163: DeepSeek V4 Explained: The Infra Whale, 1M-Token Context Made Real, and Extreme Efficiency Optimization
Summary
刘逸峰 argues that DeepSeek V4 is not an R1-style “new paradigm,” but takes 1M-token context from theoretical feasibility to acceptable cost along the test-time scaling path. 赵晨阳 argues that, rather than debating methodological paradigm shifts, the question is whether a model opens up a new capability domain; V4’s real commercial value is coupling hybrid sparse attention, MHC, Muon, FP4, and a new inference stack and getting them to run end to end—a systems-engineering agenda that may still define the next 1–2 years.
V4’s core economics come from long context, not across-the-board cost cuts that every request can enjoy. The report puts per-token FLOPs at about 27% of V3.2 and KV Cache use at about 10%, but the advantage only becomes material as context lengthens; an 8K-input, 4K-output test will struggle to show the architecture’s real value, while Agents approaching 1M tokens are the primary beneficiaries.
Lower per-token cost is being partly offset by higher token consumption. 刘逸峰 speculates that training rewards care more about task completion than token count, while aggressive token-wise compression may force the model into longer reasoning to recover lost information; he calls this “the aesthetic of watering flowers with a pressure washer”（拿着高压水枪浇花的美感） and argues that “1M tokens should be spent on more important things”（100万token应该拿去做更重要的事情）.
With 1.6T total parameters and roughly 49B active parameters, V4 pushes MoE’s decoupling of knowledge capacity and inference cost to an active ratio of about 3%. Against roughly 5.5% for V3, 4.3% for MiniMax M2.7, 5.3% for GLM-5.1, and 3.2% for Kimi K2.6, V4 Pro is more aggressive; the cost is a nonlinear rise in the difficulty of balancing expert training, routing tokens, and handling low-level communication, leaving the real moat in Infra execution.
V4 abandons MLA for an SWA stack augmented with CSA/HCA, showing that LLM architecture is far from converged. CSA compresses the sequence 4:1 before selecting Top-K positions, while HCA compresses it 128:1 and then retains dense attention; with different paths preselected by layer, V4 sacrifices architectural neatness for a more controllable balance among long-range information, prefix caching, and production deployment.
Muon, MHC, TileLang, and FP4 together make up the harder-to-copy part of V4. Muon optimizes matrices as wholes but requires distributed sharding, aggregation, and Checkpoint logic to be rebuilt; MHC widens inter-layer information flow and stabilizes training with Sinkhorn-Knopp; FP4 uses fake quantization in training and real quantization at sampling to reduce an RL sampling bill that can exceed 70%. “Every term maps to a paper.”
Model capabilities have compressed into an extremely narrow competitive band, so long-term winners will be determined more by data flywheels, Agent products, and cost structures. V4’s internal online experiments put it around Opus 4.5, still behind Opus 4.6 and GPT-5.5, while its external Arena rank is around 23; 赵晨阳’s experience after switching from Claude Code to Codex was that “leaving Claude did not make the world any worse”（离开Claude，这个世界完全没有下滑）, suggesting that small leaderboard gaps do not necessarily translate into meaningful Use Case gaps.
Deep dive
1. V4 Clearly Moves Past V3, but Doesn’t Pull Away from the First Tier
刘逸峰’s hands-on experience is that V4 is much better than V3 at mathematical reasoning, code generation, and following Agent instructions, with substantially fewer hallucinations in particular; against the leading Chinese open-source models of the same period, however, it feels close to Kimi K2.6 and has not created a clear breakaway.
Code remains the easiest place to feel the gap: V4 is a clear improvement over V3, but it still trails the strongest overseas models. 程曼祺 therefore asks how much real product advantage a one-off leaderboard lead can represent if China’s first-tier models all feel broadly similar in use.
赵晨阳’s team had both the inference and RL pipelines running on the day V4 launched. He says the SGLang ecosystem may already be deployed across more than 400K GPUs globally, so the focus is not chat experience but whether the new architecture can enter production workloads.
2. DeepSeek Again Gives Open-Source Infra “Another Year of Life”
赵晨阳’s quip is that “every DeepSeek release keeps open-source research and Infra alive for another year”（DeepSeek每年的发布，都会为开源社区的研究和Infra续命一年）. The open-source community took roughly 1 year to fully digest details such as MLA and DeepSeek MoE in the V3 era; V4 leaves another substantial implementation burden.
From V2 to V3, the inference-side work was mainly extending MLA and DeepSeek MoE. In V4, the attention structure itself changes materially: sliding-window attention is combined with dual compression, forcing extensive optimization of prefix caching, speculative sampling, and KV Cache management.
“Infra whale” is therefore not rhetoric about model size but a judgment about the workload that spills over. Every combinatorial innovation in the algorithm report becomes a long-term engineering task for inference engines, RL systems, kernels, and hardware adaptation.
3. V4 Extends the R1 Paradigm; Its Value Is Engineering, Not Renaming an Era
刘逸峰 says the term “paradigm shift” has become somewhat narratively overloaded in AI. If a paradigm is reserved for sparse events such as Transformer, Scaling Law, RLHF, Constitutional AI, and Test-Time Scaling, V4 has not opened a new one.
R1’s historical significance was making Long-Scale Reasoning and test-time scaling work in the open-source world. V4 likewise says it is following the R1 route, tackling 1M-token context, low-cost inference, and complex systems integration—harder problems, but more engineering-heavy ones.
刘逸峰 believes the existing LM paradigm still has “a very large amount of room.” The main theme of the next 1–2 years may be breaking 1M tokens into many granular optimizations and getting the whole system to run stably. “Every step will be difficult,” but the commercial value will be just as large.
赵晨阳 adds that, rather than explaining paradigm shifts methodologically, it is more useful to ask whether a model has opened a new capability domain. V4 is far less shocking than R1, and less shocking than Kimi’s initial demonstration of long-text capability.
4. What’s Truly Scarce Is Discovering New Capabilities, Not Merely Saturating Old Tests
刘逸峰 offers another definition of “paradigm”: finding a new capability domain. Once long context, Agent capability, or lower hallucination rates were clearly articulated, leading models typically caught up within 6–12 months. “It’s not that we don’t know whether it can be done; we don’t know what else needs to be done.”
V4 does not introduce an entirely new capability domain; it makes known targets more complete and cheaper. When 程曼祺 raises AI consciousness and autonomous behavior, 刘逸峰 does not develop them as current directions, instead pivoting to the reduction of pointless reasoning.
刘逸峰 argues that many tasks do not need excessive reasoning. The Agent community needs to figure out how to use context more efficiently instead of assuming that more than 1M tokens is automatically better.
5. The Missing Cost Figure Signals That DeepSeek No Longer Needs “Cheap” to Define Itself
V3 disclosed $5.57M for its final training run, but neither the V4 report nor the official blog gives a comparable figure. 刘逸峰 sees the silence as a signal that DeepSeek may want the model’s capabilities to do the talking rather than continue relying on a low-cost narrative.
He notes that $5.57M covered only the final training run. Frontier technology exploration, comparative validation, headcount, and data make up the bulk of R&D, so total development spending was “certainly dozens of times” the final training cost; an isolated disclosure of the last run has limited meaning.
V4 was originally scheduled for the Spring Festival but arrived roughly 2-plus months late. 赵晨阳 suspects that launching four mutually coupled new Features—hybrid attention, MHC, Muon, and FP4—created a “combination explosion problem,” with any single component requiring large-scale Debugging to productionize.
On domestic chips, the report explicitly says the parallelization scheme was validated on Huawei Ascend, which is an inference-side technical validation. It does not say whether training used domestic chips. 刘逸峰 notes that the outside world still mostly speculates that Nvidia was used for training, but that is not a fact confirmed by the report.
6. Skipping Multimodality Is a Deliberate Trade-Off Under a Complexity Budget
刘逸峰 believes V4 may have chosen to stay focused on text because pursuing multimodality at the same time would materially increase model and training complexity. With the architecture already complicated and domestic-chip adaptation still required, adding multimodal structure would further multiply the engineering workload.
This is not a rejection of multimodality’s value. When the discussion later turns to StepFun, 赵晨阳 still credits it for investing in multimodality early, but says the field is currently “stuck in the middle”; whether native multimodality is worth the complexity burden in a super-large model remains unsettled.
From the author acknowledgments, 刘逸峰 estimates departures at roughly 5%. He reads “not swayed by praise, not frightened by criticism”（不诱于誉，不恐于诽） and “follow the Way and hold oneself upright”（率道而行，端然正己） as a shared long-termism among the team and its researchers, not launch copy.
7. V4’s Evaluation Is Candid, but the Leaderboards Show No Dominance
DeepSeek ran an internal Arena-like online experiment in which engineers selected models for real tasks and submitted feedback. V4 was broadly around Opus 4.5, still behind Opus 4.6 and GPT-5.5; the report also says roughly 9% of engineers would not choose V4 Pro as their first choice.
External rankings are less flattering. 程曼祺 cites an Arena position around No. 23, behind Qwen 3.5 Max, MiMo V2.5, and GLM-5.1; the Artificial Analysis Intelligence Index is around 52, below MiMo V2.5 and Kimi K2.6.
赵晨阳 is unwilling to overinterpret such small gaps: “Choosing a model requires a certain amount of faith in AI.” After switching from Claude Code to Codex, he found no material collapse in his work and concluded that models have been compressed into an “extremely narrow space,” making precise rank ordering increasingly difficult.
8. The Real Moat in Coding Is the Usage Data Flywheel
From internal adoption rates, 刘逸峰 extends a broader judgment: only a handful of companies in the world have a programming data flywheel. To build a leading Coding model, the model must enter real programming work, and usage itself is the best way to acquire that data.
程曼祺 asks whether China’s open-source model base and user self-hosting cause model vendors to lose data. 赵晨阳 says individuals cannot shoulder deployments of 1T-scale models at the level of H200, B200, or Ascend 910B; most calls still run through third-party clouds and APIs, so they usually leave a Trace.
赵晨阳 also acknowledges that, to his knowledge, many users in real domestic workflows use U.S. closed-source models.
For Claude Code to remain tied to its own model over the long run, that base model must stay in the first tier for Coding, interaction, and Agent frameworks. After GPT-5.5 launched, some corporate executives were already considering moving To B workloads from Claude to GPT, showing that product leadership is not secure.
9. 27% FLOPs and 10% KV Cache Only Pay Off at Long Contexts
The V4 report’s headline figures are per-token inference FLOPs at about 27% of V3.2 and KV Cache usage at about 10%. 赵晨阳 stresses that the premise is sufficiently long context, because the marginal gains from compression and sparsification grow with sequence length.
If a third party tests only 8K input and 4K output, V4 will not show its extreme advantage over V3.2. But a few thousand tokens may now not even fit the System Prompt, and everyday Coding and complex-text tasks are steadily moving into V4’s advantage zone.
Agent is the most direct deployment scenario: multi-step tasks can easily approach 1M tokens. Lowering compute and cache costs per token is what turns a long-range workflow that was theoretically feasible into a cost-acceptable product.
10. Cheaper Tokens Don’t Mean Cheaper Task Completion
刘逸峰 adds the key counterpoint to the efficiency narrative: many users feel that V4 consumes more tokens to solve the same problem. Even if FLOPs and KV Cache fall at equal token counts, end-to-end costs may not improve as much as the headline figures suggest.
His 2 hypotheses are that the training objective favors “complete the task and receive the reward” without sufficiently penalizing token length, and that more aggressive token-wise attention compression may force the model to spend additional tokens reconstructing discarded information and reasoning.
刘逸峰 describes the situation as “the aesthetic of watering flowers with a pressure washer.” A report similar to K1.5 had already tried length penalties, yet responses continued to grow inexorably, suggesting that the model is faithfully copying the verbosity embedded in its training data and reward mechanisms.
刘逸峰 is not rejecting 1M tokens; his position is that “1M tokens should be spent on more important things.” Infra can make context throughput faster, but the Agent community still needs rules for saving, reusing, and organizing context.
11. An ~3% Active Ratio Pushes MoE Toward a New Engineering Limit
V4 retains the 2 main lines of DeepSeek MoE and MTP. V4 Pro has roughly 1.6T total parameters, about 49B active parameters, and an active ratio of about 3%. MoE continues to decouple knowledge capacity from inference cost: make total parameters large while keeping each individual computation as small as possible.
The cross-model comparison makes the aggressiveness clear: roughly 5.5% for V3, 4.3% for MiniMax M2.7, 5.3% for GLM-5.1, and 3.2% for Kimi K2.6, with V4 Pro the lowest. 程曼祺 emphasizes that moving from 5% to 4% is not as difficult as moving from 4% to 3%; the relationship is not linear.
赵晨阳 prefers to look at the reciprocal: the system has moved from roughly 5:1, 10:1, or 20:1 to the 30s, approaching 40:1. A large pool of redundant experts still has to be trained relatively evenly, while tokens must be routed quickly, communicated across devices, and reaggregated.
Low activation is not always better; it can introduce load imbalance, undertrained experts, and routing jitter. V4’s ability to train stably at this ratio looks more like proof of joint Infra-and-algorithm capability than a standalone model-size game.
12. Hash Routing in Early Layers Solves Premature Expert Clustering
赵晨阳 points out that V4 uses hash routing in the first few MoE layers rather than relying entirely on conventional routing. The algorithmic goal is to prevent shallow-layer tokens from concentrating in a small number of experts and to spread early training signals more evenly across the expert pool.
This design, combined with the very low active ratio, raises the engineering bar. The more experts there are and the fewer that activate per pass, the faster any routing skew becomes undertrained experts or device load imbalance; low-level kernels must also ensure that tokens reach experts and complete aggregation without creating communication hotspots.
The ratio itself is therefore not a copyable recipe. Different teams can choose a more conservative active rate; there is no simple right or wrong. DeepSeek is trading higher engineering risk for lower inference cost and greater knowledge capacity.
13. DeepSeek Also Abandons Its Own Techniques: Engram Didn’t Make V4
V4 does not include every technology DeepSeek has previously proposed. Engram combines 2 or more consecutive tokens into a new encoding, whereas traditional Input Embedding encodes a single token; in theory, this can strengthen long-range signals in “needle-in-a-haystack” tasks.
赵晨阳 says their experiments found that the method required a large number of additional parameters while delivering limited capability gains, and also created extra Infra challenges. It looks more like signal enhancement than something likely to become a general choice for Frontier Labs in the near term.
The trade-off is consistent with V4’s overall style: the team is willing to launch 4 highly coupled innovations at once, but also willing to leave a self-developed method out of V4 when the gains are insufficient. The real filter remains whether scaling up contributes capability or lowers cost.
14. Hybrid Attention Lets Different Layers Capture Key Details and Global Semantics
Every V4 layer runs Sliding Window Attention and adds 1 long-range attention mechanism, either CSA or HCA. CSA takes the sparse route, compressing the sequence 4:1 before selecting Top-K key positions; HCA is more aggressive, compressing 128:1 and then retaining dense attention.
Whether a layer uses CSA or HCA is predefined, so the same context appears through different views at different layers. 赵晨阳’s interpretation is that sparse layers lock onto key tokens, while dense layers provide a semantic overview after compression.
This is also important context for V4’s decision to abandon MLA. 刘逸峰 speculates that stacking large-scale token-wise compression on top of MLA’s low-rank KV compression would make the low-level implementation too complex, so the team returned to a foundational structure closer to conventional MQA and recombined it.
15. Large Models Still Bet on Sparse Attention Because Linear Attention Has a Long-Range Ceiling
赵晨阳 believes sparse attention is more controllable from an engineering standpoint and easier to make compatible with existing KV Cache and Prefix Cache. There is still no Frontier model at a comparable scale proving that pure linear attention can stably push the capability ceiling.
刘逸峰 uses 1024 tokens as an example: if linear attention updates the hidden state sequentially, information from token 1 may be compressed 1023 times. With an SWA window of 128, only about 8 hops are needed, allowing 2 positions to exchange information with near-zero loss.
Linear attention still suits speed-focused small and midsize models such as Qwen 3.5. Larger models such as V4 and GLM-5.1 continue to use sparse or sliding-window attention. Kimi may also consider a hybrid structure, because full attention is still needed as a backstop against compression loss.
MiniMax offers a case with a commensurate trade-off: M1 used linear attention, while the first model in M2 returned to full attention. Linear is faster; sparse is harder to implement but has a higher ceiling. Hardware and task distributions will continue to reshape the optimum.
16. Shadow Redis Translates Attention Innovation into a Deployable Cache System
For inference teams, the hardest part of hybrid attention is prefix-cache consistency. 赵晨阳’s team developed Shadow Redis to maintain compressed states separately for SWA, CSA, and HCA, creating 3 heterogeneous KV Cache pools.
These states must stay synchronized across Prefill, Decode, and Speculative Decode. Otherwise, the same prefix may be difficult to reuse because it takes different attention paths. This problem did not exist in the MLA era and shows that V4’s adaptation complexity has moved up another level.
The team also has HiSparse, which pre-offloads KV Cache to Host Memory for sparse attention and can deliver more than 5x throughput in long-context scenarios. 赵晨阳 believes only this kind of low-level support can turn 27% FLOPs and 10% cache usage into production commercial value.
17. Muon Upgrades from Per-Parameter Updates to Whole-Matrix Optimization
刘逸峰 explains that AdamW maintains momentum and normalizes updates element by element, which is stable but ignores relationships within a 2D weight matrix. Some elements may already have converged while others are still catching up, slowing the matrix as a whole.
Muon treats the matrix as a single object and uses matrix-level updates to keep its elements on a more synchronized trajectory, with the goal of faster convergence. It cannot handle 1D parameters, so Input Embedding, Output Embedding, LayerNorm, Bias, and other such modules still use AdamW, leaving the optimizer in a hybrid state.
Keller Jordan’s initial version required a separate Learning Rate for each Module, making it difficult to use. Kimi’s Moonlight set the Muon-to-AdamW ratio at roughly 0.2, reducing the system to a single global Learning Rate and helping move the idea from theory into large-scale use.
V4 did not copy 0.2 and instead uses 0.18. 刘逸峰 is also watching the number of Newton–Schulz iterations: whether Keller Jordan’s 5 steps or V4’s 10 steps is better remains an open experimental question.
18. Muon Is an Infra Test, Not a One-Function Swap
赵晨阳 corrects the simplified claim that AdamW does not require distributed sharding: a 1T-scale model cannot fit on a single node, so both AdamW and Muon require complex parallelism. Muon’s matrix-level computation is simply harder to partition.
Muon removes the Second Moment, taking Optimizer State from roughly 2x to 1x and saving VRAM. The trade-off is that orthogonalization after Newton–Schulz requires the full 2D weight matrix; if parameters have already been fragmented by TP or FSDP, they must first be aggregated.
In 赵晨阳’s impression, Kimi K2 mainly partitions Muon at the Data Parallelism level rather than slicing it further through Tensor Parallelism. Scheduling, communication, ZeRO/FSDP alignment, Checkpoint saving, and Resume logic all become much more complex than in the AdamW era.
Pretraining and post-training generally need to keep the optimizer consistent. If the post-training system cannot finish adapting to Muon in time, pretraining may also continue “making do” with AdamW. Both therefore agree that full-stack Muon deployment is an effective signal of a model team’s Infra capability.
19. MHC Trades Wider Information Flow for Reasoning Ability
A traditional Transformer layer passes residual information with width d. Hyper-Connection adds a Channel dimension, expanding inter-layer information flow to c×d.
Naive Hyper-Connection can destabilize forward activations and backward gradients, which limited its early reception in the community. MHC adds Sinkhorn-Knopp constraints so that horizontal and vertical information-flow scales remain controlled, mitigating exploding and vanishing gradients.
刘逸峰 believes the direct gains land mainly in reasoning ability. 赵晨阳 points out that DeepSeek likely did not precisely compare specific capability gains, focusing instead on a selection of Benchmarks, so the contribution of MHC cannot be cleanly separated from the final model score.
20. Attention Residual May Have a Higher Ceiling; MHC Is Easier to Deploy
刘逸峰 compares MHC with Kimi’s recent Attention Residual: MHC increases the width of inter-layer information flow, while Attention Residual is more like DenseNet, allowing early layers to influence later layers across the network. Both redesign Layer-Wise information flow.
For teams with limited resources, MHC is relatively straightforward to implement. 刘逸峰’s own Deep Delta Learning also explores the Hyper-Connection direction, seeking performance with little additional compute; Attention Residual describes cross-layer relationships more precisely and may have a higher ceiling, but its Infra requirements are greater.
The inference side also pays the bill: MHC turns a simple residual addition into Sinkhorn normalization and Mixing, and legacy kernels are not efficient enough. 赵晨阳’s team wrote a TileLang Split-K Kernel to improve GPU utilization during Small Batch Decode.
21. TileLang Is Lowering the Cost of Custom Kernels for New Algorithms
赵晨阳 explains a Kernel as the concrete tiling and scheduling of the same matrix computation on different hardware. A 4096×4096 matrix can be processed in 128×128 or 256×256 tiles, with the optimal choice determined by memory bandwidth and hardware structure.
CUDA offers the highest performance and control but is the most expensive to develop and maintain. Triton materially lowers the barrier while sacrificing some expressive power and extreme performance. TileLang sits between them: it is lower-level and more expressive than Triton, while improving development speed over hand-written CUDA.
The V4 report emphasizes that TileLang reduces some Kernel launch overhead to the microsecond scale and improves bit-wise reproducibility: repeated Forward passes on the same Prompt are easier to reproduce, which is especially important for inference-engineering Debugging.
赵晨阳 sees TileLang’s long-term value in lowering the iteration cost of producing high-performance Kernels for new algorithms. It has become one of the default options for Frontier Labs worldwide and can be supported proactively by different hardware vendors; it is not inherently Nvidia-only.
22. FP4 Finally Moves from Hardware Slides into the Full Supermodel Pipeline
From BF16 to FP8 to FP4, bit width keeps halving, directly easing compute, memory capacity, and data-read bandwidth constraints. FP4’s numerical range is extremely narrow, making gradient overflow or collapse to zero easy during training; productionizing it at very large scale is far harder than releasing a quantized model.
V4 post-training uses QAT: the training side retains FP32 master weights, compresses them into the FP4 range before computation, then dequantizes them to FP8 for calculation so the model adapts to quantization error in advance. The sampling side uses actual FP4 weights, directly reducing memory traffic.
This directly benefits the Rollout stage of RL. The larger the model and the longer the output, the more likely sampling is to account for more than 70% of total time. The FP4 weights used for sampling during training also match the final release Checkpoint, avoiding the additional precision loss of training in FP8 and quantizing only before release.
赵晨阳’s team has completed full-process FP8 RL and INT4 QAT, but sampling is still mostly W4A16. DeepSeek went further to W4A8—4-bit weights and 8-bit activations. He therefore believes FP4 has “officially left hardware vendors’ PPT” and begun to become an industrial standard for open-source supermodels.
23. The Value of Open-Source Inference Is Supporting a Proliferation of Models in One Stack
Even if DeepSeek opens its own inference implementation, general-purpose open-source engines have a different job: one system must support DeepSeek, Xiaomi, MiniMax, Kimi, and other models at the same time. Integration, local deployment, and portability are the core value.
赵晨阳 believes the gap between open- and closed-source inference is no longer large, while training still lags clearly. In the past, a new model might launch in February while an open-source RL system did not run end to end until May or June; V4 completed the full inference and RL pipeline on launch day, materially narrowing the reproduction lag.
Open source also feeds back into closed-system performance and makes technical paths more transparent. 赵晨阳 sees the ecosystem’s advances in quantized training and expert aggregation as a substantive contribution by Chinese labs to global AI.
24. Post-Training Resolves Multi-Capability Conflicts by “Splitting First, Then Distilling”
V4 post-training first has multiple experts learn different skills separately, then distills them back into a unified model. Unlike R1’s distillation from a strong model into a small model, this approach first expands expert capacity, searches for local optima for different objectives, and then extracts the useful parts while controlling the final parameter count.
赵晨阳 frames it as a multi-objective optimization problem: push Coding too hard and math may deteriorate; repair math and instruction following may suffer. On a complex Loss Surface, directly searching for a joint Pareto optimum creates severe gradient conflict.
“Splitting first, then distilling”（先分裂、再蒸馏）rewrites the problem as first obtaining multiple converged discrete points, then having a student model fit multiple teacher distributions—roughly an interpolation among those points. Expert aggregation is not a new concept, but V4 offers a public industrial implementation.
25. Benchmarks Age, but Evaluation Determines Whether the Industry Is Fooling Itself
刘逸峰 quotes a well-known NLP researcher: “We cannot optimize what we cannot evaluate.” If a capability cannot be quantified, there is no way to tell whether a new Feature improves or damages the product.
赵晨阳 argues for fewer Benchmarks and more Evaluation. A single Benchmark can be saturated in 1–2 years, but the need to evaluate multi-step task stability, long-horizon dialogue, persona consistency, and tool-call robustness will not disappear.
After a Claude update, users often decide that the new version is worse based on a handful of conversations, creating Vibe Checking or Vibe Benchmarking. When public scores broadly exceed 90 while real-world experience still differs materially, the industry has entered a “Benchmark credibility crisis.”
刘逸峰 adds that capability domains matter more than individual test sets. Once Agentic capability or long-range attention is identified as a Domain, it may be saturated within 6–12 months. The absence of Degradation in old capabilities is good news, but context and evaluation complexity expand with it.
26. Agent Is the Common Direction; Commercialization Still Points to Real Workflows
From MiMo 2.5, GLM-5.1, MiniMax 2.7, and Kimi K2.6 to Opus 4.7 and GPT-5.5, 刘逸峰 sees a common trend: every major player is prioritizing Agent capability. Products such as OpenClaw and Claude Code show model companies that commercialization may come from workflows consuming tokens continuously, not just membership services.
He also believes open-source methods and architectures still converge in some respects, while optimizers are gradually moving from AdamW toward Muon and its variants. But V4’s abandonment of MLA shows that convergence does not mean the architecture has settled.
程曼祺 adds that the underlying model bound to a To B product can still change quickly. After GPT-5.5 launched, some corporate executives were already considering switching their model from Claude to GPT; the next phase remains to be watched.
27. China Bets on Architecture and Cost-Performance; the U.S. Bets on Capability Ceilings and Data Loops
赵晨阳 finds U.S. models particularly impressive on long-horizon Agent work, with OpenClaw and Claude Code standing out. Since Claude 4.5, multi-turn Agentic Coding has improved substantially, while RLHF/RLAIF and high-quality human feedback have formed a powerful data loop.
He summarizes the strengths of Chinese teams as high architecture-innovation density and strong engineering follow-through: V4 replaced and ran multiple components at once, including hybrid sparse attention, MHC, Muon, and FP4 training. The U.S. and Chinese paths and styles differ, but the industry as a whole is still advancing.
赵晨阳 also believes U.S. models generally focus more on opening new domains such as long context, multimodal fusion, and Agent capability. Chinese models focus more on cost-performance; at equivalent capability, their prices may be an order of magnitude lower than U.S. models.
On sparsification, he believes the U.S. has enough compute to push the performance ceiling first and address cost later; sparsification sacrifices some ceiling. 程曼祺 reflects that only China and the U.S. can afford this round of AI’s “space race.”
28. V4 Will Be Remembered for Its Combination Recipe, Not Any Single Structure Lasting Forever
刘逸峰 believes V4’s most durable algorithmic legacy is likely to be extreme token-wise compression. Earlier work mostly reduced and then expanded dimensions within an individual token; V4 pushes that compression into industrial-scale supermodels.
赵晨阳 places more weight on the combination of long context, an ultra-low active ratio, and low per-token cost. MHC or hybrid attention may not survive in its current form, just as MLA may have been optimal only for one stage; hardware will also drive more elegant replacements.
V4 has nevertheless validated this engineering recipe first and may become the default starting point for future open-source large models. DeepSeek’s influence is not limited to leaderboards; it keeps resetting the reference standard for how open-source models should be implemented.
Their next steps therefore diverge: 刘逸峰 wants to continue studying CSA, HSA, and other token-wise compression methods, Hyper-Connection, Attention Residual, and Muon parameters; 赵晨阳 plans to study how to identify the quality of AI-generated PRs and port MTP and other LM optimizations to open-source speech systems that still lag significantly.