
Inkling 2 Mini Ships at 20: a 20GB Model That Thinks in Loops
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 183 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 568 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 385 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 150 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 84 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 121 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 107 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 214 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
The checkpoint is about 20GB, and that number is the argument. Inkling 2 Mini is a 26.4-billion-parameter mixture of experts with 3.4 billion parameters active per token, published by Thinking Machines Lab on 20 November 2026 under Apache 2.0 as a single download that runs on a Mac with 32GB of unified memory or a single NVIDIA DGX Spark. On the same day, Artificial Analysis timed it and scored it: 20.1 on the Intelligence Index v5.0.1, $0.15 per million input tokens and $0.50 per million output, $0.59 to finish one index task, 280 output tokens per second. The interesting part is not the score. It is that the model reaches it by running the same twelve layers three times instead of storing more parameters, which is what makes a 26B checkpoint behave like something larger without needing more memory than a laptop has.
Mix the two kinds of number carelessly and the release reads as a clean win; keep them apart and it reads as something more specific, which is what this piece tries to do. The index, the cost per task, the speed, the verbosity and the licence are Artificial Analysis measurements and metadata read on 21 November 2026. Every coding and reasoning row further down is Thinking Machines' own harness on a checkpoint the company trained, and says so where it appears. Nobody outside the lab has reproduced those rows.
What Thinking Machines published
The release is weights-first. The Hugging Face repository was created on 17 November and last modified on the 20th, which is what a staged checkpoint looks like rather than a surprise drop, and the model card names 20 November 2026 as the date of release. There is a real licence file situation worth noting: the card declares Apache 2.0 in its metadata and the repository tags carry the same, so this is a permissive release rather than the promise-of-one that Inkling 2 Small shipped with on 5 November.
• Parameters — 26.4B total, 3.4B active, and 5.4B effective weight activations per token once looping is counted; Thinking Machines puts each extra loop pass at roughly 1.0B more active parameters.
• Footprint — one checkpoint of about 20GB, 20 layers in NVFP4 and 4 in MXFP8, shipped in the precision it was trained in rather than quantised afterwards.
• Modalities — text and images in through a vision transformer projected into the shared hidden space, text out.
• Context — the model card states up to 1M tokens, and Artificial Analysis records 1M.
• Serving — Hugging Face, SGLang, vLLM, TokenSpeed, NVIDIA Dynamo, Unsloth and llama.cpp GGUFs, Llama App, plus Ollama and Exo Labs on Apple silicon. Post-training recipes exist for Miles and SkyRL.
• Hosted — Tinker, for inference and customisation, with a playground. That is the vendor's own endpoint and the only one; Inkling 2 Mini is not on OrcaRouter, and nothing below implies otherwise.
A companion model ships alongside it. TML-d24-varloop-research is the same architecture trained to accept between one and five loop passes, and it exists so researchers can vary the depth that the production model fixes at three. It is on Hugging Face too, and it is the object most of the monitorability work in this release was done on.

The loop is the product
Inkling 2 Mini is a 24-layer decoder-only transformer with sparse mixture-of-experts feed-forward layers, divided into four blocks of six layers. The first block runs once, the last block runs once, and the two middle blocks — twelve layers — run three times with shared weights. A token therefore passes through 48 layer applications drawn from 24 stored layers of weights.
The repeat is not a copy. Each pass through the middle blocks reads the output of the prelude, and the second and third passes also read what Thinking Machines calls the message: the middle blocks' own output from the previous pass, merged through a learned gate. That gate starts by blocking the message entirely, so training begins as an ordinary non-looping model and the gain on the message is raised over time. It is a training-stability trick, and it is also the reason the loop can be added mid-run rather than trained from zero.
Which layers loop was not arbitrary. The company measured looping the middle twelve against looping the last twelve across 28 evaluations and found the middle won on 27 of them, with a compute-equivalent gain of 1.51–1.91× against 1.24–1.36×. The intuition it gives is that early and late layers are busy translating between tokens and internal representations, while the middle carries the abstract reasoning that benefits from more serial steps.
The experts were pruned rather than designed. Training started at sparsity 48, routing each token to 8 of 384 experts; at one point during midtraining the company kept the 128 most-used experts per layer and healed the damage on 1T tokens distilled from a larger internal Inkling model, then ran SFT and an RL phase totalling 2.2T tokens over 2,000 steps. That is the same post-training mix as Inkling 2 Small, which is why the two models' strengths land in the same places.
Twenty, and what sits under it
The headline score is real and independently measured, and it is also close to the floor of what the index resolves usefully. Artificial Analysis places 20 as well above average for the model's comparison class on the open-weights board, where the median is 10, but the band immediately around 20 is crowded with models a point either side, and the index does not order models that finely. A score of 20 means the model clears a bar, not that it beats everything within a point of it.
What the score does support is the placement that matters for a model this size. The Pareto claim in the launch coverage — frontier intelligence per active parameter — is the one Artificial Analysis's own summary backs, calling it among the leading models in intelligence for its size class and reasonably priced against other open-weights models of similar size. It gets there because of the loop, which buys serial compute without buying parameters, and the claim stops holding the moment you ask the model to know something rather than work something out.
The cost and speed columns are where the release looks strongest. At $0.15 in and $0.50 out it is priced at the bottom of the reasoning tier, and $0.59 per index task is a quarter of what a 27B dense model in the same score band costs. It is fast in the two ways that show up in an agent loop: 9.72 seconds to first answer token and 280 output tokens per second. The compensating weakness is verbosity — 161M output tokens to complete the index run against a peer median of 85.7M, roughly 1.9 times the output of a typical model in its class. On a per-token bill that is fine. On a fixed token budget it means you get about half as many tasks.

Coding rows, vendor-reported
The table Thinking Machines published is the strongest part of its case, and it is entirely its own. These are unreproduced numbers from a vendor harness, several of them run with the company's own agent scaffold rather than a neutral one, so read them as claims about a checkpoint the lab trained.
• SWE-bench Verified — 79.2%, bash only.
• SWE-bench Pro v2, Hard — 77.9%.
• SWE-bench Multilingual — 73.7%.
• DeepSWE v1.1 — 48.2%, against 42.2% reported for Qwen3.8-27B.
• Terminal Bench v2.1 — 71.6%, behind Qwen3.8-27B's 73.0; on Terminal Bench v4.0 both print 5.6%.
• FrontierCode v2.0 — 15.1%, against 14.9% for Qwen3.8-27B.
• Agentic search and tools — BrowseComp 79.0%, DeepSearchQA 85.3%, MCP Atlas 82.7%, each ahead of the Qwen3.8-27B figure in the same table.
• Vision and documents — MMMU Pro 74.9%, OmniDocBench v1.6 92.8%, ScreenSpot Pro 76.5% plain and 87.7% with Python tooling.
• Reasoning — HLE-Diamond 15.2%, CritPt v2 31.4%, ARC-AGI 84.5% on v1 and 49.3% on v2, LongBench v2 66.9%.
The shape of that table is consistent with the architecture, and it is the most useful thing in the release. The rows where the model looks strong are the ones that reward persistence and tool use over a long horizon: search, repository-scale patching, document parsing. The rows where it looks ordinary are the ones that need a single correct recall or a hard scientific inference. Looping adds serial steps. It does not add stored facts.
Looping buys reasoning, not facts
That distinction is the release's own thesis, and the company tests it rather than asserting it. On AA-Omniscience, the factuality benchmark, the vendor table prints −8.6 for Inkling 2 Mini with a 63.4% non-hallucination rate. A negative score there means the model is worse than declining to answer: it produces confident errors. The company's framing is that the knowledge gap is in line with other models of the same total size, and that the model covers it by searching well — which is exactly what the BrowseComp and DeepSearchQA rows show, and exactly why the honest summary of this model is "good at finding, weak at remembering".
The variable-loop companion makes the point more sharply. Its no-chain-of-thought time horizon — how long a task it can finish at 50% reliability without writing reasoning down — rises from 16 seconds at one pass to 37 seconds at three, then flattens, while serial reasoning on an N-Hop lookup task keeps improving out to ten passes, twice the depth it was trained at. Gains concentrate in multi-hop retrieval, math and science reasoning, and long-horizon agentic work. Closed-book factual QA barely moves.
The monitorability result is the part researchers will care about most, because it is the objection to looped models. If a model does more computation inside a token, less reasoning appears in the visible chain of thought, and a chain-of-thought monitor sees less. Thinking Machines measures this and finds the effect smaller than feared at this scale: the variable-loop model's reasoning at five passes was less than half as long as at one (1,661 against 3,703 tokens on average) with sentence length and dictionary-word share unchanged, which reads as more direct reasoning rather than less legible reasoning. On its LinuxArena stealth evaluation, Inkling 2 Mini at 3.9% sits far below Qwen3.8-27B's 20.5%. None of that is a guarantee — it is a first measurement on a class of architecture nobody has had downloadable weights for before, and the company says plainly that it released the model partly so others can check.
What it costs to run
Self-hosting is the advertised mode and the arithmetic is easy. A 20GB checkpoint on a machine you already own has no marginal token cost at all, and the vendor's own efficiency study — run by Stanford's Hazy Research and Scaling Intelligence labs on early access, at MXFP8 and low reasoning effort — puts it at 0.703 intelligence per watt on a DGX Spark against 0.430 for Qwen3.6-35B-A3B and 0.159 for Gemma 4 26B-A4B. That is a vendor-adjacent result on a single device, not a fleet measurement, but the direction is not subtle.
The honest case against running it yourself is that a looped architecture has roughly one day of serving maturity. SGLang, vLLM and the rest list support, and the model card warns in its own words that looped-model runtime support is new and speed varies. Throughput you measure on your own box is a function of which engine you picked and how recent its kernels are, and that will change faster than anything else about this model.
Where a router earns its place here is not in serving Mini — we do not host it. It is in what you put around it. An evaluation of an unproven architecture should never be the only thing between your users and an error page, and the way to arrange that is to call a model you trust for the production path while Mini runs beside it. Qwen3.8-Flash is on OrcaRouter at $0.15 per million input and $0.47 per million output, the provider's list price passed through with no markup, and the routing DSL and automatic failover let you put Mini behind it the day Mini reaches an endpoint we can reach. One key covers both sides of that experiment, which is the whole of the argument.

Four things to watch
The first is independent reproduction. Artificial Analysis has the index, cost, speed and verbosity — measured — but the coding table is the lab's. Until someone outside Thinking Machines runs SWE-bench Pro v2 or DeepSWE on the public weights, 77.9% and 48.2% are claims.
The second is whether a hosted endpoint appears beyond Tinker. A 20GB Apache 2.0 checkpoint is exactly the kind of artifact third-party providers pick up quickly, and the price competition that would make $0.59 per task look durable only starts when there is a second seller.
The third is the loop itself. If the pilot works, the technique is portable: three passes on the middle layers, a gate to merge the message, no extra parameters. Every small-model recipe in the field becomes a candidate for it, and the interesting question stops being whether this model scores 20 and becomes whether the next one scores 24 for the same memory.
The fourth is the part the launch does not settle. Looped models move reasoning out of the visible chain of thought and into the forward pass. Thinking Machines found no meaningful loss of monitorability at 5.4B active parameters. Nobody has published the same measurement at 50B. That is the question this release makes askable, and it is a better reason to care about a 20GB download than the score.
For anyone deciding this week, the split is clean. If you want a competent agentic coder that runs offline on hardware you already own, and you are willing to sandbox it and to test factual recall before trusting it, Inkling 2 Mini is the most interesting small model available right now and the price of trying it is a download. If you need an answer you can put in front of a customer without a retrieval layer behind it, this is not that model, and the 20 on the index will not make it one.
