Open-weight coding agents and efficient fine-tuning

Tess 4.5 publishes a 327 MB coding adapter for the 594B Inkling 2 Small model

A public LoRA and merged checkpoint adapt a huge sparse model for software agents; publisher benchmarks suggest a narrow gain, not universal superiority.

Open Model Weights published17 Nov 2026
Primary sourceContinuity Research / Migel Tissera — Tess-4.5-594B-A17B official model card
Source publishedNovember 2026 (exact day not stated)
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original midnight violet editorial illustration of a glowing amber adapter prism connected to a large glass neural matrix with fine code-like light trails.
Original conceptual artwork of parameter-efficient coding specialization; not an official model photograph or a performance graph. AI-generated editorial illustration; not a documentary photograph.

Continuity Research developer Migel Tissera has released Tess-4.5-594B-A17B, a coding-focused derivative of Thinking Machines' Inkling 2 Small. The public model card links both a full merged checkpoint and a separate LoRA adapter. Tess is not a newly pretrained 594-billion-parameter model: it inherits Inkling's sparse architecture, its 17 billion active parameters per token and its million-token context. The new contribution lies in a targeted post-training update intended to improve agentic engineering work. Primary source ↗

The developer describes training on 46.9 million tokens from anonymized engineering sessions. Instead of fine-tuning the routed experts, which are stored in low-precision formats, the adapter targets attention projections and the model's output head. That choice allows the update to be merged into the BF16 tensors without requiring expert-weight requantization. The separately released adapter is about 327 MB; the full merged checkpoint is approximately 394 GB. Both artifacts are far from a trivial consumer-laptop deployment.

On the author's reported 51-task SWE-Bench Pro v2 Hard evaluation, Tess resolved 88.2 percent of tasks versus 86.3 percent for the underlying Inkling checkpoint, while producing roughly nine percent fewer output tokens. LongBench v2 results were nearly even at 70.2 versus 70.6 percent. The test details provide a useful lead for replication, but the numbers should remain attributed to the publisher rather than presented as Open Model Weights measurements or a demonstrated gain across all coding and reasoning tasks.

The project describes deployment using Inkling-compatible vLLM and SGLang paths and recommends at least 500 GB of aggregated GPU memory for the merged model. The Hugging Face page declares Apache-2.0, though exact checkpoint revisions and dependencies still require verification before a formal registry addition. Tess illustrates a practical feature of open-weight ecosystems: a relatively small inspectable adapter can produce a meaningful behavioral specialization without changing most of a massive sparse base model.

OMW REGISTRY WATCH

What this changes for the evidence layer

Treat Tess 4.5 as a derivative of the existing Inkling 2 Small family. Check the public merged weights, separate 327 MB adapter, exact source revisions and license text before introducing a verified derivative entry. The developer's 88.2% vs 86.3% result is from its 51-task hard benchmark slice and is not an independent OMW evaluation.

PRIMARY SOURCE

Continuity Research / Migel Tissera — Tess-4.5-594B-A17B official model card

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗