In one seismic week, open-source AI experienced an unprecedented double shock. On July 17, Moonshot AI released Kimi K3 — a 2.8-trillion-parameter behemoth that became the world's largest open-source model and topped the Berkeley Frontend Code Arena with a score of 1679. By late July, DeepSeek V4's official release loomed, bringing 1-million-token context as a standard feature across all tiers while introducing peak/off-peak pricing that pushed inference costs to record lows. Together, these two models are rewriting the fundamental rules of AI competition.
Kimi K3: The First Open-Source Model to Top the Code Arena
Core Specifications
Kimi K3 employs a Mixture-of-Experts (MoE) architecture with a total of 2.8 trillion parameters, 896 experts (16 activated per token), native vision understanding, and a context window of 1 million tokens.
Benchmark performance highlights:
Frontend Code Arena: 1679 points, #1 globally, surpassing Claude Fable 5 and GPT-5.6 Sol
Artificial Analysis Intelligence Index: 57 points, #3 globally, behind only Claude Fable 5 and GPT-5.6 Sol
DeepSWE Long-Horizon Development: 67.5% task completion rate, #3
Code Arena Frontend Development: 76% win rate, #1
Architectural Innovation: KDA + AttnRes
K3's foundational capabilities rest on two proprietary architectural breakthroughs:
KDA (Kimi-Dense Attention) Hybrid Linear Attention: Shatters the quadratic computation bottleneck of traditional Transformers in long-sequence processing, boosting 1M-token decoding speed by 6.3×. This means K3 doesn't just "handle" ultra-long context — it handles it fast.
Attention Residuals (AttnRes): Adds cross-layer information retention on top of traditional residual connections, improving training efficiency by 25% and solving the long-standing problem of information loss in deep networks. One of the co-first authors of this innovation is a 17-year-old researcher, Chen Guangyu — a fact that itself illustrates how AI research barriers are being redefined.
Long-Horizon Coding: A Paradigm Shift in AI Autonomy
K3's most striking capability is long-horizon coding. In an official demonstration, K3 ran autonomously for 48 consecutive hours, independently completing the full-flow design of a 45nm chip — from architecture to verification — with zero human intervention. This isn't glorified code completion; it's the critical leap from AI as a "Q&A tool" to AI as an "autonomous worker."
For developers, this means you can hand K3 a complete project specification and it will work for hours or even days with minimal supervision, automatically writing code, testing, debugging, and generating documentation end-to-end.
Open Source and Pricing
Full model weights are scheduled for release before July 27, 2026. API pricing:
Input: $3/million tokens ($0.30 cache hit)
Output: $15/million tokens
Coding scenario cache hit rate exceeds 90%, reducing effective input cost to roughly 1/4 of the standard price
Important note: Due to compute overload, Moonshot AI suspended new C-end subscriptions on July 19, prioritizing existing members. However, API access and open-source weight releases remain unaffected.
DeepSeek V4: Making 1M-Token Context the New Baseline
Dual-Version Architecture
DeepSeek V4 launches in two versions spanning flagship to lightweight use cases:
SpecificationV4-ProV4-Flash
Total Parameters1.6T284B
Active Parameters49B13B
Context Length1M tokens1M tokens
MoE Experts1 shared + 384 routed1 shared + 256 routed
Inference FLOPs (vs V3.2)27%10%
KV Cache Footprint (vs V3.2)10%7%
Training Data33T tokens32T tokens
A telling comparison: V4-Pro's per-token inference compute at 1M context is just 27% of its predecessor, with KV cache compressed to 10%. One-million-token context has transitioned from a "technical curiosity" to an "affordable engineering reality."
Three Architectural Breakthroughs
1. mHC: Manifold-Constrained Hyper-Connections — Preventing Trillion-Parameter Collapse
Traditional residual connections have remained virtually unchanged since He et al.'s 2016 ResNet paper. When model depth pushes past hundreds of billions of parameters, standard residuals expose clear signal propagation instability.
V4's solution: Constrain the channel mixing matrix to the Birkhoff polytope — the manifold of doubly stochastic matrices where each row and column sums to 1. This constraint provides two mathematical guarantees:
Spectral norm ≤ 1: Hard upper bound on residual propagation, truncating gradient explosions at the source
Multiplicative closure: Such matrices are closed under matrix multiplication, maintaining stability even when stacked beyond 100 layers
This isn't about a 0.5% improvement on some benchmark — it's about laying a stable foundation for the entire architecture's scaling, making it possible to train a 1.6T-parameter model to completion without collapse.
2. CSA/HCA Hybrid Attention — The Efficiency Core of Million-Token Context
V4 alternates between two attention layer types, each serving a distinct function:
HCA (Heavily Compressed Attention): 128:1 compression ratio, providing global context — a "super-wide-angle lens" scanning the full picture
CSA (Compressed Sparse Attention): 4:1 compression ratio, preserving detail — a "macro telephoto lens" focusing on key线索
CSA's four-step information pipeline: Raw sequence → KV compression (every m tokens compressed to 1) → Lightning Indexer + Top-K selection → Core Attention computation → Grouped Output Projection. This design allows the model to maintain global awareness while dramatically reducing computational overhead.
3. Deep Domestic Compute Adaptation — The Migration from NVIDIA to Ascend Begins
V4 is the first DeepSeek model to explicitly validate Huawei Ascend NPU performance in its official technical report. V4's chosen FP4 precision format happens to be natively supported by Huawei's Ascend 950 chip. DeepSeek has stated that V4-Pro pricing will drop significantly once Ascend 950 super-nodes ship at scale in the second half of 2026.
Huawei confirmed the same day: "Ascend has consistently supported the DeepSeek model series. Through tight chip-model technical collaboration, all Ascend super-node products now support the DeepSeek-V4 series."
This marks the official start of migration from NVIDIA's CUDA ecosystem to Huawei's CANN framework — an impact on the domestic compute supply chain that cannot be overstated.
Peak/Off-Peak Pricing: AI Inference Enters the Era of Precision Operations
V4's official release introduces peak/off-peak pricing for the first time — a landmark shift from flat-rate to precision-managed inference services:
Peak hours (9:00–12:00 / 14:00–18:00 Beijing Time): 2× standard price
Off-peak hours: Standard price
VersionOutput (Off-Peak)Output (Peak)Cache-Hit Input
V4-Pro$0.87/M$1.74/M$0.435/M
V4-Flash$0.28/M$0.56/M$0.0028/M
Compared to Claude Fable 5's $50/million output tokens, V4-Flash at off-peak rates is 1/178th the cost. Even with peak pricing factored in, DeepSeek remains the undisputed price disruptor.
Legacy Model Migration Timeline
As of July 24, the legacy model names deepseek-chat and deepseek-reasoner have been retired, automatically mapping to V4-Flash's non-reasoning and reasoning modes respectively. Developers don't need to modify code, but should be aware of the pricing structure changes.
Industry Landscape: Open Source Enters an Arms Race
Three Key Signals
Signal 1: Open Source Surpasses the Most Expensive Closed Model in a Specific Capability
K3's Code Arena score surpassing Fable 5 shatters the "closed-source is always stronger" consensus. This proves the open-source path isn't about catching up — it can achieve localized superiority. For the entire open-source community, this is a powerful validation.
Signal 2: Inference Costs Enter the Era of Diminishing Marginal Returns
V4's CSA/HCA architecture reduces 1M-context inference costs to 1/10th of traditional approaches. When the cost of processing a million tokens drops from $50 to $0.28, the marginal cost of AI approaches zero. This fundamentally changes enterprise AI deployment decisions — batch processing, full-text analysis, and long-document understanding scenarios that were previously cost-prohibitive are now economically viable.
Signal 3: Competition Dimensions Are Escalating
The battleground is shifting from raw parameter counts to a multi-dimensional contest of "cost-efficiency × context length × agent capability × open-source ecosystem." K3 excels in agent capability, V4 crushes on cost efficiency, and Alibaba's Qwen3.8-Max at 2.4T parameters (preview) has also entered the fray. Three Chinese mega-models releasing in the same week is unprecedented in AI history.
Practical Implications for Enterprises and Developers
Heavy API users: Migrate batch tasks to V4 off-peak hours (avoid 9–12 / 14–18 Beijing Time) to save 50%+ on costs
Agent developers: K3's long-horizon coding + 1M context makes complex multi-step agents a reality; V4-Flash's ultra-low pricing suits high-frequency agent invocations
On-premise deployment: K3 weights go open-source on 7/27, enabling self-hosted deployment with data never leaving your network — though 2.8T parameters requires a 64+ GPU cluster
AI Search & GEO: Longer context windows mean AI search engines can more completely understand and cite web content. GEO strategy must shift from "keyword density" to "content completeness and authority"
Real-World Testing Data
Based on developer feedback from the V4 gray-release testing:
V4 Pro overall performance is close to Claude Opus 4.8 level, with coding capabilities approaching GPT-5.6 Sol
V4 shows significant improvement in 3D and SVG generation; single-shot game and interactive app outputs are reasonably playable
For the same task, V4 requires more iteration rounds than Fable 5, but given the 5×+ price difference, the cost-performance ratio remains compelling
K3 excels in long-horizon autonomous tasks, but still trails Fable 5 and GPT-5.6 Sol in deep reasoning
Outlook: The Competition Has Only Just Begun
The week of July 2026 may well be the watershed moment in open-source AI history. When a 2.8T-parameter model goes open-source, million-token context becomes standard equipment, and inference costs drop to 1/100th of previous levels, the barriers to AI adoption are being dismantled entirely.
The next round of competition won't center on "whose model is bigger" but on "who can make AI cheaper, more stable, and more autonomous in completing real work." On this dimension, K3 and V4 are just the beginning.
Key dates to watch:
July 27: Kimi K3 full weight release; community distillations and quantized versions expected within a week
Late July: DeepSeek V4 official launch with peak/off-peak pricing going live
H2 2026: Huawei Ascend 950 super-nodes ship at scale; V4-Pro pricing expected to drop significantly
The second half of open-source AI has only just kicked off.