linked note ·
Kimi K3 Brings the Open-Model Gap Down to Weeks
Kimi K3's benchmarks put it within touching distance of the strongest proprietary models, with its full weights due by July 27.
source: Kimi K3: Open Frontier Intelligence · Moonshot AI
#ai #models #open-source #competition
In yesterday’s note on Kimi K3, I wrote that we still needed benchmarks and real use to know whether its capability justified the frontier-tier price.
Moonshot has now published the benchmarks. It also confirmed that K3’s full weights will be released by July 27, 2026.
At 2.8 trillion parameters, K3 is set to become the first open model in the 3T class. The more consequential claim, however, is frontier intelligence.
In Moonshot’s own benchmarks, K3 still trails Claude Fable 5 and GPT-5.6 Sol overall, but the gap is narrow, and K3 leads on a few individual tests.
These are not entirely clean comparisons. The evaluations mix harnesses and include internal benchmarks, so independent testing and real use still matter. But this no longer looks like an open model catching up to an old frontier. It looks like an open model arriving at the current one.
K3 matters for agent infrastructure too. Its dynamic tool loading brings the progressive-disclosure pattern from Your Agent Has Too Many Tools to an open model, another proprietary advantage that is beginning to close.
The timing makes that clearer. Claude Fable 5 first shipped on June 9, was suspended three days later under US export controls, and returned globally on July 1. GPT-5.6 launched on July 9. K3’s official launch arrived eight days later, with its weights due ten days after that.
Depending on where you start the clock, the gap between the latest proprietary frontier and an open release is now measured in weeks, not months.
K3 is not the strongest model, but it may be the first open model that can reasonably be called frontier.
Side note: I liked how Moonshot’s chart shows the size of open-weight models changing over time, so I recreated it below in Vega-Lite. The original chart is from Moonshot’s Kimi K3 post.
{
"$schema": "https://vega.github.io/schema/vega-lite/v5.json",
"description": "Open frontier model size over time, reconstructed from the inline SVG chart in Moonshot AI's Kimi K3 launch post.",
"usermeta": {
"source": "https://www.kimi.com/blog/kimi-k3",
"source_chart_title": "Open frontier model size over time",
"chart_as_of": "2026-07-16",
"units": "billions of total parameters",
"extraction_note": "Release dates and parameter counts were recovered from the inline SVG's plotted coordinates and labels. The series dataset adds synthetic as-of rows solely to extend each step line to the chart date; release_events contains only model releases."
},
"title": {
"text": "Open frontier model size over time",
"subtitle": [
"Total parameters of each company's flagship open model, Jul 2025 to Jul 2026",
"Only open models and flagship models are included"
],
"anchor": "start",
"fontSize": 22,
"subtitleFontSize": 13,
"subtitleColor": "#6e6e6e",
"offset": 18
},
"width": 1000,
"height": 520,
"background": "#ffffff",
"padding": {
"left": 12,
"right": 60,
"top": 12,
"bottom": 12
},
"datasets": {
"release_events": [
{
"company": "DeepSeek",
"model": "DeepSeek V2",
"date": "2024-04-22",
"parameters_billion": 236
},
{
"company": "DeepSeek",
"model": "DeepSeek V3",
"date": "2024-12-25",
"parameters_billion": 671
},
{
"company": "Alibaba",
"model": "Qwen3 235B",
"date": "2025-04-28",
"parameters_billion": 235
},
{
"company": "Moonshot AI",
"model": "Kimi K2",
"date": "2025-07-11",
"parameters_billion": 1040
},
{
"company": "Z.AI",
"model": "GLM 4.5",
"date": "2025-07-28",
"parameters_billion": 355
},
{
"company": "MiniMax",
"model": "MiniMax M2",
"date": "2025-10-27",
"parameters_billion": 230
},
{
"company": "Xiaomi",
"model": "MiMo V2 Flash",
"date": "2025-12-16",
"parameters_billion": 309
},
{
"company": "Z.AI",
"model": "GLM 5",
"date": "2026-02-11",
"parameters_billion": 744
},
{
"company": "Alibaba",
"model": "Qwen 3.5",
"date": "2026-02-16",
"parameters_billion": 397
},
{
"company": "DeepSeek",
"model": "DeepSeek V4 Pro",
"date": "2026-04-22",
"parameters_billion": 1600
},
{
"company": "Xiaomi",
"model": "MiMo V2.5 Pro",
"date": "2026-04-27",
"parameters_billion": 1020
},
{
"company": "MiniMax",
"model": "MiniMax M3",
"date": "2026-06-01",
"parameters_billion": 428
},
{
"company": "Thinking Machines",
"model": "Inkling",
"date": "2026-07-15",
"parameters_billion": 975
},
{
"company": "Moonshot AI",
"model": "Kimi K3",
"date": "2026-07-16",
"parameters_billion": 2800
}
],
"series": [
{
"company": "DeepSeek",
"model": "DeepSeek V2",
"date": "2024-04-22",
"parameters_billion": 236,
"record_type": "release"
},
{
"company": "DeepSeek",
"model": "DeepSeek V3",
"date": "2024-12-25",
"parameters_billion": 671,
"record_type": "release"
},
{
"company": "DeepSeek",
"model": "DeepSeek V4 Pro",
"date": "2026-04-22",
"parameters_billion": 1600,
"record_type": "release"
},
{
"company": "DeepSeek",
"model": "DeepSeek V4 Pro",
"date": "2026-07-16",
"parameters_billion": 1600,
"record_type": "as_of",
"latest_label": "DeepSeek · 1.6T"
},
{
"company": "Alibaba",
"model": "Qwen3 235B",
"date": "2025-04-28",
"parameters_billion": 235,
"record_type": "release"
},
{
"company": "Alibaba",
"model": "Qwen 3.5",
"date": "2026-02-16",
"parameters_billion": 397,
"record_type": "release"
},
{
"company": "Alibaba",
"model": "Qwen 3.5",
"date": "2026-07-16",
"parameters_billion": 397,
"record_type": "as_of",
"latest_label": "Alibaba · 397B"
},
{
"company": "Moonshot AI",
"model": "Kimi K2",
"date": "2025-07-11",
"parameters_billion": 1040,
"record_type": "release"
},
{
"company": "Moonshot AI",
"model": "Kimi K3",
"date": "2026-07-16",
"parameters_billion": 2800,
"record_type": "release",
"latest_label": "Moonshot AI · 2.8T"
},
{
"company": "Z.AI",
"model": "GLM 4.5",
"date": "2025-07-28",
"parameters_billion": 355,
"record_type": "release"
},
{
"company": "Z.AI",
"model": "GLM 5",
"date": "2026-02-11",
"parameters_billion": 744,
"record_type": "release"
},
{
"company": "Z.AI",
"model": "GLM 5",
"date": "2026-07-16",
"parameters_billion": 744,
"record_type": "as_of",
"latest_label": "Z.AI · 744B"
},
{
"company": "MiniMax",
"model": "MiniMax M2",
"date": "2025-10-27",
"parameters_billion": 230,
"record_type": "release"
},
{
"company": "MiniMax",
"model": "MiniMax M3",
"date": "2026-06-01",
"parameters_billion": 428,
"record_type": "release"
},
{
"company": "MiniMax",
"model": "MiniMax M3",
"date": "2026-07-16",
"parameters_billion": 428,
"record_type": "as_of",
"latest_label": "MiniMax · 428B"
},
{
"company": "Xiaomi",
"model": "MiMo V2 Flash",
"date": "2025-12-16",
"parameters_billion": 309,
"record_type": "release"
},
{
"company": "Xiaomi",
"model": "MiMo V2.5 Pro",
"date": "2026-04-27",
"parameters_billion": 1020,
"record_type": "release"
},
{
"company": "Xiaomi",
"model": "MiMo V2.5 Pro",
"date": "2026-07-16",
"parameters_billion": 1020,
"record_type": "as_of",
"latest_label": "Xiaomi · 1.02T"
},
{
"company": "Thinking Machines",
"model": "Inkling",
"date": "2026-07-15",
"parameters_billion": 975,
"record_type": "release"
},
{
"company": "Thinking Machines",
"model": "Inkling",
"date": "2026-07-16",
"parameters_billion": 975,
"record_type": "as_of",
"latest_label": "Thinking Machines · 975B"
}
],
"as_of_marker": [
{
"date": "2026-07-16",
"label": "July 16, 2026"
}
]
},
"layer": [
{
"data": {
"name": "series"
},
"mark": {
"type": "line",
"clip": true,
"interpolate": "step-after",
"strokeWidth": 3
},
"encoding": {
"x": {
"field": "date",
"type": "temporal",
"scale": {
"domain": [
"2025-07-01",
"2026-07-21"
]
},
"axis": {
"title": null,
"format": "%b %Y",
"tickCount": "quarter",
"labelAngle": 0,
"grid": true
}
},
"y": {
"field": "parameters_billion",
"type": "quantitative",
"scale": {
"domain": [
0,
3000
],
"nice": false
},
"axis": {
"title": "Total parameters",
"values": [
0,
500,
1000,
1500,
2000,
2500,
3000
],
"labelExpr": "datum.value >= 1000 ? format(datum.value / 1000, '.1~f') + 'T' : datum.value + 'B'"
}
},
"color": {
"field": "company",
"type": "nominal",
"legend": null,
"scale": {
"domain": [
"Z.AI",
"Alibaba",
"MiniMax",
"DeepSeek",
"Moonshot AI",
"Thinking Machines",
"Xiaomi"
],
"range": [
"#454a66",
"#a8621f",
"#824b50",
"#345077",
"#b92622",
"#6e6e73",
"#a89f80"
]
}
},
"order": {
"field": "date",
"type": "temporal"
},
"tooltip": [
{
"field": "company",
"type": "nominal",
"title": "Company"
},
{
"field": "model",
"type": "nominal",
"title": "Model"
},
{
"field": "date",
"type": "temporal",
"title": "Date",
"format": "%Y-%m-%d"
},
{
"field": "parameters_billion",
"type": "quantitative",
"title": "Parameters (B)"
}
]
}
},
{
"data": {
"name": "release_events"
},
"mark": {
"type": "point",
"clip": true,
"filled": true,
"size": 85,
"stroke": "#ffffff",
"strokeWidth": 1.5
},
"encoding": {
"x": {
"field": "date",
"type": "temporal",
"scale": {
"domain": [
"2025-07-01",
"2026-07-21"
]
}
},
"y": {
"field": "parameters_billion",
"type": "quantitative",
"scale": {
"domain": [
0,
3000
],
"nice": false
}
},
"color": {
"field": "company",
"type": "nominal",
"legend": null,
"scale": {
"domain": [
"Z.AI",
"Alibaba",
"MiniMax",
"DeepSeek",
"Moonshot AI",
"Thinking Machines",
"Xiaomi"
],
"range": [
"#454a66",
"#a8621f",
"#824b50",
"#345077",
"#b92622",
"#6e6e73",
"#a89f80"
]
}
},
"tooltip": [
{
"field": "company",
"type": "nominal",
"title": "Company"
},
{
"field": "model",
"type": "nominal",
"title": "Model"
},
{
"field": "date",
"type": "temporal",
"title": "Release date",
"format": "%Y-%m-%d"
},
{
"field": "parameters_billion",
"type": "quantitative",
"title": "Parameters (B)"
}
]
}
},
{
"data": {
"name": "as_of_marker"
},
"mark": {
"type": "rule",
"color": "#999999",
"strokeDash": [
6,
6
],
"strokeWidth": 1.5
},
"encoding": {
"x": {
"field": "date",
"type": "temporal",
"scale": {
"domain": [
"2025-07-01",
"2026-07-21"
]
}
}
}
},
{
"data": {
"name": "as_of_marker"
},
"mark": {
"type": "text",
"align": "center",
"baseline": "bottom",
"dy": -5,
"color": "#6e6e6e",
"fontSize": 12
},
"encoding": {
"x": {
"field": "date",
"type": "temporal",
"scale": {
"domain": [
"2025-07-01",
"2026-07-21"
]
}
},
"y": {
"datum": 3000,
"type": "quantitative"
},
"text": {
"field": "label"
}
}
},
{
"data": {
"name": "series"
},
"transform": [
{
"filter": "isValid(datum.latest_label)"
}
],
"mark": {
"type": "text",
"align": "left",
"baseline": "middle",
"dx": 9,
"fontSize": 12,
"fontWeight": 600
},
"encoding": {
"x": {
"field": "date",
"type": "temporal",
"scale": {
"domain": [
"2025-07-01",
"2026-07-21"
]
}
},
"y": {
"field": "parameters_billion",
"type": "quantitative",
"scale": {
"domain": [
0,
3000
],
"nice": false
}
},
"text": {
"field": "latest_label"
},
"color": {
"field": "company",
"type": "nominal",
"legend": null,
"scale": {
"domain": [
"Z.AI",
"Alibaba",
"MiniMax",
"DeepSeek",
"Moonshot AI",
"Thinking Machines",
"Xiaomi"
],
"range": [
"#454a66",
"#a8621f",
"#824b50",
"#345077",
"#b92622",
"#6e6e73",
"#a89f80"
]
}
}
}
}
],
"config": {
"font": "Inter, ui-sans-serif, system-ui, sans-serif",
"view": {
"stroke": null
},
"axis": {
"domainColor": "#d0d0d0",
"gridColor": "#eeeeee",
"labelColor": "#666666",
"titleColor": "#555555",
"tickColor": "#d0d0d0"
}
}
}