June 21, 2026 · 6 min read · ← All posts

MiniMax M3 and WebBrain Cloud 1.0 enter the frozen WebBrain planner benchmark

We added two hosted OpenAI-compatible planner runs to WebBrain's frozen first-tool benchmark: minimax/minimax-m3 through OpenRouter and webbrain-cloud 1.0 through https://api.webbrain.one/v1. The surprise is that MiniMax M3 does not beat the older MiniMax M2.7 result in this harness. WebBrain Cloud lands close behind M3 on Sonnet alignment, costs much less for the replay, but was much slower on this endpoint path.

What we ran

Both runs used the same 100 single-turn browser-agent prompts and the same frozen baseline we use for comparable planner results:

node test/llm/run-llamacpp.mjs \
  --base https://openrouter.ai/api/v1 \
  --model minimax/minimax-m3 \
  --tag 2026-06-21-openrouter-minimax-m3 \
  --concurrency 2 \
  --timeout 180000 \
  --no-save-request \
  --freeze test/llm/freeze/baseline-2026-05-23.json

node test/llm/run-llamacpp.mjs \
  --base https://api.webbrain.one/v1 \
  --model "webbrain-cloud 1.0" \
  --tag 2026-06-21-webbrain-cloud-final \
  --concurrency 1 \
  --timeout 180000 \
  --no-save-request \
  --freeze test/llm/freeze/baseline-2026-05-23.json

The compatibility anchor is the frozen May 23, 2026 Sonnet 4.6 WebBrain prompt and 41-tool schema, system hash 5c4fac1387025050. These hosted endpoints both support native OpenAI structured tools, so the published runs used the old frozen toolset as native tools, not the text-call fallback used for local chat-template compatibility experiments.

Result files:

test/llm/results/2026-06-21-openrouter-minimax-m3_chrome_minimax_minimax-m3_frozen
test/llm/results/2026-06-21-webbrain-cloud-final_chrome_webbrain-cloud_1.0_frozen

Headline results

MetricMiniMax M3 via OpenRouterWebBrain Cloud 1.0
Completed cases100/100100/100
Transport errors00
Parsed tool calls85/10090/100
Strict exact first-call match17/10016/100
Tool-name match vs ideal32/10035/100
Sonnet match, all cases75.0%73.0%
Sonnet match, when Sonnet tooled73.9%72.8%
Median latency3.1s8.8s
p95 latency8.2s47.1s
Total wall time214s at concurrency 21,388s at concurrency 1
Reported run cost$1.06$0.12

The clean read: MiniMax M3 is a solid hosted planner, but it is not an upgrade over our saved MiniMax M2.7 result for this specific first-tool benchmark. WebBrain Cloud 1.0 is slightly weaker on Sonnet alignment, but it is cheaper in the reported usage fields and produced more parsed tool calls.

Latency needs one caveat. MiniMax M3 ran with concurrency 2, matching the older OpenRouter-style run. WebBrain Cloud had to run with concurrency 1; a concurrency-2 attempt produced transport failures from this environment. So compare the median per-case latency directly, but treat wall time as endpoint-path-specific.

MiniMax M3 vs MiniMax M2.7

This is the comparison I cared about most.

ModelParsed callsExactNameSonnet allSonnet tooledMedianCost
MiniMax M2.788/10023%36%77.0%76.1%3.1s$0.16
MiniMax M385/10017%32%75.0%73.9%3.1s$1.06

M3 did not regress catastrophically. It still tracks Sonnet better than most local runs. But the older M2.7 result remains stronger on every quality metric we publish here: more parsed tool calls, more exact first actions, higher tool-name agreement, and better Sonnet alignment. M3 also reported a much higher cost for this replay.

That does not mean M3 is the weaker model in general. This is a narrow first-tool browser-agent routing harness, frozen to an older WebBrain schema. It does mean we should not silently replace M2.7 with M3 in the planner table just because the model name is newer.

WebBrain Cloud 1.0

WebBrain Cloud 1.0 came in just below MiniMax M3 on Sonnet alignment:

ModelParsed callsExactNameSonnet allSonnet tooledMedianCost
MiniMax M385/10017%32%75.0%73.9%3.1s$1.06
WebBrain Cloud 1.090/10016%35%73.0%72.8%8.8s$0.12

The profile is different. WebBrain Cloud produced more tool calls and a slightly better ideal tool-name score than M3, but it aligned less often with Sonnet's first-tool choices and had much higher latency in this run. It also reported a much lower cost for the 100-case replay.

For product use, this is a reasonable default-cloud profile: cheap, OpenAI-compatible, native tool-capable, and close enough to the top hosted planner tier to be useful. For the benchmark leaderboard, it sits below MiniMax M3 and above the Qwen 3.5 4B row on the all-case Sonnet score.

Updated Sonnet 4.6 comparison

This table uses the saved Claude Sonnet 4.6 run as the first-tool reference. "Match all" counts all 100 prompts. "Match when Sonnet tooled" counts only the 92 prompts where Sonnet emitted a tool call. "Exact" and "Name" are the stricter replay against expected/NNN.json, not the Sonnet reference.

#ModelMatch allMatch when Sonnet tooledTool-call rateValid-name rateExactNameMedian
refClaude Sonnet 4.6100.0%100.0%92%92%19%41%2.8s
1Gemma 4 31B QAT w4a1677.0%78.3%95%95%19%37%0.55s
2Qwen 3.6 27B77.0%77.2%92%92%18%37%10.2s
3MiniMax M2.777.0%76.1%88%88%23%36%3.1s
4MiniMax M375.0%73.9%85%85%17%32%3.1s
5Intel Gemma 4 31B int4 AutoRound74.0%72.8%88%88%14%34%0.63s
6WebBrain Cloud 1.073.0%72.8%90%90%16%35%8.8s
7Qwen 3.5 4B73.0%71.7%82%82%12%33%5.5s
8Gemma 4 26B-A4B71.0%70.7%87%87%13%30%1.4s
9Qwen 3.6 35B-A3B70.0%70.7%90%90%18%38%10.3s
10Qwen 3.5 9B70.0%69.6%90%90%15%35%0.91s
11Gemma 4 E4B68.0%68.5%87%87%14%35%4.5s
12Nemotron Omni 30B67.0%68.5%93%93%16%36%2.6s
13Gemma 4 12B QAT w4a1667.0%67.4%92%92%14%33%0.43s
14DiffusionGemma 26B-A4B67.0%64.1%79%79%10%26%0.35s
15Gemma 4 E2B63.0%60.9%76%76%12%31%4s
16Gemma 4 12B Coder Fable5 Composer 2.561.0%62.0%94%94%9%26%1.9s
17Cohere North-Mini-Code 1.059.0%58.7%93%93%9%24%3.2s
18Browser-Use Qwen 30B-A3B Q443.0%45.7%93%88%12%35%0.48s
19LFM 2.540.0%38.0%83%83%4%23%6s
20Qwen 3.5 0.8B37.0%34.8%90%90%7%15%0.45s
21Qwen 3.5 2B36.0%34.8%89%89%4%7%0.78s
22VibeThinker 3B BF1633.0%32.6%84%83%2%19%5s
23Molmo2 8B8.0%0.0%2%1%0%0%1.7s

The top of the table is getting crowded, but the conclusion did not change. The best saved Sonnet-alignment result is still the Gemma 4 31B QAT / Qwen 3.6 27B / MiniMax M2.7 cluster. MiniMax M3 joins just below that cluster. Because M3 is roughly double the size of M2.7, that result is hard to justify for this job: larger, more expensive, and weaker on the frozen planner run is not a worthwhile trade.

WebBrain Cloud 1.0 sits one row lower, but it is the more useful product story here. It is a good entrant with a free tier, and it gives people a working WebBrain path when they do not have a local LLM ready. It should also keep improving as we tune the browser-side optimizations around the hosted route.

What changes

For hosted planner routing, I would keep MiniMax M2.7 in the table ahead of MiniMax M3 until M3 wins a rerun with a newer prompt or a different tool format. M3 is good, but the frozen WebBrain result does not justify a default-model change, especially when the newer model is about twice the size.

For WebBrain Cloud, the result is more product-facing than leaderboard-facing. It passed the old structured tool interface cleanly, completed all 100 cases without transport errors in the final run, and stayed close to the hosted MiniMax tier on Sonnet alignment. The latency needs endpoint work, but the quality/cost shape is usable, and the free tier makes it a practical default for anyone who wants to try WebBrain before setting up a local model.

The next fair test is the current production prompt and tool schema, not the May 23 frozen one. The frozen run answers "how does it compare with our historical rows?" The live-schema run answers "what should WebBrain route to today?"

Tags: #MiniMaxM3 #WebBrainCloud #OpenRouter #ToolCalling #BrowserAgent #LLMBenchmark

Written by Emre Sokullu. WebBrain is MIT-licensed and open on GitHub.