A 276B Model Just Beat the 975B That Taught It to Code

TL;DR

Thinking Machines released the full weights of Inkling-Small, a 276B-parameter Mixture-of-Experts model with 12B active parameters, under Apache 2.0. It scores 80.2% on SWE-bench Verified, which is higher than the Inkling flagship that helped train it (77.6%), at roughly a quarter of the parameters and 30% of the output price. It also takes Terminal-Bench 2.1, IFBench and GPQA Diamond off its own teacher. The download is 532 GB.

What actually shipped

When Thinking Machines put out Inkling on July 15, the announcement mentioned a smaller sibling and said the weights would arrive once testing finished. Testing finished. The Hugging Face repo now holds 32 safetensors shards plus a 4.46 GB multi-token-prediction module, 532 GB in total, Apache 2.0, no field-of-use rider attached.

The architecture: a 42-layer decoder-only transformer with sparse MoE routing and hybrid local plus global attention, 276B total parameters with 12B active per token, a context window up to 1M tokens, and native inputs for text, images and audio. Output is text only. Training ran on NVIDIA GB300 NVL72 systems, the same hardware behind the flagship. You can fine-tune it on Tinker or poke at it in the Tinker Playground.

The student beat the teacher

Thinking Machines says it post-trained an earlier Inkling-Small preview checkpoint partly through on-policy distillation with Inkling as the teacher, then scaled agentic coding reinforcement learning on top for about two more weeks. On-policy distillation is the part worth understanding: instead of grading a student's finished homework against a textbook, you sit the teacher beside the student while it works and correct each move as it happens, so the student learns on the distribution of its own mistakes rather than someone else's transcript.

The awkward result for the flagship, and the good result for you, is that the pupil now outscores the teacher on most of the board:

• SWE-bench Verified: 80.2% vs 77.6%
• Terminal-Bench 2.1: 64.7% vs 63.8%
• HLE, text only: 31.6% vs 29.7%
• IFBench: 82.2% vs 79.8%
• GPQA Diamond: 89.5% vs 87.2%
• MMMU Pro: 74.0% vs 73.5%

The flagship keeps competition math and audio. Inkling still leads AIME 2026 (97.1% vs 95.5%), Audio MC (56.6% vs 54.9%), MMAU (77.2% vs 77.0%) and VoiceBench (91.4% vs 90.1%). Every number above was run at reasoning effort 0.99 and temperature 1.0.

Against the rest of the open field

Thinking Machines benchmarked it against the current sparse-MoE crowd. On SWE-bench Verified it leads Qwen3.5-397B-A17B (76.4%), MiMo V2.5 (71.0%), Minimax M2.7 (79.9%), DeepSeek V4 Flash (79.0%) and Nemotron 3 Ultra (70.7%). On Terminal-Bench 2.1 the gap is wider: 64.7% against 63.7% for the next best and 51.3% for Qwen3.5.

It does not sweep. DeepSeek V4 Flash edges it on HLE text-only (32.1%) and AIME 2026 (95.8%), and both Qwen3.5 (77.3%) and MiMo V2.5 (75.4%) beat it on MMMU Pro. Note also that Inkling-Small is the smallest model in that comparison except Minimax M2.7, and it is running 12B active against Nemotron 3 Ultra's 55B.

Standard disclaimer, and it matters more than usual on a day-zero release: these are the vendor's own harnesses and its own numbers. Nobody outside the lab has replicated them yet.

The price cut is the other half of the story

Serving cost tracks active parameters, not total, so a 12B-active model is dramatically cheaper to run than a 41B-active one. Thinking Machines passed that through.

The announcement quotes output pricing only and does not break out an input rate, and it calls both figures a limited-time discount without saying when the discount ends. Budget accordingly.

What it takes to actually run

532 GB across 32 shards works out to two bytes per parameter, so what you are downloading is a bf16 checkpoint. The model card lists F32, BF16 and NVFP4 as supported formats, which is the interesting one: a 4-bit cast drops the resident footprint by roughly a factor of four and puts this inside a single dense GPU node rather than a rack.

The sparsity is what makes it worth the trouble. Only 12B parameters fire per token, so you get the throughput of a small dense model while paying the memory bill of a large one. Think of a 276-lawyer firm where exactly 12 people ever touch your case: fast billing, but you are still paying rent on the whole building.

For scale, this is about a third the download of Kimi K3's 1.56 TB, and it ships under plain Apache 2.0, which after the last two weeks of license tripwires in open-weight land counts as a feature rather than a formality.

The caveats

• Vendor-run benchmarks, day zero, zero independent replication.
• SWE-bench Verified is crowded near the top now. A 1.2-point lead over Minimax M2.7 is inside the noise you would expect from harness differences.
• Multimodality is input-only. It reads images and listens to audio, it writes text.
• The audio benchmarks are the one place the flagship still clearly holds ground, so if speech is your workload, do not assume the small model is a drop-in.

Key Takeaways

• Inkling-Small is out with full weights on Hugging Face under Apache 2.0: 276B total, 12B active, 1M context, 532 GB.
• It beats its own 975B teacher on SWE-bench Verified (80.2% vs 77.6%), Terminal-Bench 2.1, HLE text-only, IFBench, GPQA Diamond and MMMU Pro.
• The teacher keeps AIME 2026 (97.1%) and the audio suite.
• Output pricing is $1.20 per 1M tokens against $4.05 for Inkling, both marked as a limited-time discount.
• On-policy distillation plus roughly two weeks of agentic coding RL is the recipe Thinking Machines credits for the jump.
• All benchmark figures are self-reported and unreplicated. Treat the leaderboard position as provisional.

Sources: Thinking Machines: Introducing Inkling-Small, Hugging Face model card, Thinking Machines: Inkling, MarkTechPost, TechCrunch