Forge Daily · Sat Jun 27, 2026ttl-forge · v1.3.0 · Sat 12:21 WEST
Two drops in 48 hours pushed the frontier in opposite directions
Ornith-1.0-9B compresses a 31B-class agentic coding model into 9B via self-scaffolding RL — beats Qwen3.5-9B by 22 points on SWE-Bench Verified, MIT licensed, fits in ~6 GB. Wan-Streamer collapses 5 cascaded multimodal modules into a single Transformer with sub-second ~550 ms full-duplex audio-visual latency — the only end-to-end system publicly disclosed. Plus Qualcomm's $4B Modular grab, US-Iran 14-point peace framework status, and the regulatory stack's next wedge.
SWE-Bench Verified · Ornith-9B69.4
SWE-Bench Verified · Qwen3.5-9B53.2
Terminal-Bench 2.1 · Ornith-9B43.1
Terminal-Bench 2.1 · Gemma 4-31B42.1
Wan-Streamer · end-to-end latency0.55s
GPT-4o Realtime · latency~0.8s
Qualcomm→Modular deal~$4B
Story 1 · Frontier-Model / Reasoning
Ornith-1.0-9B: a 9B model that writes its own RL scaffold
Self-scaffolding RL is the trick. Most agentic systems use a fixed, human-authored harness (OpenHands, Claude Code, Terminus-2). Ornith instead lets the model author its own scaffold during RL — each step runs a scaffold-refinement stage plus a solution-rollout stage, both backpropagated with reward. Per-task-category strategies emerge without hand-tuning per benchmark. Asynchronous pipeline-RL with staleness-weighted GRPO loss handles long agent rollouts. Reward-hacking defenses: a fixed outer trust boundary, deterministic monitor, and a frozen LLM judge as a veto layer.
Published 2026-06-25, MIT licensed, with official GGUF Q4_K_M quantizations that fit in ~6 GB VRAM/RAM. Runs via llama.cpp, Ollama, vLLM, SGLang, Unsloth, Docker Model Runner, and Lemonade out of the box.
Story 2 · Consumer / Voice-Multimodal
Wan-Streamer: full-duplex A/V in one Transformer
~550 ms
Total end-to-end latency · only public system with all 5 capabilities
Single-Transformer, fully causal. Wan-Streamer collapses what is normally 5–6 cascaded modules (VAD / ASR / LM / TTS / animation / video-gen) into one streaming causal Transformer with a Thinker-Performer two-GPU inference pipeline. Paper dropped 2026-06-24 from the Wan Team at Alibaba. Initialized from a Qwen LLM backbone, joint causal audio + video VAEs, flow-matching loss for continuous modalities.
Capabilities: sub-second full-duplex audio + video I/O, model-side ~200 ms response latency, persistent world/dialogue state, turn-taking learned jointly with response timing. The only architecture publicly disclosed that hits all five of: end-to-end (no external LM/ASR/TTS), audio + video out, full-duplex, sub-1s latency, real video output.
Ornith-9B beats its base class on every agentic benchmark
+22 pts
SWE-Bench Verified lead over Qwen3.5-9B (69.4 vs 53.2)
SWE-Bench VerifiedOrnith-9B · 69.469.4
SWE-Bench ProOrnith-9B · 42.942.9
Terminal-Bench 2.1Ornith-9B · 43.143.1
ClawEval (avg)Ornith-9B · 63.163.1
Compare: Qwen3.5-9B53.2 / 31.3 / 21.3 / 53.2
Terminal-Bench 2.1 numbers tell the headline story: 9B dense ≈ Gemma 4-31B on agentic coding (43.1 vs 42.1). For local-coding-agent use cases this is the most interesting small-model drop of June 2026 — Q4_K_M quant in ~6 GB, MIT licensed, GGUF ready. ⚠️ SWE-Bench Verified is increasingly noisy (~20% of "resolved" patches semantically incorrect under strengthened test suites, per March 2026 arXiv:2603.00520).
Capabilities held simultaneously — no competitor scores 4/5
System
Video in
Video out
Full-duplex
End-to-end
Sub-1s
Wan-Streamer
✓
✓
✓
✓
✓
Doubao Realtime
✓
✗
✓
✗
~
GPT-4o Realtime
✓
✗
~
✗
✓
LPM 1.0
~
✓
✓
✗
~
Hallo-Live
~
✓
~
✗
✗
Most systems perceive A/V but only respond in audio (Doubao, GPT-4o Realtime). Others generate A/V but rely on external LM/ASR/TTS (LPM 1.0, Hallo-Live, StreamAvatar) — their reported latencies exclude those dependencies. Wan-Streamer's 550 ms is the first truly end-to-end number.
The latency ladder — Wan-Streamer is the only one under 0.6s end-to-end
3.5× faster
vs Gemini Live worst case (0.55s vs 1.2-3.6s)
Wan-Streamer0.55s
GPT-4o Realtime~0.8s
Doubao Realtime~1.0s
Hume EVI 30.9s
Gemini Live1.2–3.6s
Streaming unit 160 ms, 25 fps, model-side response latency ~200 ms. The Thinker-Performer two-GPU pipeline pipelines current-frame perception + state update, previous-frame decoding, KV/latent communication, and next-frame denoising across adjacent units — the key trick for getting sub-second end-to-end from a single unified model.
Same frontier, opposite moves: compress vs collapse
2026 = post-training year
Heavy lifting moves from base scaling → training recipe + architecture
Compress · Ornith-1.0
"What if the model writes its own training harness?"
Self-scaffolding RL → per-task harness emerges
Pipeline-RL with staleness-weighted GRPO
3-layer reward-hacking defense
9B ≈ 31B on Terminal-Bench 2.1
MIT license, GGUF Q4 in ~6 GB
Collapse · Wan-Streamer
"What if one Transformer does it all, causally?"
5–6 modules → 1 single Transformer
Fully causal A/V stack
Thinker-Performer 2-GPU pipeline
Flow-matching loss for continuous modalities
Sub-second full-duplex A/V
Three macro-trends both releases sit inside: (1) post-training is doing the heavy lifting in 2026 — Ornith's 9B → 30B-equivalent jump comes almost entirely from RL recipe, not base-model scaling. (2) Pipeline thinking is being abandoned in favor of single-model unification. (3) Open-source models are closing the gap with closed frontier in vertical domains — Ornith-397B beats Claude Opus 4.7 on Terminal-Bench 2.1 and SWE-Bench Verified (still loses to Opus 4.8).
Qualcomm grabs Modular for ~$4B — biggest Nvidia CUDA challenge yet
~$4B
All-stock deal · targets developer-tool moat, not just chips
Qualcomm disclosed the all-stock acquisition of Modular, the portable-AI-workloads startup that lets developers run AI models across processors without rewriting code. Strategic read: Nvidia's dominance has rested on CUDA as the default AI runtime; Modular makes that runtime portable across Qualcomm's expanding footprint (smartphones, data centers, cars, edge). The same week OpenAI received first samples of its Broadcom-designed "Jalapeño" inference chip and SK Hynix filed a $29.4B Nasdaq ADR — the inference stack is fragmenting from "Nvidia-only" to multi-supplier faster than most expected.
~19.78% of patches labeled "resolved" by top-30 agents are semantically incorrect under strengthened test suites. The top agent's score drops 78.80% → 62.20% with stronger eval. >32% of SWE-Bench Verified instances contain shortcuts.
arXiv:2603.00520 · March 2026
Read every 2026 SWE-Bench Verified leaderboard number — including Ornith's 82.4 on the 397B flagship — through this lens. The benchmark is increasingly noisy; per-parameter efficiency arguments (Ornith-9B at ~3× fewer params matching Gemma 4-31B) survive the noise better than absolute leaderboard claims.