Key takeaways
- Hugging Face published its harness-adapter recipe on November 20, 2026, with Qwen3.8-27B LoRAs for pi, Hermes Agent and OpenCode and a merged super-adapter.
- The release includes downloadable artifacts rather than an announced future weights release. The merged repository lists Apache-2.0 in its license metadata; deployment still requires the base model and compatible serving software.
- The publisher reports better coding-task pass rates and fewer tool calls within each tested harness. These are not AccessAllGPT measurements or a ranking of the harnesses.
- Try a specialist for one harness with thinking off, or the merged adapter for several harnesses or thinking on. Evaluate your own tasks before treating lower tool counts as lower total cost.
What is available now
Hugging Face published a reusable harness-adapter training recipe on November 20, 2026, alongside downloadable Qwen3.8-27B adapters for coding agents. Instead of changing the agent application alone, the work trains a small model-weight update with reinforcement learning inside the actual harness. The target is a model that finishes software-engineering and terminal tasks more accurately with less work.
The released identifiers are harness-hub/Qwen3.8-27B-pi-adapted, harness-hub/Qwen3.8-27B-hermes-adapted and harness-hub/Qwen3.8-27B-opencode-adapted, plus harness-hub/Qwen3.8-27B-super-adapter. Each uses Qwen/Qwen3.8-27B as its base. The public collection also lists GGUF variants, training tasks and a safeagents plugin. We verified the collection listing and the merged repository’s adapter files, not a successful weight download or local run.
The merged model card declares Apache-2.0 and the repository API reports it as public and ungated. This is an adapter release, not a complete new base model or a managed API launch. We did not verify a hosted inference endpoint or an API token price. Download access does not remove the cost of running the base model, and operators should review the base model’s terms and all deployment dependencies separately.
The model learns inside the agent loop it will use
The authors trained one LoRA per harness using TRL’s AsyncGRPOTrainer, OpenEnv and Harbor. Their published setup uses pi 1.5.2, OpenCode 2.2.7 and Hermes Agent v0.22.6, with thinking disabled during training. The reward combines verifier correctness with a small tool-efficiency bonus. These details matter: an adapter tuned against one tool protocol and task distribution is not automatically optimized for every agent.
According to the announcement, each training setup used two NVIDIA B200 GPUs, one for the vLLM server and one for training. That is the reported training arrangement, not a minimum inference requirement. The final super-adapter combines the three rank-32 updates at half weight into a rank-96 adapter; the merge itself requires no additional training.
The practical change for builders is another optimization option between swapping the whole model and rewriting the harness. A team operating an open model can evaluate a task-specific weight update while retaining its agent interface. That is AccessAllGPT analysis, not evidence that a production application will improve without retesting.
What the reported gains do and do not establish
In the pi specialist model card, the publisher reports SWE-bench Pro hard-51 pass rates of 39.2% for the base model and 64.7% for the adapter, and Terminal-Bench 2.1 rates of 39.5% and 55.8%. These are publisher-reported benchmark results, not independent tests. The comparisons use the same harness and settings on each side; they must not be read as a leaderboard comparing pi, Hermes Agent and OpenCode.
The announcement reports that the merged adapter with thinking enabled raises the mean pass rate across six harness–benchmark pairs from 55.5% to 63.5%. Its shared-task comparison reports 53–73% fewer tool calls with thinking off. Efficiency is measured on tasks that both runs solve, not as an unconditional reduction across every incoming request or a measured dollar saving.
The small evaluation sets make individual rows uncertain. The merged card explicitly says most individual paired results are within noise, while pooled results are not. Its headline, shared-task table and full paired tables also use different task intersections, so some percentages differ within the card and from the blog. Do not combine a baseline from one table with an adapter score from another. The authors acknowledge that choosing the merge weight on Terminal-Bench makes those results slightly optimistic.
Out-of-distribution testing is a useful warning against overgeneralizing. On 300 AutomationBench business-workflow tasks using a different agent loop, the blog reports 23.0% for the base model and 23.7% for the merged adapter; the Hermes specialist reaches 17.7%. The authors describe most differences as noise. Coding-task gains are not evidence of a general automation upgrade.
Serving and evaluation constraints to check first
The merged card’s vLLM instructions load the base model with LoRA enabled, a maximum supported rank of 128 for the rank-96 adapter, the qwen3_xml tool-call parser and the qwen3 reasoning parser. Its example sets a 131,072-token context and uses --mamba-cache-mode align with prefix caching. These are documented configuration requirements, not commands we executed or a claim of compatibility with every vLLM version.
The card’s benchmark setup is text-only: the vision tower was not served. It uses one attempt per task, temperature 0.8 and a 100-step budget, with time limits differing by benchmark. The blog also notes harness differences in context handling. This release does not establish vision performance, unrestricted long-running reliability or what happens after a harness upgrade.
The authors observed short reasoning-like text appearing before tool calls even when the chat template’s thinking mode was disabled. Thinking off therefore does not mean the model necessarily produces no reasoning-like content. Treat token accounting and visible output as measured behaviors, not assumptions based on a mode label.
Which adapter should a builder try?
The publisher recommends a specialist when using one harness with thinking off, and the super-adapter for several harnesses or thinking on. That is a useful starting hypothesis. Pin the base revision, adapter revision, harness, serving version, tool parser, context limit and thinking setting before comparing against your current model.
Start with a disposable repository and non-sensitive tasks. Measure verified task completion, regressions, wall-clock time, generated tokens and tool calls on the same task set. Include hardware utilization and serving overhead before claiming cost savings. Keep permission controls and independent tests in place; a shorter trajectory is not itself proof that a change is correct or safe.
Adapter selection for a first pilot
A proposed evaluation plan, not hands-on results.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Compare its specialist against the unchanged base model using identical settings and independently graded tasks.
Try the super-adapter, preserve separate results per harness and record context-overflow failures.
Require task-specific evidence; the reported AutomationBench results do not show a clear general transfer benefit.
Primary sources
Browse the publication-wide evidence index →
- Introducing harness adapters: adapt your favorite model to your harness to improve performance and costHugging Face · Reviewed: Publication date; training inside harnesses; training setup; adapter merging; evaluation setup and results; out-of-distribution evaluation; reasoning leakage discussion · Retrieved · Supports: Announcement date, released recipe, harness versions, training hardware, publisher-reported performance and efficiency claims, benchmark-selection and transfer limitations.
- Qwen3.8-27B super-adapter model cardharness-hub on Hugging Face · Reviewed: License and base metadata; specialist selection; shared-task and full paired results; efficiency denominators; caveats; vLLM use · Retrieved · Supports: Adapter identity, Apache-2.0 metadata, rank, serving configuration, text-only evaluation and different result denominators. Reports the authors’ measurements, not independent validation.
- Qwen3.8-27B pi specialist model cardharness-hub on Hugging Face · Reviewed: Base and license metadata; headline paired results; evaluation settings; available deployment paths · Retrieved · Supports: Specialist identifier, exact quoted specialist pass rates, thinking setting and benchmark comparability warning.
- Qwen3.8-27B harness adapters collectionharness-hub on Hugging Face · Reviewed: Public collection API model and dataset listings · Retrieved · Supports: Discoverable specialist, merged and GGUF repository identifiers, training artifacts and plugin; not execution or file-integrity verification.
- Merged adapter repository metadataHugging Face Hub API · Reviewed: id; sha; private; gated; cardData; siblings · Retrieved · Supports: Public ungated status, exact inspected revision, license metadata and presence of adapter_config.json and adapter_model.safetensors. No weights were downloaded or hashed.
Limitations
No hands-on inference, weight download, training, independent benchmark reproduction, hosted API access check or total-cost measurement. Results are publisher-reported, some comparison tables use different task intersections, and merge-weight selection used a target benchmark. We inspected the collection listings, not every GGUF file or specialist repository. Model cards and documentation are mutable.
Disclosures
Public-source reporting and AccessAllGPT analysis, not a sponsored evaluation. The release authors did not review or endorse this briefing. No claim is made that Mann Jadwani tested these adapters; publication-wide relationships appear on the disclosures page.



