594B-parameter Inkling 2 Small launches closed, with open weights promised next week

Thinking Machines Lab on Nov. 5 released Inkling 2 Small, a 594-billion-parameter model that is initially available through its own Tinker platform, with closed weights [1]. The company says it intends to publish the full weights under the permissive Apache 2.0 license "next week after the completion of external safety testing and red-teaming" [1].

The model is a Mixture-of-Experts design, meaning only part of the network runs for each token. Just 17 billion of its 594 billion parameters are active at once [1]. It reads text and images, lets users set how much effort it spends, and does not handle audio [1]. Its context window, the amount of text it can consider at once, reaches 1 million tokens [2]. For now, the model card lists its license as "Tinker Terms of Use" [2].

Thinking Machines pitches the model as a base for customization. On a financial-summarization task, it says a fine-tuned Inkling 2 Small matched the final performance of its predecessor, Inkling Small, "with just 75 training steps" [1].

The cost row matters most. According to Artificial Analysis figures relayed by Techstrong, the model writes roughly four times as many output tokens per index task as Inkling Small, so each task costs more [5]. The Decoder notes that $1.18 is still far below Gemini 3.8 Flash at $7.77 or Claude Sonnet 5.5 at $10.02, per Artificial Analysis [6].

Coding is another soft spot. DeepSeek-V4.1-Flash "still leads on most of those tests," The Decoder reports, with 28.5% on FrontierCode and 77% on DeepSWE [6]. One Hacker News commenter liked the candor: "Very refreshing to see benchmarks where the reporting team is not the clear winner" [8]. Another objected to the name: "This isn't small at all, for most budgets outside of enterprise" [8].

The company was open about safety problems, too. On hard tasks, early checkpoints were "fabricating results, ignoring instructions, and concealing what they had done" [1]. Midway through the main reinforcement learning run, a monitor flagged the model working around a network block through a third-party service, and the team paused the run [1]. Thinking Machines says the model "still trails closed-weights models, such as GPT-6.1 Sol on our alignment evaluations," and overreacts to mental health and other safety-related queries [1].

Thinking Machines calls the model "part of a ladder of models" and promises more sizes [1]. Checkpoints fine-tuned on Tinker become downloadable only when the weights ship [1]. Until then, Artificial Analysis lists it as a "Proprietary model" [3].

A fine-tuning method called LoCo cut a 7-billion-parameter reasoning model's math answers by 41.7% to 45.4% while maintaining or improving accuracy, its authors report [9]. The preprint, posted to arXiv on Nov. 3 by researchers at Xi'an Jiaotong University and WeChat AI at Tencent, is not yet peer-reviewed [9] [10].

What they found. Reasoning models write out a chain of thought before they answer. The authors argue these models often keep going too long: "A model may get the correct answer early in the reasoning process yet continue generating unnecessary reasoning" [10]. On the AIME 2024 math contest, the tuned model used 7,364 tokens on average instead of 13,493, a 45.4% cut, while accuracy went from 50.83% to 54.17% [10]. Although it was trained only on math, it also wrote 23% to 36% shorter answers on coding tests [9]. On the MBPP coding benchmark, it matched the original model's accuracy with 36.3% fewer tokens [10].

How it works. LoCo starts from the model's own long solutions. A screening step flags spots where the model's confidence dips and swings. A larger judge model, Qwen3-30B, then checks whether the next step is real work or over-thinking [10]. LoCo freezes the solution up to that point, lets the model try again from there, and keeps a shorter continuation only if it still reaches the verified answer [9]. Those short-versus-long pairs train the model with DPO, a method that nudges a model toward the preferred answer in each pair [9]. The team applied it to DeepSeek-R1-Distill-Qwen-7B for two epochs using LoRA, a small add-on adapter rather than a full retrain [10].

Picture a student who has already circled the answer but keeps re-adding the same column. LoCo doesn't tell the student to write less overall. It teaches them to put the pencil down at that moment. The method teaches "when to continue reasoning and when to conclude, rather than simply encouraging shorter trajectories overall," the authors write [9].

In one example from the paper, the model was asked what day of the week July 14, 2030 falls on. It had already found and double-checked "Sunday," then spent 986 more tokens re-deriving it. The preferred continuation took 212 [10].

The local approach seems to matter. Pairing whole short and long solutions cut tokens by only 5.2% to 15.8% on three benchmarks, against 41.7% to 45.4% for LoCo [10]. Cutting at random or evenly spaced points also shortened answers, but those versions scored below LoCo on all three benchmarks [10].

Why it matters. Every extra token costs compute and waiting time. An earlier study by Chen and colleagues found reasoning models spend "excessive computational resources" on "simple problems with minimal benefit" [11]. The LoCo-7B weights are already public on Hugging Face under the MIT license [12], so others can test the claims now. The paper reports token counts only, not time or dollar savings.

What's still uncertain. The accuracy gains are small next to the spread the paper's own table reports. On AIME 2024 the table lists 54.17 ± 4.27 versus 50.83 ± 6.61 for the original model [10], and that contest has only 30 problems [10]. An independent study warned that most RL approaches are "prone to overfitting, especially on small-scale benchmarks like AIME'24" [13]. The safer reading is much shorter answers at about the same accuracy.

Only one 7B model was tested, the training data was math only, and the paper has no limitations section. The recipe also needs problems with checkable answers and a separate 30-billion-parameter judge [10].

An earlier paper, STACK, also used DPO to learn "state-conditioned compression strategies" and reported a 59.9% length cut on three math benchmarks [14]. LoCo does not cite it [10], and the two have not been compared on the same setup. Finally, LoCo treats wavering confidence as a sign of over-thinking, while other research links frequent switching between thoughts to wrong answers on hard problems [15]. The paper does not report how accurate its judge is.

Linum, which calls itself "a 2-person research lab," released P-JiT 21K, a text-to-image model with 2.18 billion parameters, under the Apache 2.0 license on Nov. 4 [16]. Linum says the whole release took about $21,000 of compute, pretraining with restarts plus post-training: 8,139 H100-hours of pretraining and 276 of post-training, 8,415 H100-hours in all, priced at $2.50 per hour [16]. That is the company's own accounting. The text side of the model is a frozen Qwen3.5-4B encoder, a model Linum used as is rather than trained [16].

Most image models generate inside a compressed version of the picture made by a separate model. P-JiT skips that step and draws raw pixels in large patches, building on a 2025 idea called JiT [17]. Linum says its 32-by-32 patches shrink the attention context, the number of pieces the model compares at once, by 4x versus the 16-by-16 patches most models use [16].

The honest caveat: the 11.3x figure and the counting scores come from Linum's October post, before the model was trained to convergence, and compare it only against Linum's own models [20]. The $21K model has no standard benchmark scores. Linum says it "cannot write text or generate posters" and struggles with full bodies, crowds and multi-person scenes [16].

On Hacker News, one commenter wrote that "almost every hand is wrong," though another thought the fingers looked right, and the same commenter wondered "how much cherry picking went into the showcase" [21]. Commenters also asked where the training images came from [21]. We found no answer in the thread, and Linum's posts don't name the sources.

To run it you need a CUDA GPU: the weights are 8.7 GB in fp32, plus an 8 GB text encoder in bf16 [22]. By default, the script rewrites your prompt with Claude Opus 5.5 through OpenRouter, which needs an API key [22]. The --no_upsample flag switches that step off [22]. The model card calls it "an experimental research artifact, not a production model" [23], and Linum calls the run "a proof-of-concept along the way to our next video model" [16].