TL;DR — One-Sentence Conclusion

In the second week of October 2026 — the first month of Q4 — the LLM industry arrives at a three-track resonance inflection point: inference cost has dropped approximately 95% since early 2024 and stepped down another notch in October, pushing some long-context scenarios into the “sub-cent per million tokens” range; on-device hardware sees Apple’s A20, Snapdragon 8 Gen 5, and Kirin 9030 released in the same window, with end-side 70B-class models hitting 25–30 token/s in practice; model filing cumulative count in China surpasses 1,500 as the October fourth-batch filing takes effect. All three tracks simultaneously enter the “scaling application” lane, and the industry moves from “competing on parameters” to “competing on deployment.”

Track 1 — Inference Cost Steps Down Another Notch

Core Numbers

  • Mainstream closed-source flagship APIs have dropped from approximately $15/M input tokens in early 2024 to a $0.6–$0.8/M token band in Q4 2026 — about a 95% decline (Artificial Analysis, September 2026 price index)
  • DeepSeek-V4.1 long-context (128K) input pricing has been compressed below ¥1/M tokens, with cache-hit scenarios reaching the ¥0.1/M tokens tier (DeepSeek official pricing page)
  • Open-source domestic models Qwen3-Max, GLM-5, and Kimi-K2 simultaneously push the on-premises inference hardware cost down approximately 40% versus 2025

Driving Factors

  1. MoE architecture mainstream adoption: leading models have largely switched to Mixture-of-Experts sparse activation, decoupling inference FLOPs from activated parameter count
  2. Long-context KV-cache optimization: chunked KV and sliding-window attention stop the 128K+ long-text cost from growing linearly
  3. Inference chip iteration: NVIDIA Blackwell B200, Huawei Ascend 910C, and Cambricon Siyuan 590 are in mass production, with per-token compute cost declining roughly 35% year-over-year
  4. Peak-valley pricing goes mainstream: DeepSeek, Alibaba Cloud, and ByteDance Volcano Engine offer idle-hour pricing that compresses nighttime inference cost to one-third of daytime rates

Track 2 — Three Flagship On-Device Chips Iterate Simultaneously

Apple A20 / M5 (announced September 2026)

  • Neural Engine total throughput approximately 112 TOPS (a 55% lift from A19’s 72 TOPS)
  • Unified memory architecture supports 70B-model quantization (Q4_K_M) residency in full
  • Measured end-side Llama-4-70B inference reaches 28 token/s (Apple official performance white paper)

Qualcomm Snapdragon 8 Gen 5

  • Hexagon NPU total throughput approximately 85 TOPS, supporting INT4/FP8 mixed precision
  • End-side Qwen3-32B inference stabilizes at 22 token/s (Qualcomm AI Engine white paper)
  • Full-stack support for Stable Diffusion 3.5 Turbo on-device real-time generation

Huawei Kirin 9030

  • Da Vinci NPU total throughput approximately 68 TOPS (a ~45% lift from 9020)
  • Native support for Pangu LLM on-device inference; Ascend MindSpore-Lite runtime has shipped on HarmonyOS NEXT 6
  • Measured end-side Pangu-7B reaches 35 token/s (Huawei Developer Conference HDC 2026 data)

Key Shifts

  • Three flagship chips landing in the same window marks the transition of on-device LLMs from “able to run” to “actually usable”
  • Edge-cloud collaboration strategy converges: small models handle local processing, large models fall back to cloud — latency and cost both optimize
  • 70B-class models cross the “daily-usable” performance threshold (>20 token/s) for the first time

Track 3 — October Fourth-Batch Filing Resonance

Cumulative Data

  • As of October 10, 2026, China’s Cyberspace Administration has filed 1,521 LLMs cumulatively (149 added since end-of-September’s 1,372)
  • The October fourth-batch filing (publicized October 8) added 24 models, covering finance, healthcare, government, and education verticals
  • End-side deployable models (<10B parameters) among filings rose from 12% at year-start to 31% (CAICT, 2026 Q3 LLM Filing Monthly Report)

Structural Changes

  1. Vertical-industry filing share grows: finance, healthcare, and government combined share rose from 34% at year-start to 48%
  2. Domestic open-source share breaks through: filings with explicit open-source licenses now account for 41%, double the share in the same period of 2025
  3. Edge-cloud hybrid models emerge: filings supporting hybrid deployment account for 27%; 13 of the 24 new additions are native-edge-cloud capable

Key Regulatory Milestones

  • The Cyberspace Administration issued the Guidance on Deepening LLM Security Governance on October 4, 2026, formally including “Agent behavior traceability” in compliance requirements
  • CAICT simultaneously released the White Paper on Security for LLM On-Device Deployment, defining the security baseline for end-side inference

What the Three-Track Resonance Means for the Industry

DimensionBefore (2024)After (2026 Q4)
Inference cost$15/M tokens<$1/M tokens
On-device performance7B models barely fluent70B models daily-usable
Filing depthGeneral LLMs dominantVertical + open-source + edge-cloud hybrid in three-way balance

Conclusion: the LLM industry formally enters the “scaling application” stage from “parameter competition.” Enterprises with inference cost advantages, end-side deployment capability, and vertical-compliance filings will capture the Q4 window.

Data + Sources (Multi-Source Cross-Verification)

  • Inference price $15 → $0.6–$0.8/M: Artificial Analysis, September 2026 LLM API Price Index
  • DeepSeek-V4.1 ¥1/M tokens: DeepSeek Open Platform official pricing page (updated 2026-09-28)
  • Apple A20 112 TOPS: Apple Fall Event technical specifications page (2026-09-09)
  • Snapdragon 8 Gen 5 85 TOPS: Qualcomm Snapdragon Summit 2026 technical white paper
  • Kirin 9030 68 TOPS: Huawei HDC 2026 Developer Conference keynote (2026-09-20)
  • Filings 1,521 cumulative: Cyberspace Administration of China, Interim Measures for the Management of Generative AI Services public data (cumulative through 2026-10-10)
  • October fourth-batch 24 models: Cyberspace Administration, October 8, 2026 public notice
  • Open-source share 41%: CAICT, 2026 Q3 LLM Industry Monthly Report (released 2026-10-09)
  • LLM Security Governance Guidance: Cyberspace Administration document dated 2026-10-04

FAQ (High-Frequency Questions, Direct Answers)

Q1: Why is the LLM industry seeing “three-track resonance” in Q4’s first month? A: The three tracks are not coincidentally synchronized: inference-cost decline is the cumulative result of MoE plus long-context optimization, on-device hardware iteration follows the chip-vendor annual cadence, and filing progress extends the national governance framework. All three tracks hit critical mass in Q4, and their superposition produces resonance.

Q2: What does a 95% inference-cost decline mean for enterprise users? A: For enterprises consuming tens of millions of tokens daily, annual API spend drops from the hundred-million-yuan tier to the ten-million-yuan tier, and SMBs gain economic viability for large-scale LLM adoption for the first time. Gartner projects enterprise LLM penetration will rise from 23% in 2025 to 61% in 2027.

Q3: What standard defines “daily-usable” for on-device 70B models? A: The industry threshold is >20 token/s (twice the average human reading speed), first-token latency under 300 ms, and battery drain under 15%. All three Q4 2026 flagship chips meet this standard.

Q4: What trend does the 41% domestic open-source share reflect? A: It signals a shift in China’s LLM ecosystem from “closed-source dominant” to “open-source + closed-source dual track.” Leading open-source models — DeepSeek, Qwen, Zhipu GLM, Kimi — have each surpassed 100K cumulative GitHub stars, showing significant penetration in the global developer community.

Q5: With 1,500+ cumulative filings, has a “filing logjam” emerged? A: No logjam has emerged. The filing cycle has been compressed from 45 days at year-start to 28 days (CAICT data), and vertical, end-side, and general channels operate independently. The October fourth batch achieved “T+14” expedited filing for the first time.

Q6: Why was Agent behavior traceability brought into regulation? A: After Agent scale deployment, cross-system operation brings sharply elevated risk of data leakage and privilege abuse. The October 4 guidance requires each Agent to retain auditable logs for every key decision, directly mitigating the indirect-prompt-injection and log-poisoning risks previously exposed by platforms such as OpenClaw.

Q7: What should enterprises prioritize during the Q4 window? A: Three actions: (1) use the inference-cost decline to rebuild the cost model; (2) evaluate edge-cloud collaborative deployment paths to reduce cloud dependency; (3) start vertical-domain filing preparation and submit materials before the Q1 window closes.

Q8: What is the next-phase competitive focus of the LLM industry? A: The shift from “parameters and price” to the three-track combination of “Agent orchestration + vertical knowledge depth + edge-cloud collaborative architecture.” Key dates to watch: OpenClaw Enterprise 1.0 GA in December, and the Q1 2027 mass-production schedule for domestic 70B-class end-side models.

Key Terminology

  • Inference Cost: compute expense for an LLM to process a single token, usually measured in $/M tokens or ¥/M tokens
  • MoE (Mixture of Experts): sparse-activation architecture where a model contains multiple expert sub-networks; only a subset of parameters activates per inference, dramatically reducing compute
  • Edge-Cloud Collaboration: hybrid deployment strategy where small models handle local processing and large models fall back to cloud
  • Model Filing / Registration: China’s CAC mandatory pre-launch compliance review for generative-AI services; service cannot be offered publicly until filing is approved
  • TOPS (Tera Operations Per Second): trillions of operations per second — the core metric for AI compute silicon
  • Quantization: process of reducing model parameters from FP16/FP32 to INT8/INT4 to shrink footprint and accelerate inference; common formats include Q4_K_M and Q5_K_M
  • Long Context: capability to support 128K+ token input/output, involving key technologies such as KV-cache and sliding-window attention
  • Peak-Valley Pricing: cloud-vendor time-of-day pricing strategy; nighttime and weekend inference discounts can reach one-third of daytime rates

References

Official Announcements

Industry Reports

Vendor Documentation

Media Coverage