By Evan Vega
Thinking Machines Lab released Inkling 2 Mini on 20 November 2026, a free, open-weights AI model the lab says is small enough to
run on a laptop with enough memory, about 20 GB. Its trick is a loop: the middle of the model runs three times for every word-piece it writes. The lab says that
lets it “reach the right answer with fewer tokens”, meaning it writes less on the way to an answer. The independent count
went the other way. Artificial Analysis, which runs every model through the same set of tasks, found Inkling 2 Mini wrote
about 38% more than Qwen3.8 27B, the model it ties on score. So it writes more, and each word is cheap: the same count found
it used about 2.8 times less computing, finished each task about four times faster and cost about a sixth as much.
Novel Cognition read the lab’s launch post, model card and benchmark table, checked them against the independent
numbers, and ran both models on one MacBook Pro. The full analysis is at
inkling2mini.novcog.us.com, and the release itself, the licence and
what it needs to run are on what shipped.
Thinking Machines Lab says its new open-weights looped model, Inkling 2 Mini, reaches the right answer with fewer tokens. Artificial Analysis counted about 38% more output tokens than Qwen3.8 27B at t
Within a day the launch drew 1,396 points and 283 comments on Hacker News and 1,567 upvotes on r/LocalLLaMA. A commenter who ran Plotly’s 43-question
data-analytics exam on a 64 GB Mac Mini reported a B, the same grade as DeepSeek V4.1 Flash and Sonnet 5, with Qwen 3.8 27B
on a B+; one question ran for more than three hours and 180 tool calls before he cut it off, the runaway loop the model card
warns about. Another found it used up a 19,200-token thinking limit on a WordPress plugin task without finishing, and with
search turned on it still reached for a hook name that does not exist. Early local runs had plainer problems. Several
people found it ran slower than the dense Qwen3.8 27B on the same graphics card with day-one builds, and Unsloth-style
quantized files gave wrong output on mainline llama.cpp until a fix merged on 21 November; the official ggml-org files were
not affected (reactions).
Artificial Analysis puts Inkling 2 Mini at 20 on its Intelligence Index, level with Qwen3.8 27B on its highest
(“xhigh”) setting. To get there Inkling wrote about 152,000 output tokens per index task against about 110,000 for Qwen.
The firm’s compute measure multiplies the parameters a model uses per token by the tokens it writes, and on that measure
Inkling used about 2.8 times less. Per task it cost $0.59 and took 9.0 minutes; Qwen cost $3.53 and took 38.2 minutes. Those
costs use Thinking Machines’ own API prices, $0.50 per million output tokens. The lab’s “fewer tokens” line does have
support of a narrower kind in its own research: a variable-loop research model reasoned in less than half the tokens at five
passes than at one on GPQA Diamond questions (1,661 against 3,703 on average). That compares the model with itself, not with
its rivals. Detail: fewer tokens.
Under the hood it is a mixture-of-experts model with 26.4 billion parameters, of which about 3 billion are used per
token: each layer holds 128 experts and picks 8. Who counts sets the figure. Artificial Analysis counts 3.1 billion unique
weights per token, the lab 3.4 billion, or 5.4 billion once the three passes are counted. The 24 stored layers sit in four
blocks of six; the two middle blocks run three times, so each token goes through 48 layer passes. In the lab’s own
comparison, a version without the loop cost 7.5% less to train but needed 140% more inference compute to match it. The
weights were trained and shipped in 4-bit NVFP4, with the last four layers in 8-bit MXFP8, for a footprint of about 20 GB,
under the Apache 2.0 licence (the loop).
The lab’s benchmark table, which it reports itself and notes uses evaluation setups that vary by model, has Inkling
ahead of every open model under 100 billion parameters on DeepSWE v1.1 (48.2%) and the Vals Index v2.2 (46.5%), and ahead
of Qwen3.8 27B on SWE-bench Pro v2 Hard (77.9% against 73.8%). Qwen leads on Terminal Bench v2.1 (73.0% against 71.6%). On Vals’
own overall leaderboard, which includes closed frontier models, that 46.5% ranks 38th of 43. The lab names its own weak spots: multi-turn chat and instructions with many constraints (50.0% on MultiChallenge, 33.4% on
ComplexConstraints), and facts, because looping “doesn’t enable its weights to store more facts”. Its non-hallucination rate
on AA-Omniscience is 63.4% to Qwen’s 71.4%. The model card warns it can get stuck in long reasoning loops when debugging and
may delete or overwrite files without asking, and recommends a sandbox plus limits on output tokens and agent turns
(the benchmarks).
Novel Cognition ran Inkling 2 Mini against Qwen3.8 27B on a MacBook Pro with an M4 Pro chip and 48 GB of memory, the
kind of machine the lab’s laptop claim is about. Both ran on the same build of llama.cpp, built from the main code after both of its
Inkling fixes had merged (20 and 21 November), in 4-bit versions of about 17 GB each, with each publisher’s recommended settings. The
test was twelve machine-scored tasks: six reasoning puzzles with one numeric answer worked out by script, among them a
multi-step table lookup in the style of the lab’s own N-Hop test, and six coding tasks checked against hidden tests. Each
model ran each task once, with a fixed seed, with answers capped at 12,000 tokens. For every answer the script counted the tokens written and
timed it, which puts the lab’s “fewer tokens” claim and Artificial Analysis’s speed figures to a quick test on one laptop.
Both models got all twelve right. Inkling 2 Mini wrote 29% more tokens than Qwen3.8 27B (17,356 against 13,437), and most of
the gap came from one coding task, Roman numerals, where it wrote 6,053 tokens to Qwen’s 1,852 while re-checking its own work;
leave that task out and the other eleven came to 11,303 against 11,585. It also writes about twice as fast on this laptop (25.7
against 12.7 tokens a second), so it finished the set in 11.3 minutes against 18.0: 1.6 times faster, not the 4 times
Artificial Analysis measured on the lab’s hosted service.
It is a spot check, not a benchmark. The results, per-task numbers and the full method are on
our test.
Novel Cognition’s reading: the loop is real engineering, and the price and speed advantages hold up in the independent
count. The “fewer tokens” line is the lab’s, and against its closest rival the count went the other way. Anyone choosing a
model to pay for or wait on should judge it by the cost and time of a finished task rather than the words it writes. For
agent work, follow the model card: set limits on tokens and turns, run it in a sandbox, and give it search when facts matter.
The analysis is also an episode of the Frontier Watch
podcast, and every source is listed on sources.
Novel Cognition tests vendor claims like these for companies as part-time AI leadership
(novcog.com/contact). The checks behind every published number are at
hiddenstatedrift.com/verify. Novel Cognition is not affiliated with
Thinking Machines Lab, Artificial Analysis, the Qwen team, Hugging Face or Unsloth.
Part of the Frontier Watch Series: Read the previous investigation
More Coverage:
→ Read this investigation on North Denver Tribune
→ Coverage from Daily Colorado News