SmolLM2-135M-Instruct-bposit8

HuggingFaceTB/SmolLM2-135M-Instruct converted to b-posit8 (32-code blocks with a power-of-two scale, 8-bit posit codes, es = 2) for the exact profile of INVAR: every matmul accumulates in a 256-bit quire with one rounding, so a deterministic runtime produces bit-identical activations and logits on x86, CUDA and aarch64, and independent reference implementations reproduce a served answer from these weights and the token ids.

• File: SmolLM2-135M-Instruct-bposit8.gguf (0.14 GB), SHA-256 f09fe23eb1ec46dfedfd4c6603c2928b3e349190c13b39b7993be8251b52d6d0
• Format: GGUF general.file_type 42 (b-posit8), llama-cpp-et fork (deterministic backend)
• Spec: docs/EXACT-PROFILE-SPEC.md in the INVAR repo; conformance fixture and tokenizer vectors in go/crverify/testdata
• Licence: this is a re-quantised copy of HuggingFaceTB/SmolLM2-135M-Instruct; the upstream licence (apache-2.0) applies unchanged, including any use restrictions and attribution requirements.

Run it: invar serve --model SmolLM2-135M-Instruct-bposit8.gguf --binary llama-cli --spot-check --spot-check-units, then invar verify worldline.jsonl --model SmolLM2-135M-Instruct-bposit8.gguf --binary llama-cli --spot-check --units --reexec.