# K2-Horizon-0.9B Technical Appendix This appendix contains the detailed architecture, checkpoint, training, deployment, hardware, evaluation, dataset, source-code, intended-use, and safety material moved from the full model card. The short [README](README.md) follows the common K2-Horizon release format. ## Model Series Overview K2-Horizon spans compact and larger dense models for transparent foundation-model research, staged checkpoint analysis, and practical deployment. The repositories share a common release structure while preserving model-specific training and serving guidance. | Model | Architecture | Parameters | Context length | Intended use | | --- | --- | ---: | ---: | --- | | **K2-Horizon-0.9B** | Dense decoder-only | 1.08B including embeddings | 131,072 | Compact reasoning, local inference, and distillation research | | K2-Horizon-3.7B | Dense decoder-only | 3.78B core | 524,288 | Efficient research, evaluation, and single-node serving | | K2-Horizon-7B | Dense decoder-only | 7B core | 524,288 | General research, fine-tuning, and cost-conscious deployment | | K2-Horizon-32B | Dense decoder-only | 32B core | 524,288 | Stronger long-context and reasoning experiments | ## This Repository | Property | Value | |---|---| | **Architecture** | `K2HorizonForCausalLM` (`model_type: k2_horizon`) | | **Parameters** | 1,078,285,824 (released as 0.9B; counted from the safetensors tensors) | | **Hidden size / layers** | 1,536 / 28 | | **Attention heads / KV heads** | 32 / 8 | | **Context length** | 131,072 tokens with YaRN RoPE scaling; original context length 8,192 tokens | | **Vocabulary size** | 64,256 | | **Released weight dtype** | BF16 | | **Format** | Hugging Face safetensors, one weight shard, with custom configuration and modeling code in the repository root | | **Distillation checkpoint** | Step 249 of a 500-step mOPD run | The `main` revision publishes `K2HorizonForCausalLM`, `model_type: k2_horizon`, and matching `configuration_k2_horizon.py` and `modeling_k2_horizon.py` modules. The Transformers and vLLM preflights below validate that public contract before loading weights. ## Checkpoint Revisions | Revision | Stage | Max context | | --- | --- | ---: | | `main` | Distilled release checkpoint, step 249 | 128K | | `mid2_47k` | Second context-extension stage | 128K | | `mid1_75k` | First context-extension stage | 40K | | Base model | Pre-distillation specialist merge | 8K | ## Training Provenance Training examples were routed to a math-and-code teacher, a STEM teacher, or an instruction-following teacher through the example's `opd_domain` metadata. The math-and-code teacher was used as the fallback when no recognized domain was present. | Domain | Teacher checkpoint step | |---|---:| | Math and code | 2,739 | | STEM | 499 | | Instruction following | 1,499 | The base context window was extended in stages from 8,192 to 40,960 and then to 131,072 tokens. The `mid1_75k` and `mid2_47k` repository revisions preserve the corresponding intermediate checkpoints. The distilled release checkpoint is on `main`; all stages use a vocabulary of 64,256 tokens. ## Training Loss K2-Horizon-0.9B begins with a task-arithmetic merge of three specialist checkpoints. mOPD then trains that merged student against math-and-code, STEM, and instruction-following teachers at the same time. The training objective combines an on-policy distillation loss with a reference-model KL term so the student can learn specialist behavior while remaining close to the merged base model. The resulting checkpoint retains most of the specialist teachers' performance on the reported math and coding tasks. It also improves every reported IFEval submetric over the pre-distillation merge. This makes the model useful for research on compact reasoning models, local inference, distillation, and task-specific adaptation. ## Deployment Guide K2-Horizon-0.9B emits a reasoning segment before its final answer when the chat template is used. The template supports `reasoning_effort` values `high`, `medium`, and `low`, which select the model's full, fast, and faster reasoning modes respectively. For a deterministic runtime check, use `temperature=0` and generate 20 to 50 tokens. For general sampled generation, `temperature=0.6` and `top_p=0.95` reproduce the GPQA evaluation setting and are reasonable starting points. IFBench used `temperature=0.8`. Long math and coding tasks may require several thousand output tokens; choose limits from application measurements rather than treating an evaluation limit as a universal default. ### vLLM The validated serving image was reconstructed into the following manual runtime contract: | Component | Validated value | |---|---| | Operating system | Ubuntu 24.04, Linux x86-64 | | Python | 3.12.13 | | CUDA toolkit | 12.9 | | PyTorch | 2.13.0+cu129 | | Transformers | 5.16.1 | | Safetensors | 0.8.0 | | FlashInfer | 0.6.17 | | Attention backend | vLLM FlashAttention 3; Triton 3.7.1 | | vLLM | `0.26.1rc1.dev1212`, [PR #53806](https://github.com/vllm-project/vllm/pull/53806) source commit [`d9fd5f11`](https://github.com/vllm-project/vllm/commit/d9fd5f11423a1a5628fe29e7296ceb9de91aac3c) | That source revision contains the native `K2HorizonForCausalLM` implementation and the built-in `k2_horizon` reasoning and tool parsers. Other vLLM revisions have not been validated for this checkpoint. Pin the exact commit until the integration is available in an upstream release. Use Linux x86-64 with a CUDA 12.9-compatible NVIDIA driver and Git. The setup uses vLLM's precompiled extension path while keeping the Python package on the exact reviewed source commit.
Show the pinned vLLM environment setup ```bash git clone --filter=blob:none --no-checkout \ https://github.com/vllm-project/vllm.git cd vllm git fetch origin pull/53806/head:refs/remotes/origin/pr-53806 git checkout --detach d9fd5f11423a1a5628fe29e7296ceb9de91aac3c test "$(git rev-parse HEAD)" = \ "d9fd5f11423a1a5628fe29e7296ceb9de91aac3c" python3.12 -m venv .venv source .venv/bin/activate python -m pip install --upgrade pip uv export UV_LINK_MODE=copy VLLM_USE_PRECOMPILED=1 uv pip install --upgrade --editable . \ --torch-backend=auto uv pip install "transformers==5.16.1" "safetensors==0.8.0" python -m pip check python - <<'PY' from vllm import ModelRegistry from vllm.reasoning import ReasoningParserManager from vllm.tool_parsers import ToolParserManager assert "K2HorizonForCausalLM" in ModelRegistry.get_supported_archs() assert ReasoningParserManager.get_reasoning_parser("k2_horizon") is not None assert ToolParserManager.get_tool_parser("k2_horizon") is not None PY ```
Download the repository chat template explicitly and start the server with one GPU. The 8,192-token profile below is a conservative starting point. Increase `MAX_MODEL_LEN` only after measuring KV-cache capacity; the checkpoint supports up to 131,072 tokens.
Show the vLLM serving command ```bash source .venv/bin/activate export MODEL_ID="IFM/K2-Horizon-0.9B" export MODEL_REVISION="main" export MAX_MODEL_LEN=8192 export CHAT_TEMPLATE="$(hf download "$MODEL_ID" chat_template.jinja \ --revision "$MODEL_REVISION")" vllm serve "$MODEL_ID" \ --revision "$MODEL_REVISION" \ --model-impl vllm \ --trust-remote-code \ --dtype bfloat16 \ --tensor-parallel-size 1 \ --max-model-len "$MAX_MODEL_LEN" \ --max-num-seqs 1 \ --gpu-memory-utilization 0.85 \ --served-model-name "$MODEL_ID" \ --chat-template "$CHAT_TEMPLATE" \ --reasoning-parser k2_horizon \ --tool-call-parser k2_horizon \ --enable-auto-tool-choice ```
The reasoning parser moves ``, ``, or `` text into the OpenAI-compatible response's `reasoning_content` field. The tool parser converts generated `` blocks into structured tool calls when tools are supplied in the request.
Show an OpenAI-compatible request ```python from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="unused") response = client.chat.completions.create( model="IFM/K2-Horizon-0.9B", messages=[{"role": "user", "content": "What is the square root of 2?"}], max_tokens=50, temperature=0, extra_body={"chat_template_kwargs": {"reasoning_effort": "high"}}, ) message = response.choices[0].message print(getattr(message, "reasoning_content", None)) print(message.content) ```
### SGLang Native K2 Horizon support is provided by [sgl-project/sglang#37654](https://github.com/sgl-project/sglang/pull/37654). Use a `lmsysorg/sglang:dev` image built after that PR is merged. Once support is included in a tagged SGLang release, use the corresponding versioned image.
Show the SGLang serving command ```bash docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --ipc=host \ lmsysorg/sglang:dev \ python3 -m sglang.launch_server \ --model-path "IFM/K2-Horizon-0.9B" \ --revision main \ --tp 1 \ --dtype bfloat16 \ --context-length 8192 \ --attention-backend fa3 \ --reasoning-parser k2_horizon \ --tool-call-parser k2_horizon \ --mem-fraction-static 0.85 \ --host 0.0.0.0 \ --port 30000 ```
This uses SGLang's native `K2HorizonForCausalLM` implementation; no `--trust-remote-code`, source patch, or external parser plugin is required. The 8,192-token limit is a conservative starting point; increase it only after measuring KV-cache capacity. ### Transformers The checkpoint can also be loaded directly from the Hugging Face repository. Its `configuration_k2_horizon.py` and `modeling_k2_horizon.py` files are loaded through `trust_remote_code=True`. Use a clean environment so the direct path does not inherit vLLM's build dependencies. Transformers 4.57.x is not compatible with this remote configuration class; use the validated 5.14.1 version below. Transformers may print nonfatal `cache_position` documentation diagnostics while loading the remote code, but BF16 loading and generation complete normally.
Show the Transformers environment setup ```bash python3.12 -m venv .venv-transformers source .venv-transformers/bin/activate python -m pip install --upgrade pip python -m pip install "torch==2.11.0" \ --index-url https://download.pytorch.org/whl/cu128 python -m pip install \ "transformers==5.14.1" \ "safetensors==0.8.0" python -m pip check ```
This deterministic sample loads the released weights as BF16 on one CUDA GPU and generates only 50 tokens, making it suitable as an end-to-end smoke test.
Show the Transformers inference example ```python import torch from transformers import AutoModelForCausalLM, AutoTokenizer MODEL_ID = "IFM/K2-Horizon-0.9B" REVISION = "main" tokenizer = AutoTokenizer.from_pretrained( MODEL_ID, revision=REVISION, trust_remote_code=True, ) model = AutoModelForCausalLM.from_pretrained( MODEL_ID, revision=REVISION, dtype=torch.bfloat16, trust_remote_code=True, ).to("cuda").eval() messages = [{"role": "user", "content": "What is the square root of 2?"}] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, reasoning_effort="high", return_dict=True, return_tensors="pt", ) inputs = {name: value.to(model.device) for name, value in inputs.items()} inputs.pop("token_type_ids", None) with torch.inference_mode(): outputs = model.generate( **inputs, max_new_tokens=50, do_sample=False, pad_token_id=tokenizer.pad_token_id, ) new_tokens = outputs[0, inputs["input_ids"].shape[-1]:] print(tokenizer.decode(new_tokens, skip_special_tokens=True)) ```
## Hardware Planning The BF16 weights occupy approximately 2.0 GiB. One CUDA GPU is sufficient for short-context inference; 8 GiB is a practical minimum for a one-request smoke test, while 16 GiB or more provides useful room for longer prompts and runtime workspaces. The BF16 KV cache is approximately 0.44 GiB per request at 8,192 tokens and approximately 7 GiB at 131,072 tokens, before allocator, activation, CUDA-graph, and framework overhead. Start with a short context and one sequence, then increase context length and concurrency from measured memory headroom. ## Source Code The model repository stores the checkpoint directly at its root. It includes: - `model.safetensors` and `model.safetensors.index.json` - `config.json`, `generation_config.json`, and the K2 architecture code - tokenizer files and three chat-template variants - `README.md` and `LICENSE` Use `chat_template.jinja` for ordinary chat and OpenAI-compatible serving. `chat_template_generation.jinja` and `chat_template_asst_tool_gen.jinja` are specialized generation and assistant-tool-generation variants. ## Evaluation - **Training:** multi-teacher on-policy distillation with OPD loss weight 0.1, reference-KL weight 0.01, learning rate `1e-7`, and a 500-step schedule. The selected checkpoint is step 249. - **Evaluation:** the AIME, coding, and instruction-following evaluations used the training evaluation path with SGLang as the rollout engine. GPQA-Diamond was evaluated with Eval360-V2 at revision `f5081bf`. - **Export:** the distributed training checkpoint was converted to Hugging Face safetensors and checked for tensor parity; all 255 expected weight tensors matched. The architecture label was later updated from `XllmForCausalLM` to `K2HorizonForCausalLM` without changing the weights. ## Intended Use K2-Horizon-0.9B is intended for compact reasoning research, local inference, evaluation dry runs, distillation studies, and task-specific adaptation. It is a research release and should be evaluated on representative prompts before deployment. ## Limitations and Safety - AIME 2025 remains below the math-and-code teacher, and AIME results have high sampling uncertainty because each benchmark contains only 30 problems. - GPQA-Diamond is statistically close to the pre-distillation base result, so the reported run does not demonstrate a clear STEM improvement. - Long-horizon tool use remains substantially weaker than single-turn tool calling in the BFCL v4 breakdown. - Benchmark scores depend on prompt templates, reasoning effort, sampling parameters, framework versions, and evaluation harness details. Validate the model on representative prompts before deployment. - As with other language models, K2-Horizon-0.9B can produce inaccurate, biased, or unsafe text. Applications should use task-specific evaluation, input and output controls, monitoring, and human review where appropriate.