Introducing Fabric 1.6

Introduction

Our mission is to make advanced intelligence efficient, open, and deployable everywhere. We have developed a family of open foundation models, an agentic platform that lets anyone run and customise them, and a body of research spanning efficient attention and reinforcement learning. Today we are advancing that mission by releasing Fabric 1.6 — our flagship open reasoning model — with the full weights available on Hugging Face so that people can make it their own.

Fabric 1.6 is a Mixture-of-Experts transformer with 35B total parameters and ~3B active per token. It supports a native context window of 262,144 tokens, extensible to 1,010,000. It reasons natively over text and images, and uses a hybrid linear-attention (DeltaNet) architecture with regularly interspersed full-attention layers to hold long contexts at sub-quadratic memory cost. It is the flagship of a family we are building on: alongside it we share an efficient serving path through llama.cpp and our benchmarks, and we will continue to release models of different sizes built on the same recipe.

Fabric 1.6 reasons before it answers, producing an internal chain of thought it can expose in a structured, streamable format. It balances cost with performance through controllable thinking effort — so a developer fine-tuning for a specialised task can trade tokens for accuracy exactly where it pays. Across 22 benchmarks spanning mathematics, coding, science, general reasoning, and agentic tool use, it posts state-of-the-art results, outperforming models many times its size while remaining efficient enough to run on-premises.

Fabric 1.6 is not the strongest model in every category today. Instead, a combination of qualities makes it a good open-weights base for real workloads: broad generalist performance, an open licence, a long native context, controllable efficiency, and clean serving on the infrastructure you may already own. Fabric 1.6 is the start of a model family we will continue to build on.

Benchmarking Fabric 1.6

Fabric 1.6 was evaluated on 22 benchmarks across mathematics, science and knowledge, coding, general reasoning, agentic tool use, and multimodal understanding. All evaluations used greedy decoding with a fixed random seed. Results are as reported by Fabric AI.

Full benchmark methodology appears in the system card.

Capabilities

Real-world applications require a model with a wide range of capabilities that can be combined — and improved — through fine-tuning. We trained Fabric 1.6 to be broad rather than narrowly optimised for one domain, and we evaluate the qualities that matter most in production: reasoning, agentic tool use, instruction following, factuality, multimodal understanding, and safety.

Generalist model

Fabric 1.6 is designed to be broad. We trained it across mathematics, science and knowledge, coding, general reasoning, agentic tool use, and vision tasks rather than squeezing a single benchmark family. That breadth matters for customisation and real-world use: different users need models that adapt to very different workflows, not just excel on a leaderboard.

On a shared 0–100 scale across representative benchmarks, Fabric 1.6 posts a strong, broad profile — 93.1 on AIME 2026, 92.8 on AIME 2025, 92.3 on C-Eval, 93.5 on MMLU-Redux, 80.2 on LiveCodeBench v6, 72.9 on SWE-bench Verified, and 67.2% pass on TAU3-Bench. The full table appears in Benchmarking below. This breadth reflects Fabric 1.6’s intended role: a practical open foundation model for deployment and customisation across domains and products.

Agentic coding and tool use

A strong base for fine-tuning needs to solve a wide variety of tasks with agentic tool use. Fabric 1.6 was post-trained for reliable tool calling in OpenAI-compatible harnesses — OpenCode, Pi Agent, Hermes Agent, and others — and its tool-calling behaviour is optimised through supervised fine-tuning on structured tool-calling traces followed by reinforcement learning that rewards reliable invocation, instruction following, appropriate reasoning depth, and calibrated refusals.

On TAU3-Bench, which evaluates long-horizon tool use, Fabric 1.6 achieves a 67.2% pass rate; on MCP-Atlas, a 62.8% completion rate; and on WideSearch, a 60.3% rubric score. Multi-Token Prediction delivers up to 50% faster generation, and the gated DeltaNet layers scale sub-quadratically with sequence length — so agents can keep large codebases and long transcripts in context without paying quadratic attention cost at every layer.

Controllable thinking effort

Test-time scaling is the core capability of a reasoning model, but that capacity is hard to capture with a single number. Developers fine-tuning for a specialised task care as much about efficiency as about a model’s max-effort score on a public benchmark; cost and latency are often the binding constraints in real applications.

Fabric 1.6 supports controllable thinking effort, balancing performance with token efficiency for a given task. Its internal chain of thought can be streamed, stored, and reasoned over by the surrounding system, and the depth of reasoning adapts to what the task needs. This lets developers choose the right operating point for each use case — full effort on hard reasoning, lean token budgets on high-volume, low-latency paths.

Multimodality

Fabric 1.6 accepts image inputs alongside text, enabling vision-grounded reasoning and agentic tasks. On multimodal evaluations it scores 74.10% on MMMU-Pro, 85.4% on RealWorldQA, and 86.6% on MathVista mini — strong results on charts, diagrams, and mathematical visual reasoning. For most agentic and document workloads, it can read and reason over images directly, and in long-running sessions it can preserve reasoning context across turns so earlier analysis remains available to later work.

Epistemics

We trained Fabric 1.6 for instruction following, calibrated reasoning, and honest uncertainty — qualities we group as the model’s epistemics. Getting facts right requires more than memorising a large corpus; a useful model must follow complex instructions and know when it does not know.

We aligned the model through reinforcement learning from feedback, optimising for reliable instruction following, helpfulness, and calibrated refusals rather than a single proxy objective. The reward function was designed to improve behaviour end-to-end — including whether it decides to reason, how long it reasons, and how it acts on the result. As with all large language models, Fabric 1.6 can still produce plausible but incorrect information, particularly in niche or rapidly changing domains; verification against authoritative sources is recommended for high-stakes use.

Safety

Safety is crucial for open-weights models. Fabric 1.6 was aligned to an internal specification of responsible model behaviour, and we evaluate it across refusal of harmful requests, dangerous-capability risk (CBRN, cyber), and human-AI threat vectors such as sycophancy and harmful manipulation.

In agentic settings, outputs should be treated as untrusted and executed only within sandboxed, permissioned environments. Tool-calling, while optimised through post-training, remains probabilistic and should be paired with schema-level validation and human oversight where actions carry real-world consequences. Bias and robustness testing across languages and demographic groups is ongoing; Fabric 1.6 was primarily evaluated on the benchmarks in this card.

Benchmarking Fabric 1.6

Fabric 1.6 was evaluated on 22 benchmarks across mathematics, science and knowledge, coding, general reasoning, agentic tool use, and multimodal undattention layers alternate within the model: ten groups, each containing three Gated DeltaNet → MoE blocks followed by one Gated Attention → MoE block. The gated DeltaNet layers provide efficient, hardware-friendly linear attention that scales sub-quadratically with sequence length, while the regularly interspersed gated attention layers supply full-attention capacity where it matters most. This is what enables the 262,144-token native context without quadratic-cost attention at every layer.

Training

Fabric 1.6 is built on Qwen3.5-35B-A3B-Base, an open 35-billion-parameter MoE base model with ~3 billion active parameters released by the Qwen team. From this foundation, Fabric 1.6 was developed through a two-stage programme: continuous pre-training, followed by post-training comprising supervised fine-tuning and reinforcement-learning-based alignment. All stages were conducted with rigorous data filtering, deduplication, and quality control.

Fabric 1.6 was continuously pre-trained primarily on a large in-house proprietary synthetic dataset spanning code, mathematics, and reasoning, complemented by open reasoning corpora — OpenThoughts3-1.2M and OpenR1-Math-220k. In total, approximately 12 billion tokens were processed across the pre-training and post-training stages. Knowledge cutoff: July 2026.

The continuously pre-trained model was then instruction-tuned on a diverse, high-quality SFT corpus combining permissively licensed data with proprietary data: smoltalk2, hermes-function-calling, oasst2, and Fabric AI's proprietary maple instruction and reasoning corpus. Following SFT, the model was aligned through reinforcement learning from feedback, optimising for the behaviours that matter most in agentic deployment: instruction following, reliable tool-calling, appropriate reasoning depth, helpfulness, and calibrated refusals.

Availability

Fabric 1.6 is available now from Fabric AI on Hugging Face, alongside its system card. The model is distributed in BF16 and GGUF formats and runs on CPU/GPU inference stacks, including llama.cpp-based servers with flash attention, quantised KV caches, and MTP-assisted speculative decoding — so it can be deployed on-premises or in your own serving infrastructure. For questions, collaborations, or access requests, contact the Fabric AI research team at research@fabricai.co.uk.

Fabric 1.6 is the first release in a model family we intend to keep building on. Follow our work on X, LinkedIn, and Hugging Face.