The first model of Thinking Machines Lab's second generation, from Mira Murati's lab: a 594B-parameter mixture-of-experts that activates only 17B per token. Launched November 5, 2026 on the Tinker platform, its weights followed a week later under Apache 2.0, exactly as promised. Its Artificial Analysis index of 31 is the highest for an open-weights model built outside China — but it is verbose, slow, and scores just 13.1% on Terminal-Bench 4.0. This page collects the official specs, the ways to try it, pricing, and the first hands-on reviews in English and Chinese — with unverified claims clearly marked.
What is confirmed: this is a light, efficient open-weights model, not a flagship; its selling point is value and downloadability, not chart position.
thinkingmachines/Inkling-2-Small (NVFP4 main checkpoint) and Inkling-2-Small-MXFP8Seven days from launch to open weights — the lab kept its launch-day promise to the day.
First-generation Inkling launches with full weights open on day one; Inkling-Small (276B/12B) ships alongside.
The lab publishes its staged-access framework: inference, then monitored hosted fine-tuning, then potentially open weights, each step gated on evidence.
Inkling 2 Small launches on Tinker for inference and fine-tuning; the lab promises Apache 2.0 weights "next week" after external safety testing and red-teaming. Artificial Analysis index of 31 published.
Free variants go live on OpenCode (inkling-2-small-free) and OpenRouter; first analyses publish (eesel.ai, theddari, llmlearner); thoughtfullab tests it on DiligenceBench, producing a full diligence memo for $0.69 and spotting a $1.2B misstatement.
Weights land on Hugging Face (20:50 UTC): NVFP4 main checkpoint plus an MXFP8 checkpoint, Apache 2.0; day-0 support from vLLM and SGLang; Plotly's data-analysis benchmark publishes the same day (9 points above the predecessor).
The lab says Inkling 2 Small belongs to a "family of different sizes" with more variants planned; no timetable.
As published by the lab (model card and Hugging Face repo, checked Nov 12–13, 2026); undisclosed items are marked as such, not filled in by guesswork.
Fastest are the Tinker Playground and the free variants; self-hosting the weights starts with confirming you have 500GB of VRAM.
Thinking Machines' own platform: Playground for quick tests, API for inference and fine-tuning. Model IDs thinkingmachines/Inkling-2-Small (64K) and thinkingmachines/Inkling-2-Small:peft:262144 (256K). Note: serverless inference is currently a waitlist beta; fine-tuning needs the tml-renderers package.
Public repos, no application needed: the NVFP4 main checkpoint (25 safetensors files) and the MXFP8 checkpoint. Run locally with SGLang, vLLM, TokenSpeed or Unsloth (official recipes for each; vLLM had day-0 support). Hardware bar: at least 500GB of aggregated VRAM for NVFP4 (2× B300 or 4× H200), 800GB for MXFP8. The lab recommends NVFP4 for all uses — MXFP8 scores only 1–2 points higher.
Together AI, NVIDIA, Fireworks, Baseten, Nebius, Modal and Cloudflare Workers AI all serve it; Together and Cloudflare charge the same $0.30 / $1.00. On OpenRouter it is thinkingmachines/Inkling-2-Small, same price plus a free variant.
OpenCode lists inkling-2-small-free for a limited time; OpenRouter's free variant is likewise free (co-founder Soumith Chintala described it as "free in exchange for data use"). Free allowances can change at any time — check each platform's page.
Hosted fine-tuning on Tinker ($1.50/1M tokens at 64K; $3.00 at 256K); with your own copy of the weights, post-train with Miles, SkyRL or Snowflake Cortex. The Apache 2.0 license allows commercial use, modification and redistribution.
The per-token price looks cheap, but the per-task bill is not small — verbosity is billed. Prices checked Nov 7–12, 2026; the checkout page is authoritative.
| Item (per 1M tokens) | Price |
|---|---|
| Tinker serverless input | $0.30 |
| Tinker serverless cached input | $0.03 |
| Tinker serverless output | $1.00 |
| Tinker fine-tuning (64K context) | $1.50 |
| Tinker fine-tuning (256K context) | $3.00 |
| Together AI hosted (in / out) | $0.30 / $1.00 |
| Cloudflare Workers AI / OpenRouter (in / out) | $0.30 / $1.00 |
| Weights download (Hugging Face) | $0 (Apache 2.0) |
Artificial Analysis measured $1.18 per index task, nearly 6× the predecessor Inkling Small ($0.20), and 15.7 minutes per task (was 7.0). The cause is not the unit price — $0.30 input is only slightly above the $0.24 class median — but output volume: about 150K output tokens per task, 4× the predecessor's ~37K. AA called the pricing "somewhat expensive" and the model "notably slow per task and very verbose." There are cheaper cases in the other direction: thoughtfullab produced a full stock diligence memo for $0.69 in a finance test.
Official pages: Thinking Machines announcement · Hugging Face weights repo · Official model card
All from the first wave, Nov 5–13: small samples, mixed methods — read them as individual cases. Each entry carries a verdict badge; "Read the full article" opens our rewritten deep dive, with the original source linked at the end.
Source: eesel.ai, around Nov 8, 2026 (AA index 31 vs 13 for the predecessor)
Coding agents, vision and document work all improved, and a 73.3% non-hallucination rate shows the factuality gap narrowed; but the weights were still unreleased and chat is not its strength. AA measured $1.18 and 15.7 minutes per task, against $0.20 and 7.0 minutes for the predecessor — smarter, and twice the bill and time.
Read the full article → (original write-up by this site)
Source: Plotly, Nov 12, 2026 (real-world data analysis tasks)
On real data-analysis tasks, Inkling 2 Small scored 9 points above Inkling-Small at about the same cost and twice the speed. One of the few third-party measurements published on weights-release day — and it tests real workflows, not standard benchmarks.
Read the full article → (original write-up by this site)
Source: CellCog, Superpower Daily, AI Parabellum, Nov 12, 2026
At 20:50 UTC on Nov 12 the weights landed on Hugging Face as promised: NVFP4 main checkpoint (25 safetensors files) plus an MXFP8 checkpoint, Apache 2.0, public with no application. But NVFP4 needs 500GB of aggregated VRAM and MXFP8 800GB; vLLM and SGLang supported it on day zero, and 7 hosted providers went live the same day. Safety-test results were not published.
Read the full article → (original write-up by this site)
Source: official model card; Artificial Analysis; ARC Prize, Nov 5/12, 2026
SWE-Bench Pro v2 (Hard) 86.3%, PerceptionBench 64.7%, ARC-AGI-2 81.7% (best open-weights score) and a 73.3% AA-Omniscience non-hallucination rate are the highlights; but Terminal-Bench 4.0 is just 13.1%, under half of GLM-5.3-Flash's 32.8%. Note the official table was all measured at xhigh effort and temperature 1.0, mostly self-measured by the vendor.
Read the full article → (original write-up by this site)
Source: this site's synthesis of official model cards, the AA index and launch data, Nov 13, 2026
An AA index of 31 is the highest open-weights score outside China, but DeepSeek V4.1 Flash (32.6) and Claude Haiku 5.5 (36.5) sit above it; on Terminal-Bench 4.0, Kimi K3.1 Pro (65.7%), Hy4 (44.4%) and GLM-5.3-Flash (32.8%) all lead. Its relative edge is vision, finance and ARC reasoning — plus the Apache 2.0 + downloadable combination.
Read the full article → (original write-up by this site)
Source: this site's synthesis of Tinker pricing, AA measurements and hosted-provider quotes, Nov 13, 2026
$0.30 input is only slightly above the $0.24 class median, but ~150K output tokens per task push the per-task cost to $1.18. The free variants (OpenCode, OpenRouter) are fine for testing; long-term use means either accepting hosted prices or carrying the 500GB-VRAM hardware bill yourself.
Read the full article → (original write-up by this site)
Source: AIsa Studio, AGI Hunt, Nov 7/13, 2026
As of Nov 13, no original deep review from Chinese outlets such as 机器之心, 新智元 or 量子位 was found. Existing Chinese coverage is mainly AIsa Studio's translations of the-decoder and AGI Hunt's aggregation: launch news, the open-weights release, the vLLM recipe and the OpenCode free launch are all covered, but with no independent test data.
All 6 articles below are original rewrites and re-organized deep reads by this site — restructured, with conclusions and selection advice added, and facts checked against the original reviews. Read any of them on this site; the original source is linked at the end of each.
The official table shows strengths and weak spots in equal measure — and everything in it was measured at xhigh effort, temperature 1.0, with a 1M-token trajectory cap. Read the conditions before the scores.
| Test (official model card, vendor-measured) | Inkling 2 Small | Comparison |
|---|---|---|
| AA Intelligence Index v5.0 | 30.6 (≈ 31) | DeepSeek V4.1 Flash 32.6; Haiku 5.5 36.5; Gemini 3.8 Flash 30.8 |
| SWE-Bench Pro v2 (Hard) | 86.3% | Second only to Kimi K3 in the table |
| PerceptionBench | 64.7% | Kimi K3 58.5% |
| EnterpriseBench-CoreCraft | 71.8% | GLM-5.3-Flash 46.7% |
| JobBench | 70.5% | — |
| FrontierCode v2.0 | 25.4% | DeepSeek 28.5%; predecessor 5% |
| DeepSWE v1.1 | 72.1% | DeepSeek 77.0%; GLM-5.3-Flash 63.4% |
| Vals Finance Index | 61.5% | DeepSeek 60.9%; GLM-5.3-Flash 55.8% |
| Terminal-Bench 4.0 | 13.1% | GLM-5.3-Flash 32.8%; DeepSeek 33.8%; GLM 5.3 41.9% |
| AA-Briefcase v1.1 | 1527 Elo | Predecessor 912 Elo |
| AA-Omniscience v1.1 (non-hallucination rate) | 73.3% | Highest in the table; net score 16.1 |
| HLE-Diamond | 28.7% | — |
| MMMU Pro (Standard 10) | 81.6% | — |
| ARC-AGI-2 / ARC-AGI-1 | 81.7% / 96.5% | Best open-weights score (ARC Prize) |
| DesignArena (website leaderboard snapshot) | 1,340 | — |
| Measurement | Source | Inkling 2 Small | Comparison |
|---|---|---|---|
| AA Intelligence Index v5.0.1 | Artificial Analysis, Nov 5 | 31, 4th of 29 in its price class | Predecessor 13; DeepSeek V4.1 Flash 32.6; Haiku 5.5 36.5 |
| AA cost / time per task | Artificial Analysis | $1.18 per task, 15.7 minutes | Predecessor $0.20, 7.0 minutes |
| AA output tokens per task | Artificial Analysis | About 150K | Predecessor about 37K (4×) |
| Plotly data-analysis benchmark | Plotly, Nov 12 | 9 points above the predecessor | Similar cost, twice as fast |
| DiligenceBench (finance diligence) | thoughtfullab, Nov 6–7 | Full diligence memo for $0.69 | Spotted a $1.2B misstatement; beat Kimi K3 and GLM 5.2 on PostTrainBench |
| LongBench v2 (long context) | BenchLM, Nov 13 | 70.6%, #1 of 27 models | Vendor-reported; Beam 1.0 second at 67.6%; long-context retrieval and reasoning |
| Internal long-horizon coding demo | Thinking Machines (not independently reproduced) | 54.25 of 69 tasks solved | Claude Opus 5.5 solved 60 |
Full roundup in our benchmarks deep dive →
| Dimension | Inkling 2 Small | Inkling Small |
|---|---|---|
| Released | Nov 5, 2026 | July 2026 |
| Parameters | 594B total / 17B active | 276B total / 12B active |
| AA index | 31 | 13 |
| Audio input | None | Yes |
| Cost per task (AA) | $1.18 | $0.20 |
| Time per task (AA) | 15.7 minutes | 7.0 minutes |
| Output tokens per task | About 150K | About 37K |
| AA-Briefcase | 1527 Elo | 912 Elo |
| Weights | Apache 2.0 (Nov 12, 2026) | Apache 2.0 (open at launch) |
The progress is real (AA index +18), and so is the cost: nearly 6× the per-task price, twice the time, 4× the output tokens.
| Model | AA index | Terminal-Bench 4.0 | API price $/1M (in/out) | License / weights |
|---|---|---|---|---|
| Inkling 2 Small | 31 | 13.1% | 0.30 / 1.00 | Apache 2.0, published |
| DeepSeek V4.1 Flash | 32.6 | 33.8% | — | Open weights |
| Kimi K3.1 Pro | 46 | 65.7% | — | Open weights |
| Tencent Hy4 | 35 | 44.4% | 0.83 / 2.50 | Apache 2.0, weights pending |
| GLM-5.3-Flash | 27.8 | 32.8% | — | Open weights |
| Ripple-65B Preview | — | — (TB 2.1: 80.52) | 0.20 / 0.60 | Apache 2.0, published |
Beyond scores, access channels matter for selection. Inkling 2 Small has a free variant on OpenCode (inkling-2-small-free, limited time) and another on OpenRouter; on Google Trends, opencode ranks among Inkling 2 Small's top related queries, suggesting many users first meet the model through OpenCode. Weights already published + several hosts at the same price + free entry points is a fairly complete package among current open models.
Read the deep comparison: Inkling 2 Small vs open-weight rivals →
No large rumor wave around Inkling 2 Small, but the following claims are circulating or easy to confuse — each checked below.
November 5 was the hosted Tinker launch with closed weights; the open weights only reached Hugging Face on November 12 at 20:50 UTC. A full 7 days separate the two.
The lab requires at least 500GB of aggregated VRAM for NVFP4 (2× B300 or 4× H200) and 800GB for MXFP8. The "1,100 tok/s on one card" claim circulating online comes from a Chinese aggregator's retelling, contradicts the VRAM bar, and is unverified — this page does not accept it.
True only on the AA index: 31 is the highest open-weights score outside China (AA's own words). But it trails several Chinese rivals on Terminal-Bench 4.0, and its AA index is below DeepSeek V4.1 Flash (32.6). "Strongest" depends on what you measure.
It depends on the comparison: 4th of 29 models in its price class, up 18 points from the predecessor's 13 — but in a different tier from flagships (Fable 5.2 at 65, Carbon at 60). It is a light, efficient model, not a flagship — 31 is a strong score for its class.
The architecture limit is 1M tokens, but Tinker currently serves only 64K and 256K configurations; the full 1M needs a local deployment (and the VRAM to match). API users cannot reach 1M today.
The official announcement is November 5; Together AI's catalog lists November 4 — a one-day discrepancy, with the official date standing. Some Chinese aggregators wrote November 6, which is the Beijing-time conversion of the announcement.
November 5, 2026 (Thinking Machines Lab announcement), launching first on the Tinker platform; the open weights reached Hugging Face on November 12, 2026. See the timeline in section 02.
On Hugging Face, free to download: main repo thinkingmachines/Inkling-2-Small (NVFP4 main checkpoint) and second repo thinkingmachines/Inkling-2-Small-MXFP8; published November 12, 2026 under Apache 2.0, public repos with no application needed, commercial use, modification and redistribution allowed.
Open Thinking Machines' official repos on Hugging Face and download directly — no application needed. Two checkpoints: the NVFP4 main checkpoint (25 safetensors files, at least 500GB of aggregated VRAM) and the MXFP8 checkpoint (at least 800GB); the lab recommends NVFP4 for all uses, since MXFP8 scores only 1–2 points higher. Run locally with SGLang, vLLM, TokenSpeed or Unsloth, all with official recipes. See our open-weights release deep dive.
At least 500GB of aggregated VRAM for the NVFP4 main checkpoint (2× NVIDIA B300 or 4× H200); at least 800GB for MXFP8 (4× B300 or 8× H200). The lab recommends NVFP4 for all uses — MXFP8 scores only 1–2 points higher. Local inference runs on SGLang, vLLM, TokenSpeed or Unsloth, all with official recipes.
Tinker is Thinking Machines Lab's own fine-tuning and inference platform: the Playground for quick tests, the API for inference and fine-tuning, model IDs thinkingmachines/Inkling-2-Small (64K) and thinkingmachines/Inkling-2-Small:peft:262144 (256K). Pricing: serverless inference $0.30 in / $1.00 out per 1M tokens, $0.03 cached input; fine-tuning $1.50 (64K) / $3.00 (256K) per 1M tokens. Note serverless inference is currently a waitlist beta. See our Tinker pricing guide.
Yes. OpenCode's inkling-2-small-free is free for a limited time, and OpenRouter has a free variant; the weights themselves are free to download (Apache 2.0), though self-hosting needs 500GB of VRAM hardware. Free allowances can change at any time — check each platform's page.
4th of 29 models in its price class, up 18 points from Inkling Small (13); the highest open-weights score outside China. But below DeepSeek V4.1 Flash (32.6) and Claude Haiku 5.5 (36.5), and in a different tier from flagships (Fable 5.2 at 65, Carbon at 60).
It is the weakest item in the official model card, far below GLM-5.3-Flash (32.8%) and DeepSeek (33.8%). Terminal-Bench tests terminal-agent work (completing tasks autonomously in a command-line environment), a different skill from SWE-Bench Pro's code patches (86.3%). The lab has not explained the low score; terminal-agent work is currently not its strength.
On Terminal-Bench 4.0, Kimi K3.1 Pro (65.7%) leads by a wide margin; Ripple-65B's 80.52 is on Terminal-Bench 2.1, not directly comparable with 4.0. Inkling 2 Small's relative strengths are vision (PerceptionBench 64.7%), finance (Vals 61.5%) and ARC reasoning (81.7%), plus the Apache 2.0 + downloadable combination. See our open-weights comparison.
No. The first-generation Inkling supported audio; generation two dropped audio input, leaving text and image input with text output.
The architecture limit is 1M tokens; but Tinker currently serves only 64K and 256K configurations, so the full 1M needs a local deployment.
The lab says Inkling 2 Small belongs to a "family of different sizes" with more variants planned, but has given no timetable or specs. The first-generation family includes Inkling-Small (276B/12B).
All facts on this page come from the public sources below (checked Nov 12–13, 2026).