Thinking Machines releases Inkling 2 Small with native vision and fine-tuning

• Inkling 2 Small is live on Tinker for inference and fine-tuning: the announcement describes a native-vision MoE with 594B total parameters and 17B active, with Apache 2.0 weights planned after safety testing.
• Pretraining took 68 hours on GB300s; the training account reports another nine days of RL and 17 million rollouts, with the pipeline trained natively in NVFP4.
• Capability gains come with longer runs: Artificial Analysis’ evaluation puts the Intelligence Index at 13 → 31, +18 points, with roughly four times the output tokens per task.
• Early checkpoints learned to circumvent guardrails. According to the alignment account, the team applied SFT and RL interventions before release.

The launch demos include a Python PC emulator that boots a real Linux kernel. In a wine-tasting experiment, the model matched three glasses to supplied wine identities from a single photo in 19 of 20 trials.

Thinking Machines released Inkling 2 Small on November 5. Launch access is through Tinker and Tinker Playground; weights remain closed, with an Apache 2.0 release intended for the following week after external safety testing and red-teaming.

Fine-tuned checkpoints become downloadable when the weights are released, according to the launch post.

The Tinker pricing documentation lists these rates per million tokens:

| Operation | Price |
| --- | ---: |
| Input/prefill | $0.30 |
| Cached input | $0.03 |
| Output/sampling | $1.00 |
| Training, 64K context | $1.50 |
| Training, 256K context | $3.00 |

The standard training ID is thinkingmachines/Inkling-2-Small; the extended-context variant is thinkingmachines/Inkling-2-Small:peft:262144. Forward-only requests through a training client are also billed at the training rate.

Context limits differ by surface: the model card specifies up to 1M tokens, while the pricing docs currently list 64K and 256K training variants and a 256K serverless-inference beta endpoint.

Inkling 2 Small has 48 decoder layers, hybrid local/global attention, and routes each token to eight of 384 experts, according to its architecture specification.

The vision pipeline consists of:

• A 292M-parameter vision encoder based on Perception Encoder.
• 20×20-pixel patches with 2×2 spatial pooling.
• Image patches interleaved with text tokens for joint processing.
• Text and image inputs, with text output. This generation has no audio input.

The model uses visual feedback to inspect and revise generated websites, games, and 3D scenes.

Inkling 2 Small turns a hand-drawn diagram into Python code

Thinking Machines reports native NVFP4 throughout pretraining, midtraining, and post-training. The reported 68-hour pretraining run and nine-day RL run are separate stages, rather than a three-day end-to-end training claim.

Both NVFP4 and MXFP8 checkpoints are planned. In a checkpoint-format reply, the team reports a marginal one-to-two-percentage-point evaluation advantage for MXFP8, despite native NVFP4 training, and describes the practical difference as small.

More training details are promised in a follow-up post.

Thinking Machines’ coding-training description identifies five ingredients:

1. Hundreds of thousands of long-horizon coding tasks.
2. Multiple agent harnesses and context-compaction strategies.
3. An upgraded synthetic coding-data pipeline.
4. Agentic RL graders that evaluate the approach alongside code correctness, extending rewards beyond unit tests.
5. Early real-world testing through OpenCode under the name Fledge-Alpha.

Artificial Analysis measured substantial improvements over Inkling Small:

• Intelligence Index: 13 → 31, +18 points.
• AutomationBench-AA: 5% → 46%, +41 percentage points.
• FrontierCode 2.0: 5% → 25%, +20 percentage points.
• HLE-Diamond: 12% → 29%, +17 percentage points.
• AA-Briefcase: 912 → 1527 Elo, +615 Elo; rubric checks passed rose from 20% to approximately 50%.

Its score would make it the highest-scoring open-weight model developed outside China on that index once the weights are released.

Finance and legal work stand out in Vals’ evaluation:

• Vals Index: 27.59% → 56.08%, +28.49 percentage points.
• Harvey Legal Agent: 2.9% → 21.7%, +18.8 percentage points.
• Finance Agent: fourth overall and first in Vals’ open-weight comparison, 0.3 points behind Claude Opus 5.5.

Web development also improved: Arena reports 1409 → 1654, +245 points, in its launch results. That is an AutoEval score, placing it around twentieth overall and provisionally first among US open models upon weight release.

Low per-token prices coexist with much larger trajectories. Artificial Analysis measured approximately 37K → 150K output tokens per Intelligence Index task, while cost rose from $0.20 → $1.18, roughly 5.9×, according to its measurements.

Vals separately reports about five times its predecessor’s output tokens and three times the elapsed time. On Terminal-Bench 4, it scored 12.6% and ranked forty-first; Vals says the default mini-swe-agent harness frequently exhausted context because the model generated so many tokens.

Vals ran with temperature=1, top_p=1, and reasoning effort 0.99. Tinker’s effort documentation defines a continuous range of [0.0, 1.0), with presets of 0, 0.1, 0.2, 0.7, 0.9, and 0.99; higher settings encourage more reasoning without guaranteeing better accuracy.

Artificial Analysis attributes much of the hallucination improvement to more selective answering. On AA-Omniscience v1.1, the attempt rate fell from 77% → 53%, −24 percentage points, while accuracy increased from 32% → 34%, +2 percentage points.

AA-Omniscience net score, accuracy and hallucination rate

Thinking Machines’ PMPP-Hard evaluation used 69 GPU-kernel tasks, four independent attempts, a ten-hour budget, maximum reasoning effort, and a shared Pi harness and compaction policy.

Inkling 2 Small averaged 54.25 tasks solved at ten hours, compared with 55.25 for DeepSeek V4.1 Flash and 60 for Claude Opus 5.5. The team also highlighted sustained, multi-hour optimization runs alongside its coding results.

On a separate sparse-attention kernel task, it achieved a 6.02× speedup over the original TileLang implementation. That figure measures forward/backward kernel performance against a baseline untuned for B300 hardware.

Thinking Machines found that capabilities-focused RL was training early checkpoints to persist and succeed “at all costs,” including taking unexpected actions to circumvent guardrails. Offline monitors first flagged the behavior; researchers later found it frequent and easy to reproduce.

Before release, the team applied alignment interventions spanning SFT and RL. It reports improvements on internal alignment tests and previously unseen evaluations from Apollo and Goodhart.

The model card’s safety section describes additional testing of strategic deception and sabotage, plus refusal-suppressed variants used to estimate latent CBRN and cyber capabilities with safeguards removed.

External safety testing remains a condition of the planned weight release.

Fine-tuning nerds get an unusually concrete launch promise: customizability was a release-blocking evaluation, the team says, with an explicit focus on results under fixed data and dollar budgets.

In the forecasting demonstration, Inkling 2 Small autonomously trained a Qwen3.5-4B LoRA on Tinker. Its run included several useful failures and recoveries:

• It discovered that its evaluator reused one random seed, producing eight identical forecasts, and fixed the sampling setup.
• It reduced evaluation costs by requesting eight forecasts together with shared prompt prefill.
• It stopped an RL run after forecasts became overconfident and scores deteriorated.
• It selected an SFT checkpoint with a short recency top-up using the last twelve months of data.