Keep the capability.
Lose the weight.
A 27-billion-parameter model, compressed to 8.48 GB. Our established recovered build retains more than 80% of its teacher’s score across five matched benchmarks.
A smaller model package.
The model stays capable. The package gets smaller. Anneal compresses existing weights, then recovers capability through further training.
Research build · Not yet releasedRetention gates passed
Above 80% of the original model’s score
Maths, knowledge, commonsense, science and code
Maths score retained
226 / 250 correct · Original: 246 / 250
GSM8K subset · 1,024-token allowance
91.9% · GSM8KTeacher score retained
97.1% · MMLUTeacher score retained
100.6% · HellaSwagTeacher score retained
92.0% · ARC ChallengeTeacher score retained
81.4% · HumanEvalTeacher score retained
Compress.Encode the weights more efficiently.
Recover.Train to bring useful capability back.
Measure.Same questions. Original model alongside.
Read the measurement details
The established recovered build scored 226/250 on GSM8K, 465/570 on MMLU, 689/1,000 on HellaSwag, 334/500 on ARC Challenge and 105/164 on HumanEval. Original model: 246/250, 479/570, 685/1,000, 363/500 and 129/164 respectively. Retention divides the recovered score by the original score. These are matched subsets, not a score for every capability.
A newer variant completed evaluation on 13 September: 90.2% maths, 96.2% knowledge, 101.2% commonsense, 92.0% science and 72.9% coding retention. It passes four of five 80% gates. Coding recovery remains active work; these results are kept separate from the established build.
Quality was evaluated with recovered weights materialized in BF16 on CUDA, separately from native compressed execution. GSM8K uses 8-shot strict match; MMLU 5-shot; HellaSwag 10-shot and ARC 25-shot normalized accuracy. HumanEval is unfiltered pass@1. Generated tasks allow 1,024 tokens. The 8.48 GB figure is file size. Total running memory and long-context optimisation remain in development. See both complete result sets ↗
Closer to
the metal.
A Rust runtime built for local AI on Apple silicon. From the computation graph to the Metal kernels, we’re making the engine fit the model.
Native compressed execution at 4K, through our own Metal runtime.
4,096 prompt tokens
64 generated tokens
125.26 tok/s prompt processing
8.48 GB file · 10.707 GiB weights · 11.936 GiB peak RSS. Memory and long-context optimisation remain in development.
Paired component tests on M1 Max against MLX 0.32.2. These are not whole-model speedups.
Built in Rust.Graph construction, dispatch and memory ownership.
Executed on Metal.Kernels shaped around local hardware.
Ready to connect.An OpenAI-compatible interface for applications.
Hardware, memory & release details
The 4K model run reports 24.07 seconds load time, 10.707 GiB resident weights and 11.936 GiB peak process RSS. RSS is not complete system-memory accounting. Smaller memory use and longer contexts remain active work.
Component benchmarks: paired medians, n=14, after 174/174 parity cases. The fused chain is 2,000 operations; the activation is 64 MB tanh-GELU; matrix multiplication is fp16 4096³. Matrix multiplication matches baseline speed; the decode microbenchmark runs at 0.93× the MLX baseline. Separate workloads from the 27B model run. See the full comparison in our investor brief ↗
Not publicly released. Code availability is planned under the same proposed terms as the Harness; the licence is still being decided.
Think it through.
Work it out.
A different boundary between model and runtime. Deterministic execution happens inside active reasoning. Returned values carry into the next thought; retained fragments become a complete file.
One value. Carried forward.
Follow a GPS clock-drift calculation.
Write. Run. Refine.
Build a file from tested fragments.
Everything in orange is the runtime’s: produced once, carried forward, never re-typed. Scripted demonstrations.
Calculate in the flow.Results enter the active generation.
Carry the result.Reuse a value instead of predicting it again.
A full file. Assembled.The runtime joins the retained fragments. The model never types the file out again.
Working protocol · Capability proof pending. The runtime executes the protocol. The current fine-tune has learned its notation; improved problem-solving and reliable execution selection remain to be demonstrated.
From the current research
The next proof: held-out problems, appropriate execution selection, correct answers, clean stopping and generalisation. Current model evaluation has not established a reasoning gain. Earlier validation contamination and conflicting training targets prompted further data repair.
Training completed 141 optimizer steps across six epochs, with validation prediction loss falling from 2.2570 to 0.6825. All 110 corpus documents replayed. These establish training and protocol behaviour, not improved answer accuracy.
Both Python and Rust hosts passed eight focused error-recovery tests and their 61-case runtime gates. The Rust controller runs inside decoding; an isolated Python/SymPy worker executes operations and feeds the results into the continuing thought.