MiniMax M3 Tested on Fireworks: 20/20 for $0.0148, Against Four Other Open-Weight Models
By Promptster Team · 2026-11-11
Is MiniMax M3 worth a place in your model shortlist? This is the first time we have tested MiniMax on this blog. We ran it on Fireworks against the four other open-weight models we call there, using the same four graded tasks.
MiniMax M3 passed 20/20 for $0.0148, with a median call of 2.4 s. That puts it in the cheap, fast group: well below GLM 5.3 ($0.0568) and Kimi K3 ($0.1601), and close to DeepSeek V4.1 Flash ($0.0119), which has the same $0.30/$1.20 list price. Every Fireworks model passed every call, so the differences below are cost, latency and token usage, not accuracy.
How we tested
Four script-graded tasks, with no LLM judge:
- json-order: strict JSON from an order email with a decoy PO reference, a cancelled line and a discount that applies only to some items.
- stock-floor: a ten-step inventory calculation that rounds down at every step (answer: 587).
- seven-word-taglines: three lowercase taglines of exactly seven words, with banned words and one required word.
- launch-notes: choose the final approved date and budget from superseded notes, replying with only
Nov 21, $37500.
We ran each task five times per model (n=5) with temperature 0.7 and max_tokens 4000, on 2026-10-23, through Promptster's own provider handlers. MiniMax M3 is accounts/fireworks/models/minimax-m3. Costs are cost_usd at list rates from shared/pricing.ts. Latency is wall-clock time from one machine, and five runs per task is a small sample.
We have restricted ourselves to what the battery measured. For background on the model, Fireworks' MiniMax M3 page, read on 2026-10-23, lists about 428B total parameters with about 23B activated, and a cached-input price of $0.06 per million tokens. The same page gives two different context lengths and does not state a licence, so we don't quote either.
The results
| Model (Fireworks) | List price (in/out per 1M) | Passed | Cost, 20 calls | Median latency, 20 calls | Input / output tokens, 20 calls |
|---|---|---|---|---|---|
| gpt-oss-120b | $0.15 / $0.60 | 20/20 | $0.0076 | 1.6 s | 3,655 / 11,714 |
| DeepSeek V4.1 Flash | $0.30 / $1.20 | 20/20 | $0.0119 | 3.2 s | 2,870 / 9,186 |
| MiniMax M3 | $0.30 / $1.20 | 20/20 | $0.0148 | 2.4 s | 4,755 / 11,162 |
| GLM 5.3 | $1.40 / $4.40 | 20/20 | $0.0568 | 6.7 s | 2,480 / 12,131 |
| Kimi K3 | $3.00 / $15.00 | 20/20 | $0.1601 | 10.3 s | 3,990 / 9,878 |
Median latency per task (json-order / stock-floor / taglines / launch-notes):
| Model | Median latency per task |
|---|---|
| gpt-oss-120b | 1.4 / 3.8 / 1.5 / 1.5 s |
| MiniMax M3 | 1.8 / 4.1 / 4.0 / 1.3 s |
| DeepSeek V4.1 Flash | 2.8 / 4.2 / 3.1 / 2.3 s |
| GLM 5.3 | 4.0 / 14.2 / 8.5 / 3.6 s |
| Kimi K3 | 8.1 / 24.5 / 11.1 / 4.8 s |
MiniMax had the fastest median of the five on launch-notes and the second fastest overall. Its slowest call took 5.6 s. Kimi K3's stock-floor median alone was 24.5 s.
MiniMax passed every diagnostic as well as every graded check. It returned no code fences, ended stock-floor with a clean Answer: 587 line, and replied to launch-notes with exactly Nov 21, $37500 on all five runs. There were no host errors, timeouts or rate limits on any of the 20 calls.
Same list price as DeepSeek V4.1 Flash, 25% more cost
MiniMax M3 and DeepSeek V4.1 Flash both list at $0.30 in and $1.20 out on Fireworks. MiniMax still cost $0.0148 to DeepSeek's $0.0119 for the same 20 calls, because it used more tokens on both sides of the call.
More prompt tokens for the same prompts. MiniMax counted 263, 224, 193 and 271 prompt tokens on the four tasks. DeepSeek V4.1 Flash counted 171, 129, 98 and 176. That is 92 to 95 extra prompt tokens per call, which points to a longer chat template on MiniMax's side rather than anything in our text.
More reasoning on the tagline task. Across five runs, MiniMax wrote 3,594 output tokens on seven-word-taglines, more than any other Fireworks model. DeepSeek V4.1 Flash wrote 2,518. One MiniMax run used 1,000 output tokens, 971 of them reasoning, to produce three seven-word lines:
refill your flask for endless adventures daily
refill and reduce waste with every sip
refill your flask keep drinks hot cold
More output on the arithmetic task too. MiniMax wrote 4,934 output tokens across the five stock-floor runs, against DeepSeek V4.1 Flash's 3,344. On the other two tasks it wrote less: 1,545 against 1,805 on json-order, and 1,089 against 1,519 on launch-notes.
So the extra cost comes from two places. One is a fixed prompt overhead on every call, which matters most when prompts and answers are short. The other is longer reasoning on the open-ended and multi-step tasks.
The cache price we did not apply
Fireworks reported cached prompt tokens on 17 of MiniMax's 20 calls, 3,910 of the 4,755 input tokens. shared/pricing.ts has no cached rate for MiniMax on Fireworks, so cost_usd prices them at the full $0.30. Fireworks' model page lists $0.06 for cached input. Input is a small part of this bill, though, so the cached rate would lower MiniMax's total only slightly. The same applies to the other Fireworks rows, all of which reported some cached tokens.
Where it sits
Within the cheap group, gpt-oss-120b was cheaper than MiniMax, at half the list price, and faster overall. DeepSeek V4.1 Flash was a little cheaper and a little slower. MiniMax M3 sits between them: under a cent and a half for 20 graded calls, and a median under 2.5 s.
Against the expensive group it is a clear step down in cost. GLM 5.3 cost 3.8× as much and Kimi K3 almost 11× as much for the same 20/20, and both were slower on every task. On this battery, choosing Kimi K3 over MiniMax M3 got the same score for more money and more waiting.
The limits of the data are the usual ones. These are four single-turn tasks that finish in seconds, and all five models passed them. They don't separate the models on long context, multi-step agent work or tool use, where the larger models are aimed. If those are your workloads, test them directly.
For how these models compare with closed ones, see our October open-weight vs closed comparison and the original open-source vs closed-source benchmark. For the same weights on different hosts, see the Kimi K3 host test.
What we'd do
Put MiniMax M3 on the shortlist for cheap, short tasks. It went 20/20, obeyed every format rule and was fast. That is the same result as the cheapest models here, at a cost in the same range.
Compare it with DeepSeek V4.1 Flash on your own prompts, not on the rate card. Their list prices are identical, but token counts are not. On short prompts, MiniMax's roughly 95-token template overhead matters.
Use GLM 5.3 or Kimi K3 only if a task needs them. On this battery neither one did better than MiniMax M3, and both cost several times more.
Run this comparison yourself — no account, no API key
- Run the seven-word tagline task: this is where MiniMax M3 spent the most tokens. Check the output token count as well as the word count.
To run MiniMax M3 next to other open-weight and closed models on your own prompts, with token counts and cost, use Promptster.
Run 2026-10-23 through Promptster's provider handlers, all on Fireworks: accounts/fireworks/models/minimax-m3, glm-5p3, kimi-k3, deepseek-v4p1-flash and gpt-oss-120b. Four script-graded tasks, five runs each, temperature 0.7, max_tokens 4000. Costs are cost_usd from list rates in shared/pricing.ts, with no cache discount. Latency is wall-clock from one machine. Model facts beyond the battery are from the Fireworks model page as read on 2026-10-23. Raw responses and usage objects for every call are in content/blog/data/q4b-2026-battery/.