过去半年,大模型的竞争一直在往“更大、更贵、更强”走。
Fable 5 这类前沿模型当然强,但问题也很现实:价格高、调用成本高、在团队或高频 agent 场景里,长期用下来是一笔不小的开支。
OpenRouter 最近做了一个很有意思的实验:不再只问“哪个单模型最强”,而是让多个模型组成一个 Panel(专家小组),并行回答,再由一个 judge / synthesizer 把答案融合成最终结果。
结果很直接:
单个模型不是终点。
多个模型合体,很多时候比单模型更强。
OpenRouter 的结论:Fusion 能逼近甚至超过前沿模型
OpenRouter 在 DRACO 深度研究基准上测试了 Fusion(融合模型)。
DRACO 是 Perplexity AI 提出的深度研究评测集,包含 100 个复杂任务,覆盖学术、金融、法律、医疗、技术、UX、产品比较等 10 个领域。它不是简单考记忆,而是考模型能不能搜索、理解、综合、引用,并产出真正有用的研究答案。
OpenRouter 的关键发现有三点:
- 多模型 Panel 稳定强于单模型。
- 前沿模型组队后,可以超过单个前沿模型。
- 平价模型组队后,也能接近 Fable 5 这类顶级模型。
原文数据里,Claude Fable 5 单跑的标准化得分是 65.3%。
而 Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro 的 Fusion 组合,得分达到 64.7%。
差距只有 0.6 个百分点。
更重要的是,OpenRouter 还指出,这组预算模型 Panel 的单任务成本大约只有 Fable 5 的一半。
这件事给我们的启发很大:
Fable 5 不是一个只能“买更贵模型”才能获得的能力。
它也可以被拆成一套本地可控的协作流程。
开源的agentfw:把 Fusion 放到本地
OpenRouter 的 Fusion 是云端服务。你调用 openrouter/fusion,它在服务端把请求分发给多个模型,再融合答案。
但对 agent 场景来说,我们更希望这个能力发生在本地。
原因很简单:
模型可以换。
供应商可以换。
路由策略可以改。
安全策略必须握在自己手里。
这正是 agentfw 适合做的事。
开源地址:https://github.com/openguardrails/agentfw
agentfw 是一个本地 AI agent firewall,站在 agent 和模型服务之间。它能看到请求、路由模型、保护敏感信息,并在本地对 agent 的模型调用做编排。
所以,Fusion 不一定非要长在云端。
它可以长在 agentfw 这一层。
本地 Fable 5 平替:DeepSeek V4 Pro + Kimi K2.6 + Qwen3.6
agentfw 推荐的本地 Fusion 阵容是:
Qwen3.6 + Kimi K2.6 + DeepSeek V4 Pro
这是一套偏工程实用的组合。
DeepSeek V4 Pro:负责推理、问题拆解、低成本高强度思考。
Kimi K2.6:负责代码、长上下文、中文理解、材料归纳和复杂文档消化。
Qwen3.6:负责稳定输出、结构化和多模态任务。
这三个模型不是简单轮流调用,而是分工协作:
同一个任务,agentfw 在本地并行发给三个模型。
三个模型各自给出答案、推理路径或候选方案。
本地 judge 对答案进行交叉检查:找共识、找冲突、找遗漏。
最后 synthesizer 生成一个统一答案。
这就是本地版 Model Fusion。
用户体感上,还是只调用一个“模型”。
但背后其实是三个开源模型在协作。
为什么 Fusion 会变强?
因为复杂任务往往不是单一能力问题。
一个模型可能事实覆盖更好,但表达一般。
一个模型可能推理更强,但容易漏掉边界条件。
一个模型可能中文更自然,但技术细节不够硬。
单模型回答时,这些弱点会直接暴露。
Fusion 的价值在于,它不是投票,而是综合。
它会把多个模型的答案拆开看:
哪些观点一致?
哪些地方互相矛盾?
谁补充了别人没看到的信息?
哪个答案看似自信但证据不足?
最终答案应该保留什么、删掉什么、修正什么?
这一步本质上是在给模型输出做“二次审稿”。
OpenRouter 原文里还有一个很有意思的现象:同一个模型和自己做 Fusion,也能涨分。Opus 4.8 与自己组成两模型 Panel 后,得分从单体 58.8% 提升到 65.5%。
这说明增益不只来自模型多样性,也来自“多次生成 + 综合裁决”这个流程本身。
agentfw 怎么做这件事?
在传统调用方式里,agent 只知道一个模型。
例如:
Claude Code 调 Fable 5。
Codex 调 GPT。
OpenClaw 调某个默认模型。
agentfw 介入后,模型调用会先经过本地防火墙层。
在这里,我们可以把“一个模型”替换成“一个 Fusion 策略”。
比如定义一个本地虚拟模型:
agent 仍然以为自己在调用一个模型。
但 agentfw 会在本地完成:
请求复制。
并行分发。
结果收集。
冲突分析。
最终融合。
敏感信息保护。
调用轨迹记录。
这就把“模型能力”从单点变成了组合能力。
人人都有 Fable 5,不是人人都买 Fable 5
这里的重点不是说某个组合在所有场景都 100% 等于 Fable 5。
真正重要的是:
很多实际工作,不需要每次都调用最贵的单模型。
调研、代码 review、方案设计、长文归纳、产品分析、投研摘要、法律条款初筛,这些任务更需要的是多视角、可复核、低幻觉和稳定产出。
Fusion 正好适合这些场景。
用 agentfw 做本地 Fusion 后,团队可以把高价模型调用从“默认选项”变成“兜底选项”。
日常任务走 Qwen3.6 + Kimi K2.6 + DeepSeek V4 Pro。
高风险任务再升级到 Fable 5。
模型故障时自动切换。
敏感信息在本地先被 agentfw 处理。
所有调用路径都有迹可查。
这比单纯追逐最强模型,更像一套真正可落地的 AI 基础设施。
最后
OpenRouter 的 Fusion 实验说明了一件事:
未来的大模型能力,不一定只来自“更大的单模型”。
它也可能来自“更好的模型协作系统”。
Fable 5 很强。
但 Fable 5 不是唯一答案。
当 Qwen3.6、Kimi K2.6、DeepSeek V4 Pro 这样的开源模型被 agentfw 在本地组织起来,每个人都可以拥有一套自己的“Fable 5 平替”。
不是云端给你的黑盒能力。
而是跑在你本地、可路由、可审计、可保护、可替换的 AI agent firewall。
这才是象信AI想做的事情:
让强模型能力变得更开放。
让 agent 的控制权回到用户手里。
让人人都有自己的 Fable 5。
开源地址:https://github.com/openguardrails/agentfw
参考来源
OpenRouter Blog: Surpassing Frontier Performance with Fusionhttps://openrouter.ai/blog/announcements/fusion-beats-frontier/
DRACO paperhttps://arxiv.org/abs/2602.11685
Over the past six months, the competition among large models has been moving toward "bigger, more expensive, and stronger."
Frontier models like Fable 5 are certainly powerful, but the practical problem is real: high prices and high call costs make them a significant expense over the long run in team or high-frequency agent scenarios.
OpenRouter recently ran a very interesting experiment: instead of only asking "which single model is strongest," it had multiple models form a Panel (expert group) that answered in parallel, with a judge/synthesizer fusing the answers into a final result.
The result was straightforward:
A single model is not the endgame.
Multiple models combined are often stronger than a single model.
OpenRouter's conclusion: Fusion can approach or even surpass frontier models.
OpenRouter tested Fusion (fused models) on the DRACO deep research benchmark.
DRACO is a deep research evaluation set introduced by Perplexity AI, containing 100 complex tasks across 10 domains including academia, finance, law, healthcare, technology, UX, and product comparison. It does not simply test memorization; it tests whether a model can search, understand, synthesize, cite, and produce genuinely useful research answers.
OpenRouter's key findings are threefold:
A multi-model Panel is consistently stronger than a single model.
Frontier models working as a team can surpass a single frontier model.
Budget models working as a team can also approach top-tier models like Fable 5.
In the original data, Claude Fable 5 running solo scored 65.3% on the normalized score.
Meanwhile, the Fusion combination of Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro scored 64.7%.
The gap is only 0.6 percentage points.
More importantly, OpenRouter also noted that this budget-model Panel costs roughly half as much per task as Fable 5.
This is very inspiring for us:
Fable 5-level capability is not something you can only obtain by "buying a more expensive model."
It can also be broken down into a locally controllable collaboration pipeline.
Open-source agentfw: bringing Fusion to your local environment
OpenRouter's Fusion is a cloud service. You call openrouter/fusion, and it distributes the request to multiple models on the server side, then fuses the answers.
But for agent scenarios, we would prefer this capability to live locally.
The reason is simple:
Models can be swapped.
Vendors can be swapped.
Routing strategies can be changed.
Security policies must remain in your own hands.
This is exactly what agentfw is suited for.
Open-source URL: https://github.com/openguardrails/agentfw
agentfw is a local AI agent firewall that sits between agents and model services. It can see requests, route models, protect sensitive information, and orchestrate agent model calls locally.
So Fusion does not have to live in the cloud.
It can live at the agentfw layer.
Local Fable 5 alternative: DeepSeek V4 Pro + Kimi K2.6 + Qwen3.6
agentfw's recommended local Fusion lineup is:
Qwen3.6 + Kimi K2.6 + DeepSeek V4 Pro
This is a practical, engineering-oriented combination.
DeepSeek V4 Pro: handles reasoning, problem decomposition, and low-cost high-intensity thinking.
Kimi K2.6: handles code, long context, Chinese understanding, material summarization, and complex document digestion.
Qwen3.6: handles stable output, structuring, and multimodal tasks.
These three models are not simply called in rotation; they collaborate with a division of labor:
For the same task, agentfw sends it to all three models in parallel locally.
Each of the three models produces its own answer, reasoning path, or candidate solution.
A local judge cross-checks the answers: finding consensus, conflicts, and omissions.
Finally, a synthesizer generates a unified answer.
This is local Model Fusion.
From the user's perspective, it still looks like calling a single "model."
But behind the scenes, three open-source models are collaborating.
Why does Fusion get stronger?
Because complex tasks are often not a matter of a single capability.
One model may have better factual coverage but mediocre expression.
One model may reason more strongly but easily miss edge cases.
One model may produce more natural Chinese but lack solid technical details.
When a single model answers, these weaknesses are directly exposed.
The value of Fusion is that it is not voting; it is synthesis.
It breaks apart the answers from multiple models and examines them:
Which viewpoints are consistent?
Where do they contradict each other?
Who contributed information the others missed?
Which answer looks confident but lacks evidence?
What should the final answer keep, remove, and correct?
Essentially, this step is a "second round of editorial review" for model outputs.
There is another very interesting phenomenon in the OpenRouter post: a model fused with itself can also gain points. After Opus 4.8 formed a two-model Panel with itself, its score rose from 58.8% as a single model to 65.5%.
This shows the gain comes not only from model diversity, but also from the process itself: "multiple generation + adjudicated synthesis."
How does agentfw do this?
In the traditional calling pattern, an agent only knows one model.
For example:
Claude Code calls Fable 5.
Codex calls GPT.
OpenClaw calls some default model.
Once agentfw is involved, model calls first pass through a local firewall layer.
Here, we can replace "a single model" with "a Fusion strategy."
For example, define a local virtual model:
The agent still believes it is calling a single model.
But agentfw completes the following locally:
Request duplication.
Parallel distribution.
Result collection.
Conflict analysis.
Final fusion.
Sensitive information protection.
Call trace logging.
This turns "model capability" from a single point into a combinatorial capability.
Everyone can have a Fable 5; not everyone has to buy Fable 5
The point here is not that some combination is 100% equal to Fable 5 in every scenario.
What truly matters is:
Many real-world tasks do not require calling the most expensive single model every time.
Research, code review, solution design, long-text summarization, product analysis, investment research summaries, and initial legal clause screening—these tasks need multi-perspective, verifiable, low-hallucination, stable output.
Fusion is exactly suited to these scenarios.
After using agentfw for local Fusion, teams can turn high-priced model calls from the "default option" into a "fallback option."
Daily tasks run on Qwen3.6 + Kimi K2.6 + DeepSeek V4 Pro.
High-risk tasks are upgraded to Fable 5.
Automatic failover occurs when a model malfunctions.
Sensitive information is first handled locally by agentfw.
Every call path is traceable.
This is more like truly deployable AI infrastructure than simply chasing the strongest model.
Finally
OpenRouter's Fusion experiment demonstrates one thing:
Future large-model capability may not come only from "bigger single models."
It may also come from "better model collaboration systems."
Fable 5 is very strong.
But Fable 5 is not the only answer.
When open-source models like Qwen3.6, Kimi K2.6, and DeepSeek V4 Pro are organized locally by agentfw, everyone can have their own "Fable 5 alternative."
Not a black-box capability handed to you by the cloud.
Rather, it is an AI agent firewall running on your local machine that is routable, auditable, protectable, and replaceable.
This is exactly what Xiangxin AI wants to do:
Make powerful model capabilities more open.
Return control of agents to users.
Let everyone have their own Fable 5.
Open-source URL: https://github.com/openguardrails/agentfw
References
OpenRouter Blog: Surpassing Frontier Performance with Fusionhttps://openrouter.ai/blog/announcements/fusion-beats-frontier/
DRACO paperhttps://arxiv.org/abs/2602.11685




