime, over how much VRAM?” — and the answer ships from the vendor’s own compression toolkit within 48 hours of launch. Expect every frontier drop to arrive with a ternary build and a patch queue from now on.

2. The verification layer is arriving faster — on a schedule you can set your watch by. Launch-plus-4-days brought the first arena ranking; launch-plus-3 brought the compression teardown. Goldman’s “ecosystem-driven training” framing and Tencent’s own supervised-loop disclosure are the same loop on two surfaces (product fleet and training pipeline) — which means the models being ranked are mid-training-cycle artifacts whose next checkpoints will be shaped by exactly the usage the rankings measure. Independent evidence is no longer scarce; stable evidence is, because the thing being evaluated is moving.

3. Per-artifact licensing is now the rule, not the exception. Qwen: qwen-community-1.0 on Flash-Next (Aug 24), Apache-2.0 on Qwen-Drive-1.0 (Aug 31). Kimi: bespoke permissive-with-thresholds. Tencent: Apache-2.0. Hy4’s Apache-2.0 vs Qwen’s community terms on the same week’s releases kills any lazy “Chinese labs license like X” shorthand for good — the only workable rule is the one our ledger enforces: read the license field of the specific artifact, every time.

---

What I’d Watch Next

1. September 9, 24:00 UTC+8. GLM-5.3-Flash’s promo ends; list prices double. Watch OpenRouter usage share through the flip — the first real price-elasticity datapoint for the record-setting launch.
2. GLM-5.3 open weights, rumored mid-to-late September (Tier-D lead; no Z.ai primary yet — do not act on it). If it lands MIT like the Flash variant, the open-weights cascade continues; the license field is the first thing to read.
3. llama.cpp upstreaming of hyv4/STQ1_0. The moment the patches merge, the 214-GiB Hy4 becomes a standard artifact rather than a build project — and the side-by-side long-context test (the open item) gets cheap enough for the community to actually run.

---

One Thing I’d Experiment With This Week

Measure the ternary tax where it bites: long-context coding. Tencent’s suite says 2.38 bpw costs 1.3 points of SWE-Bench Multilingual — single run, short context, vendor harness. Nobody has published the 128K number.

• Setup: fixed repo + task battery (e.g., 30 issues with known patches, dependencies included), three builds — BF16 (or FP8 API as proxy), Q4_K_M (435 GiB), STQ1_0 (214 GiB) — same patched llama.cpp commit, same hardware (needs ~8×32 GB class GPUs or the H20 rig), identical sampling.
• Conditions: run the battery twice per build at 8K context and once per build at 128K (fill the window with real repo files, not filler).
• Measure: pass rate per build per context; the widening of the BF16→STQ1_0 gap between 8K and 128K (the vendor delta is the short-context floor, not the slope); tokens-to-fix and cache behavior. Decision output: the $/correct-fix curve per bit-width — if STQ1_0’s gap stays ≤2 points at 128K, a 770B-class coding model fits on hardware that used to hold a 70B; if it widens 3×, ternary is a batch-inference story only.

Corrections

Issue 2.0 → 2.1, 2026-09-09. One correction, filed from the next issue’s research cycle.

1. What I’d Watch Next, item 2 (GLM-5.3 open weights) — a discovery miss, not just an unconfirmed rumor. v2.0 wrote GLM-5.3 open weights were “rumored mid-to-late September (Tier-D lead; no Z.ai primary yet — do not act on it).” In fact the weights were already public: zai-org/GLM-5.3 (753B, FP8-native) has an initial commit of August 28 and community evaluation pull requests from August 28–29 — nine days before this issue published. “No Z.ai primary yet” was therefore wrong at publication time. The hedge (“do not act on it”) limited the damage but the discovery sweep missed it. Corrected in Issue 3. Root cause: discovery checked Z.ai’s blog, docs, and release notes but not the Hugging Face organization page — the registry has been updated to include those endpoints so the org sweep is routine.

---

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 claims ledger lives at research/2026-09-06/claim-ledger.md.

References

Zhou, Xin, Zongchuang Zhao, Zhibo Yang, et al. 2026. “Qwen-Drive-1.0: An Initial Step Towards a Vision-Language Foundation Model for Autonomous Driving.” arXiv. https://arxiv.org/abs/2609.00111.

Footnotes

1. Comparator names vary by source vintage — Issue #001’s tables referenced Opus 4.6 Max / Opus 4.8 / Opus 5; SCMP’s leaderboard references Claude Fable 5. Cross-source score comparisons across comparator generations are not apples-to-apples.