October 2026. The open-weights story I expected to write this quarter was the usual rotation: Kimi and Qwen trading the top of the board while everyone else fought for bronze. That story died on September 22, when an Economic Information Daily report on the Artificial Analysis open-source board put Xiaomi's MiMo-V2.6-Pro in first place at 46 points, priced a working task at $0.13, and called it the first model into the ten-cent band — ahead of Kimi K3 and Qwen3.8 Max. As of this article's data cutoff on October 11, 2026, the field around it is the shortlist ranked below: GLM-5.3-Flash, Kimi K3, Qwen3.8 Max, MiniMax M2.7, and Inkling-Small. Every number carries its source and its date, and the whole list comes back up for review on January 15, 2027.
The Board This Ranking Stands On
Two dated sources do most of the work. The September 22 Economic Information Daily report supplies the open-source order, MiMo's 46 points, and the $0.13-per-task figure; Sina Finance's and Beijing Daily's September 22-23 syndications confirm the same Pro-flagship numbers. The wider context comes from the same index family: DataLearner's September 30 mirror of the Intelligence Index v4.3 has Claude Opus 5.5 at the top of the overall board, and an October 9 snapshot relayed through llm-stats puts GPT-5.6 Sol first at 0.590 on a $5/$30 price card. Those are different boards at different calibers, and I keep them apart on purpose. When a ranking looks strange this quarter, the usual culprit is two snapshots blended into one sentence. How the boards are built, and where they disagree, is its own subject — the methodology breakdown handles that.
Data hygiene, stated once: every figure below traces to a dated source, and anything I couldn't trace stayed out. The rule costs the piece some color. It also keeps the claims from rotting, which is what model rankings start doing about three weeks after publication.
What's On the Record, Model by Model
| Model | On the record | Price card |
|---|---|---|
| MiMo-V2.6-Pro (Xiaomi) | 46 points, first on the AA open-source board; third among open models on Code Arena, level with the high-spec build of Claude Fable 5 (same report, dated 2026-09-22) | $0.13 per task; no token card in my sources |
| GLM-5.3-Flash (Zhipu) | On the September 22 board; 320B parameters | $0.15 in / $0.50 out per MTok |
| Kimi K3 | On the September 22 board, below MiMo-V2.6 | Not in my sources |
| Qwen3.8 Max (Alibaba's Qwen line) | On the September 22 board, below MiMo-V2.6 | Not in my sources |
| MiniMax M2.7 | In the same open-source field; 205K token context | $0.30 in / $1.20 out per MTok |
| Inkling-Small (Thinking Machines Lab) | In the same open-source field; 276B parameters | Not in my sources |
Read the missing cells as findings in themselves. Half this field ships fewer public numbers than the discourse around it implies, and an empty cell I couldn't fill tells you more than a guessed one would.
MiMo-V2.6: The Takeover, On the Record
The numbers first. Forty-six points, first on the open-source board, per the September 22 report. A working cost of $0.13 per task, flagged in the same report as the first entry into the ten-cent range. And the order change: MiMo ahead of Kimi K3 and Qwen3.8 Max — the sentence that redrew the map, since those two sat at the top of this field until that date.
The domestic angle deserves care. In July, the highest-ranked Chinese model on the Artificial Analysis Intelligence Index was Alibaba's Qwen3.7 Max at 56.6, global #5 on that month's snapshot, as its own deep dive records. The September report scores a different board and calls MiMo-V2.6 the top domestic model on it. Different snapshots, different rulers, so I won't subtract 46 from 56.6 and present the remainder as a gap. The defensible sentence is narrower: the top domestic slot changed hands this quarter, and the report that moved it is dated September 22.
The Pro variant matters too. MiMo-V2.6-Pro, per the same report, runs third among open models on Code Arena, level with the high-spec build of Claude Fable 5. Coding boards are where value claims get stress-tested, and an open model standing level with a frontier closed one will age better than a points lead on any aggregate index.
Now the catches, because there are several. The $0.13 figure is a per-task cost; my sources include no per-token price card, so token-billed budgets still need the vendor's own card. The 46 points come from one report on one date, and I have no MiMo position from the later September 30 mirror to confirm the standing held through the month. And the overall boards I can date have closed models at the top (Opus 5.5 on the September 30 mirror, GPT-5.6 Sol on the October 9 relay), with nothing in my sources putting an open model above them.
GLM-5.3-Flash: The Cheapest Card on the Table
Zhipu's GLM-5.3-Flash appears on the September 22 board with a 320B parameter count and the lowest price card on this list: $0.15 in and $0.50 out per million tokens. The same report's table supplies the comparison anchors — Gemini 3.7 Flash at $0.75/$3.75 and Grok 4.5 at $2/$6, both closed models. Whatever a benchmark says, unit economics an order of magnitude under Grok 4.5's card is the open-source pitch this quarter more than any single score.
The catch: the coverage that prices GLM-5.3-Flash doesn't hand me its individual board score. I can tell you what it costs and that it held a seat on the dated board; where it landed beyond that stays an open cell, and this article leaves it open on purpose.
Kimi K3 and Qwen3.8 Max: The Former Occupants
Both models appear in the September 22 report in one role: the incumbents MiMo-V2.6 passed. That is close to everything my dated sources say about them individually, which is worth dwelling on, because it is the actual state of the record at my cutoff — no score split for either model, no price card for either, in anything that reached me.
Context I can add is dated too. The Qwen line's July flagship, Qwen3.7 Max, held the domestic top at 56.6 on the main index and Code Arena's fourth slot at 1541 Elo, the first Chinese model on that table, all pinned to July. Whether Qwen3.8 Max answers the September result is a question for a future snapshot; I have no dated answer.
Is Kimi K3 finished, then? No. Losing first place on one September board is one result in one column, a thin file to close a lab's case with. What the result does end is the pitch that still leads with "Kimi and Qwen own the open-source top." That leaderboard expired on September 22.
MiniMax M2.7: Context on a Budget
MiniMax M2.7 shows up in the same report with a 205K token context window and a $0.30 in / $1.20 out price card. The window reads differently in late 2026 than it would have a year ago — the July flagships shipped million-token contexts, which puts 205K at mid-pack for the era. On paper the trade is simple: a longer window than GLM-5.3-Flash, at roughly double the input price. Where that trade pays, in my read of the spec sheet, is work that holds large documents in one session without needing frontier-grade reasoning over them: volume contract review, transcript summarization, codebase-context chat. My sources give me the numbers but no task-level scores for M2.7, so consider that paragraph judgment, labeled as such.
Inkling-Small: The Unknown Quantity
Thinking Machines Lab's Inkling-Small appears in the same report at 276B parameters, and that, plus its seat in the open-source field, is close to the full extent of what my dated sources say. Everything else a buyer would want — card, score, benchmark names — stayed absent through my October 11 cutoff. A 276B open model from a lab people watch this closely is worth tracking precisely because so little verified detail is public; the missing numbers say something about maturity, and nothing proven about weakness. My working position: watch it, and don't build a launch plan on it until the January re-check puts numbers next to the name.
Where This Ranking Is Soft
Three limits, stated plainly. First, the open-source order rests on one report from one date; a single-source ranking is a snapshot with an expiry date, and this one was already three weeks old at my cutoff. Second, the boards don't share a scale — 46 open-source points, 56.6 index points from July, and 0.590 on the October relay are three different rulers, and mixing them produces confident nonsense. Third, half the field ships no public price card in my sources, which makes "cheapest" a per-token claim about GLM-5.3-Flash and a per-task claim about MiMo-V2.6, and those are never the same claim.
Anyone selling you "open-source beat closed models this quarter," or its mirror, "closed models buried the open field," is selling a headline. The dated record says something narrower: the open field took the value frontier at $0.13 a task behind a 46-point board lead, and the overall-board tops I can date stayed closed.
A Note From the Platform Side
One disclosure, since this is where my employer enters: I track these boards for x-rush.ai. The platform's design is the reason none of the above turns into a product claim. There is no model picker and no provider vocabulary in the product; each task is routed to a model that fits it, and which lab served a given run stays an implementation detail the interface doesn't promise. Text work is a flat 0.5 credits a unit, and chat's assistance pool takes 0.5 credits per ten successful replies — the invoice is silent on the engine. If MiMo-V2.6 or GLM-5.3-Flash ever joins that rotation, you'll meet it as output quality and price, never as a model name you had to memorize.
Verdict: A Value Crown, Dated
If you want the top of the overall boards, this list is not where you shop; every dated top I have belongs to a closed model, and pretending otherwise would be the hype this piece keeps refusing. If your bill matters more than your benchmark tab, the open-source field just had its strongest quarter that my sources can date: MiMo-V2.6 for per-task economics, GLM-5.3-Flash for the cheapest token card, MiniMax M2.7 when window size beats raw score, Inkling-Small on the watchlist, and Kimi K3 and Qwen3.8 Max worth re-checking in January before anyone writes their obituaries.
Check the date on any ranking before you quote it. Or borrow this one, and re-check it with me on January 15, 2027.
AI Chat runs that trade daily: you describe the job, and no engine name enters the conversation.