Inkling

Sources#

• Inkling: Our Open-Weights Model

What it is#

Thinking Machines Lab's first from-scratch model release (July 2026), with full weights on Hugging Face (original checkpoint plus NVFP4 for Blackwell). A Mixture-of-Experts transformer, 975B total / 41B active, context up to 1M tokens, pretrained on 45T tokens of text, images, audio and video. Alongside it: a preview of Inkling-Small — 276B total / 12B active, same recipe, weights promised after testing completes.

The positioning is explicit and unusual: "Inkling is not the strongest overall model available today, open or closed." Its pitch is being the best base for customization — multimodal, efficient, and available for fine-tuning on Tinker, TML's hosted fine-tuning platform (64K/256K context options). The model exists to feed the platform: the announcement's centerpiece demo has Inkling fine-tuning itself on Tinker (writing, running, and evaluating its own lipogram training job), and the release ships with cookbook recipes, a Playground chat console, and a tml-renderer library for sampling/post-training with tool calls and multimodal inputs. This is a third open-weight strategy, distinct from both poles in The Open-Weight Frontier Gap: not approach the frontier with sparsity (GLM, Kimi, DeepSeek) and not approach the device with efficiency (Gemma 4), but approach the fine-tuner with adaptability.

Architecture#

Departures from the common recipe, each argued from efficiency or long-context performance:

• MoE largely follows DeepSeek-V3: 256 routed + 2 shared experts per layer, 6 routed active per token, sigmoid router with auxiliary-loss-free load balancing; routed and shared scores normalized jointly.
• Attention: sliding-window and global layers interleaved 5:1 with 8 KV heads — the same ratio Gemma 4 chose for KV-cache economy, now from a second lab. Relative positional embedding instead of RoPE, which TML claims performs and extrapolates better at long context — a departure from field consensus worth watching for replication.
• Short convolutions after the key/value projections and on the attention/MLP residual-branch outputs.
• Encoder-free multimodality: audio in as dMel spectrograms, images as 40×40-pixel patches through a four-layer hMLP, both via lightweight embedding layers into the shared transformer — the third instance of Encoder-Free Early Fusion, and the second from TML itself (after TML-Interaction-Small), now at 975B open-weight scale.

Training#

Pretrained on NVIDIA GB300 NVL72 systems with a hybrid optimizer — Muon for large matrices, Adam elsewhere — and weight decay coupled to the square of the learning rate (from TML's modular-manifolds research), which they report kept weight norms stable across horizons.

Post-training was bootstrapped with SFT on synthetic data generated by open-weights models including Kimi K2.5 — a from-scratch pretrain whose post-training seed corpus comes partly from a competitor's open model; the bootstrap is described as a small fraction of compute. The majority went to large-scale asynchronous RL: 30M+ rollouts over two long continuous runs, with held-out reasoning aggregate (AIME, HLE, GPQA) improving log-linearly throughout (vendor chart).

Two RL findings worth more than the benchmarks:

• Controllable thinking effort is trained, not scaffolded: effort levels were specified per-sample via system message plus an adjusted per-token cost, teaching a continuous effort dial (0.2–0.99) settable from inside a harness.
• Emergent CoT compression: over RL the chain of thought became telegraphic — dropping articles and connectives ("We need to understand" → "We need determine") — with no reward targeting it; efficiency pressure alone drove the compression. Cognition reports the same during SWE-1.7 training. Recorded in Chain-of-Thought Monitorability, where the monitorability implications live.

Where it lands (vendor-reported, effort=0.99)#

Mid-pack among the open MoE giants on capability, ahead on breadth, safety, and token-efficiency:

• Behind GLM 5.2 / Kimi K2.6 on hard reasoning and coding: HLE text-only 29.7% (GLM 5.2: 40.1), Terminal Bench 2.1 63.8% (GLM 5.2: 82.7), SWE-bench Pro 54.3% (GLM 5.2: 62.1).
• Ahead of the compared open field on: MCP Atlas (74.1%), IFBench (79.8% — above every closed model in TML's own table, including Claude Fable 5 at 63.5), SimpleQA Verified (43.9%), audio (56.6% AudioMC vs 24–38 for open omni specialists; 91.4% VoiceBench).
• Effort curve: matches Nemotron 3 Ultra on Terminal Bench at roughly a third of the tokens. TML publishes the full effort/performance sweep against competitors' default operating points — the closest any vendor release has come to the curve-not-grid reporting Compute-Controlled Benchmarking demands, though only for its own model.
• Safety: FORTRESS Adversarial 78.0% — the strongest built-in safeguards among compared open-weights models — with 95.9% on benign look-alikes; StrongREJECT 98.6%. External testers for CBRN/cyber/loss-of-control and for sycophancy/vulnerable-user/manipulation vectors. Unlike Gemma 4's untabulated safety prose, the numbers are printed — see Open-Weight Elicitation Irreversibility for what a single-point evaluation of open weights can and can't establish.
• Calibration: ForecastBench Brier Index 61.1 (no search) — level with Gemini 3.1 Pro, above GPT-5.5 (59.1) and well above Claude Opus 4.8 (54.6) in TML's table. The training recipe behind this is Trained Calibration.
• Benchmark hygiene notes: evals at effort 0.99, temperature 1.0, 256K-token trajectory caps; externally reported (Artificial Analysis) numbers used where available; Terminal Bench rollouts found contaminated via web search were scored 0 — a small, voluntary instance of the discipline Benchmark Contamination and Decontamination argues for.

All of the above is vendor-claim: TML's own tables, some on a pre-release checkpoint (the ForecastBench results were run June 30–July 13, 2026 on a different checkpoint than shipped).

Inkling-Small and the interaction-model connection#

Inkling-Small is a 276B MoE with 12B active — exactly the shape of TML-Interaction-Small, TML's May 2026 interaction model. The announcement states Inkling's design goal is to serve as the background reasoning model in the interaction-model system (Interaction / Background Model Split). Read together: the fast, present half (TML-Interaction-Small) and the deep, asynchronous half (Inkling) of TML's dual-model architecture are now both public, and the small sibling shares the interaction model's dimensions. Inkling-Small matches or exceeds full Inkling on several benchmarks (HLE with tools 46.6 vs 46.0, IFBench 83.4 vs 79.8, GPQA 88.3 vs 87.2) — attributed to pretraining-recipe improvements made after the big run — while trailing where scale stores knowledge (SimpleQA 20.9 vs 43.9).

Ecosystem#

Day-0 partnerships across the serving stack: APIs on Together AI, Fireworks, Modal, Databricks, Baseten; inference/RL support in SGLang and Miles (RadixArk), vLLM (Inferact), TokenSpeed (Lightseek), llama.cpp (Unsloth); Hugging Face transformers integration. Tinker offers Inkling at a 50% launch discount.

Connections#

• Thinking Machines Lab — the lab; this is its pivot from research previews to shipped open weights
• TML-Interaction-Small — sibling model sharing Inkling-Small's 276B/12B shape; the interaction half of the pair
• Interaction / Background Model Split — Inkling is the named background-reasoning half of the split
• Encoder-Free Early Fusion — third instance of the design, first at ~1T open-weight scale
• The Open-Weight Frontier Gap — a Western entrant to the frontier-open MoE table, competing on a customization axis rather than peak capability
• Trained Calibration — the epistemics training recipe (proper scoring rules, abstention-aware rewards, dual graders)
• Chain-of-Thought Monitorability — the emergent RL compression of its reasoning traces
• Compute-Controlled Benchmarking — publishes its own effort/performance curve; competitors still shown as points
• Open-Weight Elicitation Irreversibility — tabulated safety numbers and a hosted fine-tuning path complicate, without resolving, the open-weight safety problem
• Gemma 4 — the other 2026 Western open-weight release; opposite strategy, convergent 5:1 attention ratio
• Kimi (Moonshot AI) — the line whose K2.5 seeded Inkling's post-training, and whose K3 (2.8T/104B, days later in the same month) takes the frontier-open scale record Inkling declined to chase — and keeps a vision encoder where Inkling drops one

Open Questions#

• Inkling-Small's 276B/12B dimensions match TML-Interaction-Small exactly. Is the interaction model an Inkling-lineage fine-tune (or vice versa), and will TML unify the two halves of the split into one family?
• Post-training was bootstrapped on Kimi K2.5 synthetic data. Does competitor-bootstrapping leave measurable fingerprints (style, refusal patterns, tokenizer-idiom echoes) that survive 30M rollouts of RL?
• TML claims relative positional embeddings beat RoPE for long-context extrapolation — against current field consensus. Does the claim replicate outside TML at 1M context?

Sources#

• Inkling: Our Open-Weights Model — TML announcement (2026-07, vendor-claim): architecture, training, benchmark tables, effort sweep, safety numbers, ecosystem