S
pending
Open Weights Isn't the Same Word as Self-Hostable
Grounded / Real
Inflated / Uruttu
Original Content
𝗞𝗶𝗺𝗶 𝗞𝟯 𝗶𝘀 𝗮 𝟮.𝟴 𝘁𝗿𝗶𝗹𝗹𝗶𝗼𝗻 𝗽𝗮𝗿𝗮𝗺𝗲𝘁𝗲𝗿 𝗼𝗽𝗲𝗻-𝘄𝗲𝗶𝗴𝗵𝘁 𝗺𝗼𝗱𝗲𝗹. 𝗧𝗵𝗲 𝗹𝗮𝗿𝗴𝗲𝘀𝘁 𝗲𝘃𝗲𝗿 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝗱.
Moonshot AI dropped the weights last night, a day ahead of schedule.
MoE, 1M-token context, native vision. Available on API since July 16 — now downloadable.
━━━━━━━━━━━━━━
Where it actually lands:
→ #3 on Artificial Analysis Intelligence Index — behind only Claude Fable 5 and GPT-5.6 Sol Max, and cheaper than both
→ #2 on the Vals AI index
→ #1 in Frontend Code Arena
→ Beats Claude Opus 4.8 and GPT-5.5 on coding and agentic evals
Not the best model in the world. Comfortably the best one you can download.
━━━━━━━━━━━━━━
The architecture is the story, not the parameter count:
→ Kimi Delta Attention — hybrid linear attention
→ Attention Residuals
→ Stable LatentMoE — 16 of 896 experts active per token, about 1.8% of the pool
→ MXFP4/MXFP8 training
→ Roughly 2.5x the scaling efficiency of K2, per Moonshot
2.8T total. A fraction of it firing on any given token.
━━━━━━━━━━━━━━
And they opened the infrastructure too:
Attention kernels, the MoE communication library, and tooling for running agent environments at scale.
That's the part that compounds. Weights get superseded in six months. Kernels and comm libraries get built on.
━━━━━━━━━━━━━━
Now the part nobody posting about this will tell you.
You cannot run this.
2.8T parameters at 4-bit is roughly 1.4TB of weights before you allocate a single token of KV cache. An 8×H100 node gives you 640GB. You need multi-node, high-speed interconnect, and disaggregated serving to make it work at all — Moonshot uses their own Mooncake stack for exactly this.
"Open weights" and "self-hostable" are not the same sentence. For most enterprises this is an API model that happens to be downloadable.
If you want genuinely deployable, the 30B-class open models are still where on-prem lives.
━━━━━━━━━━━━━━
Two things to check before you build on it:
→ Confirm the LICENSE file in the repo yourself. The previous Kimi release was Modified MIT. Terms do not automatically carry forward.
→ There is an active US policy dispute around Moonshot, including public discussion of export controls. If you're deploying for regulated or government clients, factor that into procurement now, not later.
━━━━━━━━━━━━━━
The gap between the best closed model and the best open one is now measured in weeks, not generations.
That changes the build-vs-buy maths for everyone.
We help teams work out what's actually deployable inside their own infrastructure — and what only looks like it is.
DM me if you're weighing open weights against an API for a regulated deployment.
♻️ Repost if this is useful to your network
🔔 Follow me for open model and on-prem deployment breakdowns
Moonshot AI dropped the weights last night, a day ahead of schedule.
MoE, 1M-token context, native vision. Available on API since July 16 — now downloadable.
━━━━━━━━━━━━━━
Where it actually lands:
→ #3 on Artificial Analysis Intelligence Index — behind only Claude Fable 5 and GPT-5.6 Sol Max, and cheaper than both
→ #2 on the Vals AI index
→ #1 in Frontend Code Arena
→ Beats Claude Opus 4.8 and GPT-5.5 on coding and agentic evals
Not the best model in the world. Comfortably the best one you can download.
━━━━━━━━━━━━━━
The architecture is the story, not the parameter count:
→ Kimi Delta Attention — hybrid linear attention
→ Attention Residuals
→ Stable LatentMoE — 16 of 896 experts active per token, about 1.8% of the pool
→ MXFP4/MXFP8 training
→ Roughly 2.5x the scaling efficiency of K2, per Moonshot
2.8T total. A fraction of it firing on any given token.
━━━━━━━━━━━━━━
And they opened the infrastructure too:
Attention kernels, the MoE communication library, and tooling for running agent environments at scale.
That's the part that compounds. Weights get superseded in six months. Kernels and comm libraries get built on.
━━━━━━━━━━━━━━
Now the part nobody posting about this will tell you.
You cannot run this.
2.8T parameters at 4-bit is roughly 1.4TB of weights before you allocate a single token of KV cache. An 8×H100 node gives you 640GB. You need multi-node, high-speed interconnect, and disaggregated serving to make it work at all — Moonshot uses their own Mooncake stack for exactly this.
"Open weights" and "self-hostable" are not the same sentence. For most enterprises this is an API model that happens to be downloadable.
If you want genuinely deployable, the 30B-class open models are still where on-prem lives.
━━━━━━━━━━━━━━
Two things to check before you build on it:
→ Confirm the LICENSE file in the repo yourself. The previous Kimi release was Modified MIT. Terms do not automatically carry forward.
→ There is an active US policy dispute around Moonshot, including public discussion of export controls. If you're deploying for regulated or government clients, factor that into procurement now, not later.
━━━━━━━━━━━━━━
The gap between the best closed model and the best open one is now measured in weeks, not generations.
That changes the build-vs-buy maths for everyone.
We help teams work out what's actually deployable inside their own infrastructure — and what only looks like it is.
DM me if you're weighing open weights against an API for a regulated deployment.
♻️ Repost if this is useful to your network
🔔 Follow me for open model and on-prem deployment breakdowns
Validated Content
Good news — this checks out far more than I expected. Kimi K3 is a real, very recent release and the post's technical details line up closely with reporting.
Confirmed accurate
- 2.8T parameters, MoE, open-weight, largest ever released — accurate. With 2.8 trillion parameters, Kimi K3 holds the distinction of being the largest open-weight model publicly available, according to Bloomberg.
- "Dropped the weights last night, a day ahead of schedule" — accurate. Moonshot AI released free, public download weights for Kimi K3 on July 26, 2026, at roughly 7:30 PM EDT — a day ahead of the July 27 target the company had previously communicated.
- API since July 16, now downloadable — matches: Kimi K3 is a 2.8-trillion-parameter open-weight model from Moonshot AI, released on July 16, 2026, with full weights following on July 26/27.
- 1M-token context, native vision — confirmed. The model features a 1-million-token context window, native visual understanding capabilities, and an always-on reasoning mode that the company calls "thinking mode."
- Beats Opus 4.8 and GPT-5.5 on coding/agentic evals, behind Fable 5 and GPT-5.6 Sol — accurate. Moonshot said K3 still sits behind Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol on overall performance, but it outperformed every other model in the company's evaluation suite, including Claude Opus 4.8 and GPT 5.5, across coding and agentic benchmarks.
- #1 Frontend Code Arena — confirmed. Arena ranked K3 first in its Frontend Code evaluation at 1,679 points, ahead of Fable 5, in blind developer testing.
- Architecture: Kimi Delta Attention, Attention Residuals, Stable LatentMoE, 16/896 experts (~1.8%) — confirmed. The model activates just 16 of its 896 experts per token, roughly 1.8% of the pool, and the model is built on two key architectural innovations developed internally at Moonshot AI: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals.
- ~1.4TB at 4-bit — exact match. Stored using MXFP4 quantization, the download occupies roughly 1.4 terabytes.
- 8×H100 = 640GB, multi-node needed — arithmetic checks out (8×80GB), and Moonshot's own guidance backs the "you need real infrastructure" point: Moonshot recommends serving K3 on supernodes of 64 or more accelerators, keeping expert-parallel traffic inside one high-bandwidth domain.
Real but understated in the post
- The export-control dispute is much bigger than "active policy dispute." This isn't a background footnote — it's a live international incident. White House OSTP Director Michael Kratsios alleged Moonshot "likely procured servers equipped with Nvidia GB300s and further accessed GB300s installed in Thailand to train its own AI models," and separately Kratsios claimed the US government had information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model, with Treasury Secretary Bessent warning that sanctions and Entity List designations will be on the table. Moonshot has denied the distillation allegations. This is worth knowing before advising clients — it's not a mild compliance footnote, it's an active US-China flashpoint with a named company facing potential Entity List action.
One thing to flag/verify yourself
- "#2 on Vals AI index" and "roughly 2.5x the scaling efficiency of K2" — I couldn't independently verify these two specific figures in what I found. They're plausible (consistent with the general performance story) but I'd confirm them directly against Vals AI's leaderboard and Moonshot's technical report before publishing, since I don't have a source pinning those exact numbers.
- License claim ("previous release was Modified MIT") — directionally sensible advice (always check the license yourself), but I didn't find a source confirming the specific prior license terms in this search. Worth a direct check of the K2 and K3 repos.
Bottom line: this is an unusually well-researched post — the technical specs, benchmark claims, and hardware math all check out against current reporting, and the export-control angle is actually understated rather than exaggerated. The two unverified figures (Vals AI rank, 2.5x efficiency claim) and the license claim are the only things I'd double check with primary sources before publishing.