Late October has been crowded for open-weight flagship models: DeepSeek V4.1 Flash 1017 open-sourced its weights on October 17, MiniMax released M3.1 Flash on October 19 and published weights on October 23, and before them came Zhipu GLM-5.3 in August and Moonshot's Kimi K3 in July. All four pitch "agentic capability" as their headline strength, and all four are locked in a price war. Put them at the same table — who is actually strongest, most efficient, and cheapest? This article tries to assemble an honest table from public data, and to spell out the table's limitations before reading any number off it.
Scope first. Every benchmark figure below comes from vendor model cards, official changelogs, or official blogs, as of 2026-10-24. The only third-party scale covering several of these models is the Artificial Analysis Intelligence Index (AA for short). Prices are snapshots of official pricing pages. Where sources conflict, both numbers are listed side by side without adjudication. Nothing here is purchasing or investment advice.
1. The Four Contenders and Why They Qualify
The roster matters, because the roster itself is part of the calibration.
MiniMax M3.1 Flash (428B total, 23B active, released 10-19, weights opened 10-23) is the protagonist, sold as the budget newcomer that cuts "agentic usage cost by 70 percent versus its predecessor." DeepSeek V4.1 Flash 1017 (763B total weights, 8B to 16B active depending on phase, open-sourced 10-17) is the top rival in MiniMax's own comparison table. Zhipu GLM-5.3 (744B total, 40B active, released 8-14, weights opened 8-28) fills the GLM column of that table, and Kimi K3 (2.8T total parameters, released in July, weights opened late July) fills the Kimi column.
These four sit at one table only because all four appear in MiniMax's official model card for M3.1 Flash. But "alignable" does not mean "neutral" — that caveat is the whole point of the next section.
Who is left out: Qwen, because MiniMax's comparison table has no Qwen column and no alignable official agentic benchmark for Qwen's latest flagship could be found. The Kimi trap deserves special care: Moonshot shipped K3.1 Pro in the early hours of October 24, but it does not belong in the comparison table — its numbers are all self-run on Moonshot's own harness, its weights are not downloadable yet (promised before November 4). K3.1 Pro appears only as a footnote in Section 5; it must not be conflated with K3.
2. Capability Table: Drawn by a Single Runner
Table first, caveats immediately after.
| Benchmark | M3.1 Flash | DeepSeek V4.1 Flash 1017 | GLM-5.3 | Kimi K3 |
|---|---|---|---|---|
| Terminal Bench 4.0 | 38.1 | 42.3 | 37.9 | 12.6 |
| DeepSWE v1.1 | 73.7% | 77.0% | 66.9% | 67.5% |
| Automation-Bench v1.0.6 | 62.8 | 60.5 | blank | 46.7 |
| Agents' Last Exam | 29.6 | 35.4 | blank | blank |
| CyberGym | 82.2 | 88.1 (DeepSeek self-reported) | 84.5 (Zhipu self-reported) | blank |
Three footnotes are mandatory.
First, the whole table is run by MiniMax. The model card states plainly: "All benchmarks on the card are MiniMax's own," and the harnesses differ per benchmark — BankerToolBench used OpenCode 1.17.18 with MiniMax's own reviewer, Terminal Bench 4.0 used Codex 0.149.1, DeepSWE used mini-SWE-agent 2.4.5, CyberGym used Claude Code 2.1.263. This is a single-runner table, and the runner is the protagonist. The three rival columns contain MiniMax's runs, not the vendors' official numbers, so a protagonist-favoring result should inform, not settle.
Second, mixed sources are flagged. The DeepSeek cell for Agents' Last Exam (35.4) and for CyberGym (88.1) comes from DeepSeek's own changelog, because MiniMax's card leaves those cells blank; the GLM-5.3 CyberGym figure (84.5) is Zhipu self-reported. Blank cells mean "no official result," not zero — the card says so explicitly.
Third, every M3.1 Flash score is vendor-reported with no third-party replication as of this writing. There is even an internal contradiction: MiniMax's blog gives 62.2 for Automation-Bench while the model card table says 62.8; both are listed here without adjudication. Parameter counts also fail to reconcile — the card says 428B while HF safetensors metadata reads 220B, a gap one outlet explicitly notes was "not reconciled."
Read with those caveats: DeepSeek 1017 leads the coding-flavored benchmarks under MiniMax's runs, K3 trails badly on Terminal Bench, and M3.1 Flash leads on Automation-Bench — remembering that judge and beneficiary are the same party.
3. Efficiency and Price: The Real Gaps Are Not in the Scores
Capability scores are contestable; the efficiency numbers are where the four models diverge most sharply.
KV cache per token is the key engineering figure, because it dictates the memory bill for self-hosted long-context workloads. DeepSeek 1017's official figure is 890 bytes per token. M3.1 Flash's counterpart is 34,560 bytes per token — but from vLLM's engineering blog, at NVFP4 precision. The two figures come from different sources — the former official, the latter from an inference-framework measurement — yet even so, a gap of roughly 39x is the single largest efficiency difference this comparison found. Back-of-envelope: holding a 1M-token context, DeepSeek 1017 needs about 0.9GB of KV cache, M3.1 Flash about 34GB. The right conclusion depends on whether your workload is long-context and self-hosted.
Beyond that, calibration thins out. M3.1 Flash officially claims 200+ tokens per second output (P50 of 170+); official output speeds for DeepSeek 1017, GLM-5.3, and K3 are unpublished, so throughput stays out of the table. The 200 tok/s figure floating around GLM-5.3 belongs to the FlashX variant and must not be mixed in. Checkpoint sizes: M3.1 Flash is 250.1GB at NVFP4 per vLLM (media also cite about 233GB and about 220GB at different precisions), DeepSeek 1017 ships about 510GB across 48 shards, GLM-5.3's FP8 weights run about 756GB, and public data for K3 is thin. Note that the 220GB checkpoint figure and the disputed 220B parameter count are unrelated items.
Price is the most tangible dimension. Per million tokens, official pages:
| Model | Input | Output | Cache |
|---|---|---|---|
| M3.1 Flash | $0.15 | $0.60 | $0.015 |
| DeepSeek V4.1 Flash 1017 | $0.15 off-peak (2x at peak) | $0.60 | $0.003 (cache hit) |
| GLM-5.3 | CNY 8 | CNY 28 | CNY 2 |
| Kimi K3 | $3 (media-cited) | $15 (media-cited) | not published |
Details worth unpacking. DeepSeek 1017's $0.003 cache-hit rate is the cheapest of the four and a genuine advantage for prefix-heavy agentic conversations, but weekday peak windows (UTC 01:00-04:00 and 06:00-10:00) double both input and output rates. GLM-5.3 bills in CNY with a 28-yuan output tier, among the most expensive output prices here. M3.1 Flash is flat-priced with no tiers or time windows — the simplest structure. K3's international pricing is media-cited. One official cost anecdote: running all 152 tasks of Agents' Last Exam on M3.1 Flash is estimated at about CNY 739 total (official figure; no comparable numbers for the other three).
A third-party cost data point: AA puts DeepSWE per-task cost at about $0.22 for M3.1 Flash and about $0.26 for DeepSeek 1017. That scale covers only part of the field, so GLM-5.3 and K3 have no comparable figures and it stays out of the table. As for the loudest slogan — "70 percent lower agentic usage cost" — it exists only as a MiniMax claim with no published methodology or predecessor definition; it is relayed here, not endorsed.
4. M3 to M3.1 Flash: A Generational Leap Without a Benchmark Table
How much better is M3.1 Flash than its predecessor? Here is an easily missed fact: MiniMax has published no side-by-side benchmark table for M3 versus M3.1 Flash. Official material comes in three kinds.
First, three claims: online inference nearly 3x faster than the predecessor; 200+ tok/s output (P50 170+); and agentic usage cost down 70 percent versus the predecessor. Any citation must carry the third-party qualifier that none of these comes with a published methodology, that "predecessor" means M3, and that the cost claim is unilateral.
Second, price comparison. M3 listed at $0.60 input / $2.40 output, commonly sold at half price ($0.30 / $1.20 for inputs up to 512K, $0.06 cache, doubling beyond 512K); M3.1 Flash is a flat $0.15 / $0.60 with $0.015 cache — half of M3's discounted rates, one quarter for cache.
Third, engineering data from vLLM's blog: weight files shrink from 451.5GB (M3) to 250.1GB (NVFP4, about 44.6 percent smaller); KV cache per token drops from 61,440 to 34,560 bytes (about 43.75 percent less); resident concurrency at 256K context on four B300 GPUs rises from 28 to 76 users (a capacity comparison, not a throughput measurement). Architecturally M3.1 Flash keeps M3's MoE design, changing three things: full-layer sparse attention (MSA) replaces the prior design, quantization-aware training (QAT) is added, and DSpark speculative decoding replaces the MTP scheme.
On the third-party scale, AA's Intelligence Index moves from 29 (M3) to 42 (M3.1 Flash), with per-task cost falling from about $0.51 to about $0.22. One warning: M3's 66.0 on Terminal-Bench 2.1 and M3.1 Flash's 38.1 on Terminal Bench 4.0 are different benchmark versions — do not compare them and declare a regression.
5. The Only Third-Party Scale, and the Contradiction Ledger
Outside vendor walls, options are thin. The only third-party scale covering three of these models is the AA Intelligence Index: DeepSeek 1017 scores 45, M3.1 Flash 42, M3 29. GLM-5.3 posts 59 on a different metric (Agentic Index, August, different scale and month, treat as directional) and K3 has no comparable score. A timing note: when weights landed on October 23, AA's page still flagged M3.1 Flash as proprietary with is-open-weights set to false — the 42 is a third-party read, and AA has not updated the label.
Footnote on freshness: K3.1 Pro, shipped in the early hours of October 24, self-reports 65.7 percent on Terminal-Bench 4.0 — versus the 12.6 that MiniMax's table lists for K3. The two numbers differ in version and harness and cannot be compared; this article draws no conclusion from either. When K3.1 Pro weights arrive before November 4 and third-party scales catch up, this table will finally get sturdier. The same caution applies to MiniMax's 70 percent agentic-cost claim, which ships without a published measurement method.
The contradiction ledger, all listed without adjudication: Automation-Bench 62.2 (blog) versus 62.8 (model card); M3.1 Flash at 428B (card) versus 220B (HF metadata); DeepSeek's CyberGym cell blank in MiniMax's table versus 88.1 self-reported by DeepSeek; M3.1 Flash checkpoint sizes of 250.1GB, about 233GB, and about 220GB (different precision bases). None of these rows change the logic of this review: every contested cell stays unverifiable until a third party runs it.
FAQ
Q1: Are all four models actually open source? All four are open weights, but the licenses differ in strictness: DeepSeek 1017 ships under MIT with zero strings; GLM-5.3 is free for ordinary developers and only constrains very large MaaS providers; M3.1 Flash's MiniMax Community License requires commercial attribution and written approval above USD 20 million annual revenue; K3 carries a similar revenue threshold. All four are downloadable right now — the M3.1 Flash weights live on ModelScope (repo address: modelscope.cn/models/MiniMax/MiniMax-M3.1-Flash), the other three on their official org pages. For clause-by-clause reading, see the site's license comparison.
Q2: Why does the table use Kimi K3 instead of the newer K3.1 Pro? Because MiniMax's official comparison table lists K3. K3.1 Pro arrived in the early hours of October 24 with only self-run harness numbers, and its weights are not downloadable until before November 4, so it stays a footnote.
Q3: Can M3.1 Flash's benchmark scores be trusted? They are all vendor-reported with no third-party replication, and the capability table is run by MiniMax alone. The only cross-vendor third-party ruler today is the AA Intelligence Index (42 for M3.1 Flash versus 45 for DeepSeek 1017), so use it for cross-checking.
Q4: What does the 39x KV cache gap mean? DeepSeek 1017 officially uses 890 bytes per token; M3.1 Flash uses 34,560 bytes (vLLM engineering figure, different source). For self-hosted long-context serving, the latter needs roughly 39 times the KV memory — a bill to compute before committing.
Q5: Which one should a budget-conscious agent team pick? Match the workload: heavy multi-turn cache hits favor DeepSeek 1017's $0.003 cache rate and off-peak discounts; flat, time-insensitive pricing favors M3.1 Flash; CNY settlement with output-heavy loads means GLM-5.3's 28-yuan output tier needs scrutiny. All based on official price snapshots — check live pages before deciding.
Related reading
- MiniMax M3.1 Flash Unmasked: Space Bunny Ran 200T Tokens: the unmasking story and timeline; this article expands its two-row table into a full comparison
- Open Weight, But How Open? Five Model Licenses Compared: clause-by-clause license reading; every licensing line here points there
- DeepSeek Weights Are Free, Yet Agent Task Bills Fall 60 Percent: the DeepSeek 1017 checkpoint deep-dive; this article borrows only the cells it needs